fix the implementation of the relative position encoding in conformer, optimize…

fix the implementation of the relative position encoding in conformer, optimize the code of ctc loss

egs/libri_trans/asr/conf/conformer.yaml

 macaron-style: True
 use-cnn-module: True
 cnn-module-kernel: 31
+encoder-attention-type: rel_pos
\ No newline at end of file

egs/libri_trans/asr/conf/debug.yaml

@@ -13,7 +13,7 @@ label_smoothing: 0.1
 
 subsampling-type: conv1d
 subsmapling-layers: 2
-subsampling-filter: 1024
+subsampling-filter: 512
 subsampling-kernel: 5
 subsampling-stride: 2
 subsampling-norm: none
@@ -32,3 +32,9 @@ decoder-ffn-embed-dim: 2048
 decoder-attention-heads: 4
 attention-dropout: 0.1
 activation-dropout: 0.1
+
+macaron-style: True
+use-cnn-module: True
+cnn-module-kernel: 31
+encoder-activation-fn: swish
+encoder-attention-type: rel_pos

egs/libri_trans/asr/conf/rpr.yaml

-encoder-attention-type: rel_selfattn
+encoder-attention-type: rel_pos
 #encoder-attention-type: relative
 #max-encoder-relative-length: 100

egs/librispeech/asr/conf/ConformerCTCSmall.yaml

@@ -4,14 +4,20 @@ clip-norm: 10.0
 lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-updates: 10000
-lr: 2e-3
-#adam_betas: (0.9,0.98)
+lr: 0.0015
+adam_betas: (0.9,0.98)
 
 criterion: ctc
 post-process: sentencepiece
 
-conv-kernel-sizes: 5,5
-conv-channels: 704
+subsampling-type: conv1d
+subsmapling-layers: 2
+subsampling-filter: 1024
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
 dropout: 0.1
 activation-fn: relu
 encoder-embed-dim: 176
@@ -22,4 +28,5 @@ encoder-attention-heads: 4
 macaron-style: True
 use-cnn-module: True
 cnn-module-kernel: 31
-encoder-attention-type: rel_selfattn
\ No newline at end of file
+encoder-activation-fn: swish
+encoder-attention-type: rel_pos
\ No newline at end of file

egs/librispeech/asr/conf/base.yaml

@@ -6,13 +6,19 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 lr: 2e-3
-#adam_betas: (0.9,0.98)
+adam_betas: (0.9,0.98)
 
 criterion: label_smoothed_cross_entropy_with_ctc
 label_smoothing: 0.1
 
-conv-kernel-sizes: 5,5
-conv-channels: 1024
+subsampling-type: conv1d
+subsmapling-layers: 2
+subsampling-filter: 1024
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
 dropout: 0.1
 activation-fn: relu
 encoder-embed-dim: 256

egs/librispeech/asr/conf/conformer.yaml

 macaron-style: True
 use-cnn-module: True
 cnn-module-kernel: 31
+encoder-activation-fn: swish
+encoder-attention-type: rel_pos
\ No newline at end of file

egs/mustc/asr/conf/base.yaml

@@ -6,13 +6,19 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 lr: 2e-3
-#adam_betas: (0.9,0.98)
+adam_betas: (0.9,0.98)
 
 criterion: label_smoothed_cross_entropy_with_ctc
 label_smoothing: 0.1
 
-conv-kernel-sizes: 5,5
-conv-channels: 1024
+subsampling-type: conv1d
+subsmapling-layers: 2
+subsampling-filter: 1024
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
 dropout: 0.1
 activation-fn: relu
 encoder-embed-dim: 256

egs/mustc/asr/conf/inter.yaml

+ctc-weight: 0.2
 intermedia-ctc-layers: 6,9
 intermedia-adapter: league
-intermedia-ctc-weight: 0.15
-ctc-self-distill-weight: 1
\ No newline at end of file
+intermedia-ctc-weight: 0.1
+ctc-self-distill-weight: 0
+post-process: sentencepiece
\ No newline at end of file

egs/mustc/asr/conf/purectc.yaml

@@ -5,21 +5,28 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 lr: 2e-3
-#adam_betas: (0.9,0.98)
+adam_betas: (0.9,0.98)
 
 criterion: ctc
 zero_infinity: True
 post-process: sentencepiece
-label_smoothing: 0.1
 
-conv-kernel-sizes: 5,5
-conv-channels: 1024
+subsampling-type: conv1d
+subsmapling-layers: 2
+subsampling-filter: 1024
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
 dropout: 0.1
+attention-dropout: 0.1
+activation-dropout: 0.1
+
 activation-fn: relu
 encoder-embed-dim: 256
 encoder-ffn-embed-dim: 2048
 encoder-layers: 12
 encoder-attention-heads: 4
 
-attention-dropout: 0.1
-activation-dropout: 0.1
\ No newline at end of file
+#load-pretrained-encoder-from:
\ No newline at end of file

egs/mustc/asr/train.sh

@@ -6,7 +6,6 @@ gpu_num=8
 update_freq=1
 max_tokens=40000
 
-
 extra_tag=
 extra_parameter=
 #extra_tag="${extra_tag}"
@@ -14,12 +13,12 @@ extra_parameter=
 
 exp_tag=
 
-#config_list=(base)
-#config_list=(ctc)
+config_list=(base ctc)
+config_list=(purectc)
 #config_list=(base conformer)
 
 #config_list=(pds_base_16)
-config_list=(pds_base_16 conformer rpr)
+#config_list=(pds_base_16 conformer rpr)
 
 # exp full name
 exp_name=

egs/mustc/st/conf/base.yaml

@@ -6,13 +6,19 @@ lr-scheduler: inverse_sqrt
 warmup-init-lr: 1e-7
 warmup-updates: 10000
 lr: 2e-3
-#adam_betas: (0.9,0.98)
+adam_betas: (0.9,0.98)
 
 criterion: label_smoothed_cross_entropy_with_ctc
 label_smoothing: 0.1
 
-conv-kernel-sizes: 5,5
-conv-channels: 1024
+subsampling-type: conv1d
+subsmapling-layers: 2
+subsampling-filter: 1024
+subsampling-kernel: 5
+subsampling-stride: 2
+subsampling-norm: none
+subsampling-activation: glu
+
 dropout: 0.1
 activation-fn: relu
 encoder-embed-dim: 256

egs/mustc/st/conf/inter.yaml

+ctc-weight: 0.2
 intermedia-ctc-layers: 6,9
 intermedia-adapter: league
-intermedia-ctc-weight: 0.15
-ctc-self-distill-weight: 1
\ No newline at end of file
+intermedia-ctc-weight: 0.1
+ctc-self-distill-weight: 0
+post-process: sentencepiece
\ No newline at end of file

fairseq/criterions/ctc.py

@@ -20,6 +20,8 @@ from fairseq.tasks import FairseqTask
 from fairseq.logging.meters import safe_round
 
 
+
+
 @dataclass
 class CtcCriterionConfig(FairseqDataclass):
     zero_infinity: bool = field(
@@ -35,6 +37,19 @@ class CtcCriterionConfig(FairseqDataclass):
             "See fairseq.data.data_utils.post_process() for full list of options"
         },
     )
+    ctc_entropy: float = field(
+        default=0.0,
+        metadata={"help": "weight of CTC entropy"},
+    )
+    intermedia_ctc_weight: float = field(
+        default=0.0,
+        metadata={"help": "weight of intermedia CTC loss"},
+    )
+    ctc_self_distill_weight: float = field(
+        default=0.0,
+        metadata={"help": "weight of the self distillation CTC loss"},
+    )
+
     wer_kenlm_model: Optional[str] = field(
         default=None,
         metadata={
@@ -64,12 +79,8 @@ class CtcCriterionConfig(FairseqDataclass):
 
 @register_criterion("ctc", dataclass=CtcCriterionConfig)
 class CtcCriterion(FairseqCriterion):
-    def __init__(self, cfg: CtcCriterionConfig, task: FairseqTask):
+    def __init__(self, cfg: CtcCriterionConfig, task: FairseqTask, ctc_weight=1.0):
         super().__init__(task)
-        self.blank_idx = task.target_dictionary.index(task.blank_symbol) if hasattr(task, 'blank_symbol') else 0
-        self.pad_idx = task.target_dictionary.pad()
-        self.eos_idx = task.target_dictionary.eos()
-        self.post_process = cfg.post_process
 
         if cfg.wer_args is not None:
             (
@@ -99,48 +110,151 @@ class CtcCriterion(FairseqCriterion):
         else:
             self.w2l_decoder = None
 
-        self.zero_infinity = cfg.zero_infinity
+        self.blank_idx = task.target_dictionary.index(task.blank_symbol) if hasattr(task, 'blank_symbol') else 0
+        self.pad_idx = task.target_dictionary.pad()
+        self.eos_idx = task.target_dictionary.eos()
+        self.post_process = cfg.post_process
         self.sentence_avg = cfg.sentence_avg
 
+        self.ctc_weight = ctc_weight
+        self.intermedia_ctc_weight = cfg.intermedia_ctc_weight
+        self.ctc_self_distill_weight = cfg.ctc_self_distill_weight
+        self.ctc_entropy = cfg.ctc_entropy
+        self.all_ctc_weight = self.ctc_weight + self.intermedia_ctc_weight + self.ctc_self_distill_weight + self.ctc_entropy
+    
+        if self.all_ctc_weight > 0:
+            assert getattr(task, "src_dict", None) is not None, "CTC need a source dictionary."
+            self.ctc_loss = torch.nn.CTCLoss(blank=self.blank_idx, reduction="sum", zero_infinity=True)
+
     def forward(self, model, sample, reduce=True):
         net_output = model(**sample["net_input"])
-        lprobs = model.get_normalized_probs(
-            net_output, log_probs=True
-        ).contiguous()  # (T, B, C) from the encoder
 
-        non_padding_mask = ~net_output["encoder_padding_mask"][0]
-        input_lengths = non_padding_mask.long().sum(-1)
-
-        pad_mask = (sample["target"] != self.pad_idx) & (
-            sample["target"] != self.eos_idx
-        )
-        targets_flat = sample["target"].masked_select(pad_mask)
-        target_lengths = pad_mask.sum(-1)
-
-        with torch.backends.cudnn.flags(enabled=False):
-            loss = F.ctc_loss(
-                lprobs,
-                targets_flat,
-                input_lengths,
-                target_lengths,
-                blank=self.blank_idx,
-                reduction="sum",
-                zero_infinity=self.zero_infinity,
-            )
-
-        ntokens = (
-            sample["ntokens"] if "ntokens" in sample else target_lengths.sum().item()
-        )
+        ntokens = sample["ntokens"]
 
         sample_size = sample["target"].size(0) if self.sentence_avg else ntokens
         logging_output = {
-            "loss": utils.item(loss.data),  # * sample['ntokens'],
             "ntokens": ntokens,
             "nsentences": sample["id"].numel(),
             "sample_size": sample_size,
         }
 
-        if not model.training:
+        loss, logging_output = self.compute_ctc_loss(model, sample, net_output, logging_output)
+        return loss, sample_size, logging_output
+
+    def compute_ctc_loss(self, model, sample, net_output, logging_output):
+        transcript = sample["transcript"]
+        if "ctc_padding_mask" in net_output:
+            non_padding_mask = ~net_output["ctc_padding_mask"][0]
+        else:
+            non_padding_mask = ~net_output["encoder_padding_mask"][0]
+        input_lengths = non_padding_mask.long().sum(-1)
+
+        pad_mask = (transcript["tokens"] != self.pad_idx) & (
+                transcript["tokens"] != self.eos_idx
+        )
+        targets_flat = transcript["tokens"].masked_select(pad_mask)
+        transcript_lengths = pad_mask.sum(-1)
+
+        ctc_loss = 0
+        ctc_entropy = 0
+        lprobs = None
+        if self.ctc_weight > 0 and "ctc_logit" in net_output and len(net_output["ctc_logit"]) > 0:
+            ctc_logit = net_output["ctc_logit"][0]
+            lprobs = model.get_normalized_probs(
+                [ctc_logit], log_probs=True
+            ).contiguous()  # (T, B, C) from the encoder
+            lprobs.batch_first = False
+
+            with torch.backends.cudnn.flags(enabled=False):
+                ctc_loss = self.ctc_loss(
+                    lprobs,
+                    targets_flat,
+                    input_lengths,
+                    transcript_lengths,
+                )
+            if self.ctc_entropy > 0:
+                from torch.distributions import Categorical
+                # ctc_logit = ctc_logit.sort(dim=-1, descending=True)[0][:, :, 0:100]
+                # ctc_logit = ctc_logit / ctc_logit.sum(dim=-1, keepdim=True)
+                cut_ctc_logit = ctc_logit.sort(dim=-1, descending=True)[0][:, :, 0:100]
+                ctc_entropy = Categorical(logits=cut_ctc_logit).entropy().sum()
+                # ctc_entropy = Categorical(logits=ctc_logit).entropy().sum()
+                logging_output["ctc_entropy"] = utils.item(ctc_entropy.data)
+            logging_output["ctc_loss"] = utils.item(ctc_loss.data)
+
+        intermedia_ctc_num = 0
+        intermedia_ctc_loss = 0
+        if "intermedia_ctc_logits" in net_output:
+            intermedia_ctc_num = len(net_output["intermedia_ctc_logits"])
+
+        # calculate the intermedia CTC loss
+        if self.intermedia_ctc_weight > 0 and intermedia_ctc_num > 0:
+            for i in range(intermedia_ctc_num):
+                out = net_output["intermedia_ctc_logits"][i]
+                if type(out) == list:
+                    inter_ctc_logit = out[0]
+                    padding = ~out[1]
+                    input_lengths = padding.long().sum(-1)
+                else:
+                    inter_ctc_logit = out
+
+                inter_lprobs = model.get_normalized_probs(
+                    [inter_ctc_logit], log_probs=True
+                ).contiguous()  # (T, B, C) from the encoder
+                inter_lprobs.batch_first = False
+
+                with torch.backends.cudnn.flags(enabled=False):
+                    loss = self.ctc_loss(
+                        inter_lprobs,
+                        targets_flat,
+                        input_lengths,
+                        transcript_lengths,
+                    )
+                intermedia_ctc_loss += loss
+            intermedia_ctc_loss /= intermedia_ctc_num
+            logging_output["intermedia_ctc_loss"] = utils.item(intermedia_ctc_loss.data)
+
+            if lprobs is None:
+                lprobs = inter_lprobs
+
+        # calculate the self distillation CTC loss
+        ctc_self_distill_loss = 0
+        ctc_self_distill_num = 0
+        if self.ctc_weight > 0 and self.ctc_self_distill_weight > 0 and intermedia_ctc_num > 0:
+            for i in range(intermedia_ctc_num):
+                out = net_output["intermedia_ctc_logits"][i]
+                if type(out) == list:
+                    inter_ctc_logit = out[0]
+                    padding = ~out[1]
+                    input_lengths = padding.long().sum(-1)
+                else:
+                    inter_ctc_logit = out
+
+                if inter_ctc_logit.size() != ctc_logit.size():
+                    continue
+
+                ctc_self_distill_num += 1
+                loss = F.kl_div(
+                    F.log_softmax(inter_ctc_logit, dim=-1),
+                    F.softmax(ctc_logit, dim=-1),
+                    reduction="none",
+                )
+                loss = loss.sum(-1).transpose(0, 1).masked_fill_(~non_padding_mask, 0.0)
+                loss = loss.sum()
+                ctc_self_distill_loss += loss
+
+            ctc_self_distill_loss /= ctc_self_distill_num
+            logging_output["ctc_self_distill_loss"] = utils.item(ctc_self_distill_loss.data)
+
+        loss = \
+            self.ctc_weight * ctc_loss + \
+            self.intermedia_ctc_weight * intermedia_ctc_loss + \
+            self.ctc_self_distill_weight * ctc_self_distill_loss + \
+            self.ctc_entropy * ctc_entropy
+
+        logging_output["all_ctc_loss"] = utils.item(loss.data)
+
+        if not model.training and self.ctc_weight > 0:
             import editdistance
 
             with torch.no_grad():
@@ -153,9 +267,7 @@ class CtcCriterion(FairseqCriterion):
                 wv_errs = 0
                 for lp, t, inp_l in zip(
                     lprobs_t,
-                    sample["target_label"]
-                    if "target_label" in sample
-                    else sample["target"],
+                    sample["transcript"]["tokens"] if "transcript" in sample else sample["target"],
                     input_lengths,
                 ):
                     lp = lp[:inp_l].unsqueeze(0)
@@ -207,13 +319,28 @@ class CtcCriterion(FairseqCriterion):
                 logging_output["c_errors"] = c_err
                 logging_output["c_total"] = c_len
 
-        return loss, sample_size, logging_output
+        return loss, logging_output
 
     @staticmethod
     def reduce_metrics(logging_outputs) -> None:
         """Aggregate logging outputs from data parallel training."""
 
-        loss_sum = utils.item(sum(log.get("loss", 0) for log in logging_outputs))
+        ctc_loss_sum = utils.item(
+            sum(log.get("ctc_loss", 0) for log in logging_outputs)
+        )
+        ctc_entropy_sum = utils.item(
+            sum(log.get("ctc_entropy", 0) for log in logging_outputs)
+        )
+        inter_ctc_loss_sum = utils.item(
+            sum(log.get("intermedia_ctc_loss", 0) for log in logging_outputs)
+        )
+        ctc_self_distill_loss_sum = utils.item(
+            sum(log.get("ctc_self_distill_loss", 0) for log in logging_outputs)
+        )
+        all_ctc_loss_sum = utils.item(
+            sum(log.get("all_ctc_loss", 0) for log in logging_outputs)
+        )
+        # loss_sum = utils.item(sum(log.get("loss", 0) for log in logging_outputs))
         ntokens = utils.item(sum(log.get("ntokens", 0) for log in logging_outputs))
         nsentences = utils.item(
             sum(log.get("nsentences", 0) for log in logging_outputs)
@@ -221,15 +348,56 @@ class CtcCriterion(FairseqCriterion):
         sample_size = utils.item(
             sum(log.get("sample_size", 0) for log in logging_outputs)
         )
+        if all_ctc_loss_sum > 0:
+            if "loss" not in logging_outputs[0]:
+                metrics.log_scalar(
+                    "loss",
+                    all_ctc_loss_sum / sample_size / math.log(2),
+                    sample_size,
+                    round=3,
+                )
+            else:
+                if all_ctc_loss_sum != ctc_loss_sum:
+                    metrics.log_scalar(
+                        "all_ctc_loss",
+                        all_ctc_loss_sum / sample_size / math.log(2),
+                        sample_size,
+                        round=3,
+                    )
+        if ctc_loss_sum > 0:
+            metrics.log_scalar(
+                "ctc_loss",
+                ctc_loss_sum / sample_size / math.log(2),
+                sample_size,
+                round=3,
+            )
+        if ctc_entropy_sum > 0:
+            metrics.log_scalar(
+                "ctc_entropy",
+                ctc_entropy_sum / sample_size / math.log(2),
+                sample_size,
+                round=3,
+            )
+        if inter_ctc_loss_sum > 0:
+            metrics.log_scalar(
+                "intermedia_ctc_loss",
+                inter_ctc_loss_sum / sample_size / math.log(2),
+                sample_size,
+                round=3,
+            )
+        if ctc_self_distill_loss_sum > 0:
+            metrics.log_scalar(
+                "ctc_self_distill_loss",
+                ctc_self_distill_loss_sum / sample_size / math.log(2),
+                sample_size,
+                round=3,
+            )
 
-        metrics.log_scalar(
-            "loss", loss_sum / sample_size / math.log(2), sample_size, round=3
-        )
         metrics.log_scalar("ntokens", ntokens)
         metrics.log_scalar("nsentences", nsentences)
         if sample_size != ntokens:
             metrics.log_scalar(
-                "nll_loss", loss_sum / ntokens / math.log(2), ntokens, round=3
+                "nll_loss", ctc_loss_sum / ntokens / math.log(2), ntokens, round=3
             )
 
         c_errors = sum(log.get("c_errors", 0) for log in logging_outputs)

fairseq/criterions/label_smoothed_cross_entropy_with_ctc.py

@@ -13,41 +13,28 @@ from fairseq.data.data_utils import post_process
 from fairseq.logging.meters import safe_round
 
 from .label_smoothed_cross_entropy import LabelSmoothedCrossEntropyCriterion
+from .ctc import CtcCriterion, CtcCriterionConfig
 
 
 @register_criterion("label_smoothed_cross_entropy_with_ctc")
 class LabelSmoothedCrossEntropyCriterionWithCTC(
     LabelSmoothedCrossEntropyCriterion
 ):
-    def __init__(self, task, sentence_avg, label_smoothing, post_process="letter",
-                 ctc_weight=0.0, intermedia_ctc_weight=0.0, ctc_self_distill_weight=0.0):
+    def __init__(self, task, label_smoothing,
+                 sentence_avg,
+                 cfg: CtcCriterionConfig,
+                 ctc_weight=0.0):
         super().__init__(task, sentence_avg, label_smoothing)
-        self.blank_idx = task.target_dictionary.index(task.blank_symbol) if hasattr(task, 'blank_symbol') else 0
-        self.pad_idx = task.target_dictionary.pad()
-        self.eos_idx = task.target_dictionary.eos()
 
         self.report_accuracy = True
-
-        assert 0 <= ctc_weight
-        self.top_ctc_weight = ctc_weight
-        self.intermedia_ctc_weight = intermedia_ctc_weight
-        self.ctc_self_distill_weight = ctc_self_distill_weight
-        self.ctc_weight = ctc_weight + intermedia_ctc_weight + ctc_self_distill_weight
-        if self.ctc_weight > 0:
-            assert getattr(task, "src_dict", None) is not None, "CTC need a source dictionary."
-            self.post_process = post_process
-            self.ctc_loss = torch.nn.CTCLoss(blank=self.blank_idx, reduction="sum", zero_infinity=True)
+        self.ctc_weight = ctc_weight
+        self.ctc_criterion = CtcCriterion(cfg, task, ctc_weight)
 
     @staticmethod
     def add_args(parser):
         """Add criterion-specific arguments to the parser."""
         LabelSmoothedCrossEntropyCriterion.add_args(parser)
-        parser.add_argument(
-            "--zero-infinity",
-            default=True,
-            type=bool,
-            help="zero inf loss when source length <= target length",
-        )
+        CtcCriterion.add_args(parser)
         parser.add_argument(
             "--ctc-weight",
             default=0.0,
@@ -55,33 +42,6 @@ class LabelSmoothedCrossEntropyCriterionWithCTC(
             metavar="D",
             help="weight of CTC loss",
         )
-        parser.add_argument(
-            "--intermedia-ctc-weight",
-            default=0.0,
-            type=float,
-            metavar="D",
-            help="weight of intermedia CTC loss",
-        )
-        parser.add_argument(
-            "--ctc-self-distill",
-            action="store_true",
-            help="use self distillation for intermedia CTC loss",
-        )
-        parser.add_argument(
-            "--ctc-self-distill-weight",
-            default=0.0,
-            type=float,
-            metavar="D",
-            help="weight of the self distillation CTC loss",
-        )
-        parser.add_argument(
-            "--post-process",
-            default="letter",
-            type=str,
-            help="how to post process predictions into words. can be letter, "
-            "word-piece, BPE symbols, etc. "
-            "See fairseq.data.data_utils.post_process() for full list of options",
-        )
 
     def forward(self, model, sample, reduce=True):
         """Compute the loss for the given sample.
@@ -114,173 +74,13 @@ class LabelSmoothedCrossEntropyCriterionWithCTC(
             logging_output["n_correct"] = utils.item(n_correct.data)
             logging_output["total"] = utils.item(total.data)
 
-        if self.ctc_weight > 0:
-            ctc_loss, logging_output = self.compute_ctc_loss(model, sample, encoder_out, logging_output)
-            loss = (1 - self.top_ctc_weight - self.intermedia_ctc_weight) * loss + ctc_loss
+        if self.ctc_criterion.all_ctc_weight > 0:
+            ctc_loss, logging_output = self.ctc_criterion.compute_ctc_loss(model, sample, encoder_out, logging_output)
+            loss = (1 - self.ctc_weight) * loss + ctc_loss
         logging_output["loss"] = utils.item(loss.data) if reduce else loss.data
 
         return loss, sample_size, logging_output
 
-    def compute_ctc_loss(self, model, sample, encoder_out, logging_output):
-        transcript = sample["transcript"]
-        if "ctc_padding_mask" in encoder_out:
-            non_padding_mask = ~encoder_out["ctc_padding_mask"][0]
-        else:
-            non_padding_mask = ~encoder_out["encoder_padding_mask"][0]
-        input_lengths = non_padding_mask.long().sum(-1)
-
-        pad_mask = (transcript["tokens"] != self.pad_idx) & (
-                transcript["tokens"] != self.eos_idx
-        )
-        targets_flat = transcript["tokens"].masked_select(pad_mask)
-        transcript_lengths = pad_mask.sum(-1)
-
-        ctc_loss = 0
-        lprobs = None
-        if self.top_ctc_weight > 0 and "ctc_logit" in encoder_out and len(encoder_out["ctc_logit"]) > 0:
-            ctc_logit = encoder_out["ctc_logit"][0]
-            lprobs = model.get_normalized_probs(
-                [ctc_logit], log_probs=True
-            ).contiguous()  # (T, B, C) from the encoder
-            lprobs.batch_first = False
-
-            with torch.backends.cudnn.flags(enabled=False):
-                ctc_loss = self.ctc_loss(
-                    lprobs,
-                    targets_flat,
-                    input_lengths,
-                    transcript_lengths,
-                )
-            logging_output["ctc_loss"] = utils.item(ctc_loss.data)
-
-        intermedia_ctc_num = 0
-        intermedia_ctc_loss = 0
-        if "intermedia_ctc_logits" in encoder_out:
-            intermedia_ctc_num = len(encoder_out["intermedia_ctc_logits"])
-
-        # calculate the intermedia CTC loss
-        if self.intermedia_ctc_weight > 0 and intermedia_ctc_num > 0:
-            for i in range(intermedia_ctc_num):
-                out = encoder_out["intermedia_ctc_logits"][i]
-                if type(out) == list:
-                    inter_ctc_logit = out[0]
-                    padding = ~out[1]
-                    input_lengths = padding.long().sum(-1)
-                else:
-                    inter_ctc_logit = out
-
-                inter_lprobs = model.get_normalized_probs(
-                    [inter_ctc_logit], log_probs=True
-                ).contiguous()  # (T, B, C) from the encoder
-                inter_lprobs.batch_first = False
-
-                with torch.backends.cudnn.flags(enabled=False):
-                    loss = self.ctc_loss(
-                        inter_lprobs,
-                        targets_flat,
-                        input_lengths,
-                        transcript_lengths,
-                    )
-                intermedia_ctc_loss += loss
-            intermedia_ctc_loss /= intermedia_ctc_num
-            logging_output["intermedia_ctc_loss"] = utils.item(intermedia_ctc_loss.data)
-
-            if lprobs is None:
-                lprobs = inter_lprobs
-
-        # calculate the self distillation CTC loss
-        ctc_self_distill_loss = 0
-        ctc_self_distill_num = 0
-        if self.top_ctc_weight > 0 and self.ctc_self_distill_weight > 0 and intermedia_ctc_num > 0:
-            for i in range(intermedia_ctc_num):
-                out = encoder_out["intermedia_ctc_logits"][i]
-                if type(out) == list:
-                    inter_ctc_logit = out[0]
-                    padding = ~out[1]
-                    input_lengths = padding.long().sum(-1)
-                else:
-                    inter_ctc_logit = out
-
-                if inter_ctc_logit.size() != ctc_logit.size():
-                    continue
-
-                ctc_self_distill_num += 1
-                loss = F.kl_div(
-                    F.log_softmax(inter_ctc_logit, dim=-1),
-                    F.softmax(ctc_logit, dim=-1),
-                    reduction="none",
-                )
-                loss = loss.sum(-1).transpose(0, 1).masked_fill_(~non_padding_mask, 0.0)
-                loss = loss.sum()
-                ctc_self_distill_loss += loss
-
-            ctc_self_distill_loss /= ctc_self_distill_num
-            logging_output["ctc_self_distill_loss"] = utils.item(ctc_self_distill_loss.data)
-
-        loss = \
-            self.ctc_weight * ctc_loss + \
-            self.intermedia_ctc_weight * intermedia_ctc_loss + \
-            self.ctc_self_distill_weight * ctc_self_distill_loss
-
-        if self.intermedia_ctc_weight > 0 or self.ctc_self_distill_weight > 0:
-            logging_output["all_ctc_loss"] = utils.item(loss.data)
-
-        if not model.training and self.ctc_weight > 0:
-            import editdistance
-
-            with torch.no_grad():
-                lprobs_t = lprobs.transpose(0, 1).float().contiguous().cpu()
-
-                c_err = 0
-                c_len = 0
-                w_errs = 0
-                w_len = 0
-                wv_errs = 0
-                for lp, t, inp_l in zip(
-                    lprobs_t,
-                    sample["transcript"]["tokens"] if "transcript" in sample else sample["target"],
-                    input_lengths,
-                ):
-                    lp = lp[:inp_l].unsqueeze(0)
-
-                    decoded = None
-                    p = (t != self.task.target_dictionary.pad()) & (
-                        t != self.task.target_dictionary.eos()
-                    )
-                    targ = t[p]
-                    targ_units = self.task.target_dictionary.string(targ)
-                    targ_units_arr = targ.tolist()
-
-                    toks = lp.argmax(dim=-1).unique_consecutive()
-                    pred_units_arr = toks[toks != self.blank_idx].tolist()
-
-                    c_err += editdistance.eval(pred_units_arr, targ_units_arr)
-                    c_len += len(targ_units_arr)
-
-                    targ_words = post_process(targ_units, self.post_process).split()
-
-                    pred_units = self.task.target_dictionary.string(pred_units_arr)
-                    pred_words_raw = post_process(pred_units, self.post_process).split()
-
-                    if decoded is not None and "words" in decoded:
-                        pred_words = decoded["words"]
-                        w_errs += editdistance.eval(pred_words, targ_words)
-                        wv_errs += editdistance.eval(pred_words_raw, targ_words)
-                    else:
-                        dist = editdistance.eval(pred_words_raw, targ_words)
-                        w_errs += dist
-                        wv_errs += dist
-
-                    w_len += len(targ_words)
-
-                logging_output["wv_errors"] = wv_errs
-                logging_output["w_errors"] = w_errs
-                logging_output["w_total"] = w_len
-                logging_output["c_errors"] = c_err
-                logging_output["c_total"] = c_len
-
-        return loss, logging_output
-
     @staticmethod
     def reduce_metrics(logging_outputs) -> None:
         """Aggregate logging outputs from data parallel training."""
@@ -291,19 +91,6 @@ class LabelSmoothedCrossEntropyCriterionWithCTC(
         nll_loss_sum = utils.item(
             sum(log.get("nll_loss", 0) for log in logging_outputs)
         )
-        ctc_loss_sum = utils.item(
-            sum(log.get("ctc_loss", 0) for log in logging_outputs)
-        )
-        inter_ctc_loss_sum = utils.item(
-            sum(log.get("intermedia_ctc_loss", 0) for log in logging_outputs)
-        )
-        ctc_self_distill_loss_sum = utils.item(
-            sum(log.get("ctc_self_distill_loss", 0) for log in logging_outputs)
-        )
-        all_ctc_loss_sum = utils.item(
-            sum(log.get("all_ctc_loss", 0) for log in logging_outputs)
-        )
-
         ntokens = utils.item(sum(log.get("ntokens", 0) for log in logging_outputs))
         sample_size = utils.item(
             sum(log.get("sample_size", 0) for log in logging_outputs)
@@ -319,38 +106,11 @@ class LabelSmoothedCrossEntropyCriterionWithCTC(
         metrics.log_scalar(
             "nll_loss", nll_loss_sum / ntokens / math.log(2), ntokens, round=3
         )
-        if ctc_loss_sum > 0:
-            metrics.log_scalar(
-                "ctc_loss",
-                ctc_loss_sum / sample_size / math.log(2),
-                sample_size,
-                round=3,
-            )
-        if inter_ctc_loss_sum > 0:
-            metrics.log_scalar(
-                "intermedia_ctc_loss",
-                inter_ctc_loss_sum / sample_size / math.log(2),
-                sample_size,
-                round=3,
-            )
-        if ctc_self_distill_loss_sum > 0:
-            metrics.log_scalar(
-                "ctc_self_distill_loss",
-                ctc_self_distill_loss_sum / sample_size / math.log(2),
-                sample_size,
-                round=3,
-            )
-        if all_ctc_loss_sum > 0:
-            metrics.log_scalar(
-                "all_ctc_loss",
-                all_ctc_loss_sum / sample_size / math.log(2),
-                sample_size,
-                round=3,
-            )
-
         metrics.log_derived(
             "ppl", lambda meters: utils.get_perplexity(meters["nll_loss"].avg)
         )
+        if "ctc_loss" in logging_outputs[0] or "all_ctc_loss" in logging_outputs[0]:
+            CtcCriterion.reduce_metrics(logging_outputs)
 
         total = utils.item(sum(log.get("total", 0) for log in logging_outputs))
         if total > 0:
@@ -368,44 +128,6 @@ class LabelSmoothedCrossEntropyCriterionWithCTC(
                 else float("nan"),
             )
 
-        c_errors = sum(log.get("c_errors", 0) for log in logging_outputs)
-        metrics.log_scalar("_c_errors", c_errors)
-        c_total = sum(log.get("c_total", 0) for log in logging_outputs)
-        metrics.log_scalar("_c_total", c_total)
-        w_errors = sum(log.get("w_errors", 0) for log in logging_outputs)
-        metrics.log_scalar("_w_errors", w_errors)
-        wv_errors = sum(log.get("wv_errors", 0) for log in logging_outputs)
-        metrics.log_scalar("_wv_errors", wv_errors)
-        w_total = sum(log.get("w_total", 0) for log in logging_outputs)
-        metrics.log_scalar("_w_total", w_total)
-
-        if c_total > 0:
-            metrics.log_derived(
-                "cer",
-                lambda meters: safe_round(
-                    meters["_c_errors"].sum * 100.0 / meters["_c_total"].sum, 3
-                )
-                if meters["_c_total"].sum > 0
-                else float("nan"),
-            )
-        if w_total > 0:
-            metrics.log_derived(
-                "wer",
-                lambda meters: safe_round(
-                    meters["_w_errors"].sum * 100.0 / meters["_w_total"].sum, 3
-                )
-                if meters["_w_total"].sum > 0
-                else float("nan"),
-            )
-            metrics.log_derived(
-                "raw_wer",
-                lambda meters: safe_round(
-                    meters["_wv_errors"].sum * 100.0 / meters["_w_total"].sum, 3
-                )
-                if meters["_w_total"].sum > 0
-                else float("nan"),
-            )
-
     @staticmethod
     def logging_outputs_can_be_summed() -> bool:
         """

fairseq/models/speech_to_text/modules/adapter.py

@@ -85,34 +85,45 @@ class InterAdapter(nn.Module):
 
         if self.adapter_type == "shrink":
             self.ctc_compress = getattr(CTCCompressStrategy, strategy)
+            logger.info("CTC Compress Strategy: %s" % strategy)
+        elif self.adapter_type == "league":
+            self.distribution_cutoff = strategy
+            if self.distribution_cutoff != -1:
+                logger.info("Distribution cutoff: %d" % int(strategy))
 
     def forward(self, x, padding):
 
         representation, distribution = x
         dim1, dim2, dim = representation.size()
         org_distribution = distribution
-        if distribution is not None:
-            distribution = distribution.view(-1, distribution.size(-1))
         lengths = (~padding).long().sum(-1)
 
         if self.adapter_type == "linear":
             out = self.linear_adapter(representation)
 
         elif self.adapter_type == "context":
+            distribution = distribution.view(-1, distribution.size(-1))
             out = torch.mm(distribution, self.embed_adapter.weight).view(dim1, dim2, -1)
 
         elif self.adapter_type == "league":
             linear_out = self.linear_adapter(representation)
+            if self.distribution_cutoff != -1:
+                cutoff = min(int(self.distribution_cutoff), distribution.size(-1) - 1)
+                threshold = distribution.sort(dim=-1, descending=True)[0][:, :, cutoff:cutoff+1]
+                distribution = torch.where(distribution > threshold, distribution, torch.zeros_like(distribution))
+            distribution = distribution.view(-1, distribution.size(-1))
             soft_out = torch.mm(distribution, self.embed_adapter.weight).view(dim1, dim2, -1)
             out = linear_out + soft_out
 
         elif self.adapter_type == "gated_league":
             linear_out = self.linear_adapter(representation)
+            distribution = distribution.view(-1, distribution.size(-1))
             soft_out = torch.mm(distribution, self.embed_adapter.weight).view(dim1, dim2, -1)
             coef = (self.gate_linear(torch.cat([linear_out, soft_out], dim=-1))).sigmoid()
             out = coef * linear_out + (1 - coef) * soft_out
 
         elif self.adapter_type == "inter_league":
+            distribution = distribution.view(-1, distribution.size(-1))
             soft_out = torch.mm(distribution, self.embed_adapter.weight).view(dim1, dim2, -1)
             out = representation + soft_out
 

fairseq/models/speech_to_text/s2t_ctc.py

@@ -15,73 +15,23 @@ from fairseq.models import (
     register_model_architecture,
 )
 from fairseq.models.speech_to_text.modules import InterAdapter, CTC
-from fairseq.models.transformer import Embedding, TransformerDecoder
 from fairseq.modules import (
     FairseqDropout,
     LayerNorm,
     PositionalEmbedding,
+    RelPositionalEncoding,
     S2TTransformerEncoderLayer,
     DynamicLinearCombination,
 )
+from fairseq.modules.speech_to_text import (
+    subsampling
+)
+
 from torch import Tensor
 
 logger = logging.getLogger(__name__)
 
 
-class Conv1dSubsampler(nn.Module):
-    """Convolutional subsampler: a stack of 1D convolution (along temporal
-    dimension) followed by non-linear activation via gated linear units
-    (https://arxiv.org/abs/1911.08460)
-
-    Args:
-        in_channels (int): the number of input channels
-        mid_channels (int): the number of intermediate channels
-        out_channels (int): the number of output channels
-        kernel_sizes (List[int]): the kernel size for each convolutional layer
-    """
-
-    def __init__(
-            self,
-            in_channels: int,
-            mid_channels: int,
-            out_channels: int,
-            kernel_sizes: List[int] = (3, 3),
-    ):
-        super(Conv1dSubsampler, self).__init__()
-        self.n_layers = len(kernel_sizes)
-        self.conv_layers = nn.ModuleList(
-            nn.Conv1d(
-                in_channels if i == 0 else mid_channels // 2,
-                mid_channels if i < self.n_layers - 1 else out_channels * 2,
-                k,
-                stride=2,
-                padding=k // 2,
-            )
-            for i, k in enumerate(kernel_sizes)
-        )
-
-    def get_out_seq_lens_tensor(self, in_seq_lens_tensor):
-        out = in_seq_lens_tensor.clone()
-        for _ in range(self.n_layers):
-            out = ((out.float() - 1) / 2 + 1).floor().long()
-        return out
-
-    def forward(self, src_tokens, src_lengths):
-        bsz, in_seq_len, _ = src_tokens.size()  # B x T x (C x D)
-        x = src_tokens.transpose(1, 2).contiguous()  # -> B x (C x D) x T
-        inner_x = []
-        for conv in self.conv_layers:
-            x = conv(x)
-            x = nn.functional.glu(x, dim=1)
-            inner_x.append(x)
-        _, _, out_seq_len = x.size()
-        # x = x.transpose(1, 2).transpose(0, 1).contiguous()  # -> T x B x (C x D)
-        out_inner_x = []
-        for x in inner_x:
-            out_inner_x.append(x.transpose(1, 2).transpose(0, 1).contiguous())
-        return out_inner_x, self.get_out_seq_lens_tensor(src_lengths)
-
-
 @register_model("s2t_ctc")
 class S2TCTCModel(FairseqEncoderModel):
 
@@ -91,18 +41,43 @@ class S2TCTCModel(FairseqEncoderModel):
     @staticmethod
     def add_args(parser):
         """Add model-specific arguments to the parser."""
-        # input
+        # subsampling
         parser.add_argument(
-            "--conv-kernel-sizes",
+            "--subsampling-type",
             type=str,
-            metavar="N",
-            help="kernel sizes of Conv1d subsampling layers",
+            help="subsampling type, like conv1d and conv2d",
         )
         parser.add_argument(
-            "--conv-channels",
+            "--subsampling-layers",
             type=int,
-            metavar="N",
-            help="# of channels in Conv1d subsampling layers",
+            help="subsampling layers",
+        )
+        parser.add_argument(
+            "--subsampling-filter",
+            type=int,
+            help="subsampling filter",
+        )
+        parser.add_argument(
+            "--subsampling-kernel",
+            type=int,
+            help="subsampling kernel",
+        )
+        parser.add_argument(
+            "--subsampling-stride",
+            type=int,
+            help="subsampling stride",
+        )
+        parser.add_argument(
+            "--subsampling-norm",
+            type=str,
+            default="none",
+            help="subsampling normalization type",
+        )
+        parser.add_argument(
+            "--subsampling-activation",
+            type=str,
+            default="none",
+            help="subsampling activation function type",
         )
         # Transformer
         parser.add_argument(
@@ -153,6 +128,9 @@ class S2TCTCModel(FairseqEncoderModel):
                 "reduced",
                 "rel_selfattn",
                 "relative",
+                "rel_pos",
+                "rope",
+                "abs"
             ],
             help="transformer encoder self-attention layer type"
         )
@@ -320,6 +298,13 @@ class S2TCTCModel(FairseqEncoderModel):
 
         # Conformer setting
         parser.add_argument(
+            "--encoder-activation-fn",
+            type=str,
+            default="relu",
+            choices=utils.get_available_activation_fns(),
+            help="activation function to use",
+        )
+        parser.add_argument(
             "--macaron-style",
             default=False,
             type=bool,
@@ -390,6 +375,12 @@ class S2TCTCModel(FairseqEncoderModel):
             type=str,
             help="type of intermedia adapter",
         )
+        parser.add_argument(
+            "--intermedia-distribution-cutoff",
+            default=-1,
+            type=int,
+            help="cutoff of the distribution",
+        )
         pass
 
     @classmethod
@@ -422,13 +413,15 @@ class S2TCTCModel(FairseqEncoderModel):
 
     def get_normalized_probs(
             self,
-            net_output: Tuple[Tensor, Optional[Dict[str, List[Optional[Tensor]]]]],
+            net_output,
             log_probs: bool,
             sample: Optional[Dict[str, Tensor]] = None,
     ):
         # net_output['encoder_out'] is a (T, B, D) tensor
-        logits = net_output["ctc_logit"][0]
-        # logits = logits.transpose(0, 1)
+        if isinstance(net_output, list):
+            logits = net_output[0]
+        else:
+            logits = net_output["ctc_logit"][0]
         if log_probs:
             return utils.log_softmax(logits.float(), dim=-1)
         else:
@@ -461,17 +454,20 @@ class S2TCTCEncoder(FairseqEncoder):
             self.embed_scale = 1.0
         self.padding_idx = 1
 
-        self.subsample = Conv1dSubsampler(
-            args.input_feat_per_channel * args.input_channels,
-            args.conv_channels,
-            dim,
-            [int(k) for k in args.conv_kernel_sizes.split(",")],
-        )
+        self.subsample = subsampling(args)
 
         self.attn_type = getattr(args, "encoder_attention_type", "selfattn")
-        self.embed_positions = PositionalEmbedding(
-            args.max_source_positions, dim, self.padding_idx
-        )
+
+        if self.attn_type == "rel_pos":
+            self.embed_positions = RelPositionalEncoding(
+                args.max_source_positions, args.encoder_embed_dim
+            )
+        elif self.attn_type == "rope":
+            self.embed_positions = None
+        else:  # Use absolute positional embedding
+            self.embed_positions = PositionalEmbedding(
+                args.max_source_positions, args.encoder_embed_dim, self.padding_idx
+            )
 
         self.layers = nn.ModuleList(
             [S2TTransformerEncoderLayer(args) for _ in range(args.encoder_layers)]
@@ -513,6 +509,8 @@ class S2TCTCEncoder(FairseqEncoder):
             strategy = None
             if args.intermedia_adapter == "shrink":
                 strategy = getattr(args, "ctc_compress_strategy", None)
+            elif args.intermedia_adapter == "league":
+                strategy = getattr(args, "intermedia_distribution_cutoff", -1)
             self.adapter = InterAdapter(dim, args.intermedia_adapter,
                                         task.source_dictionary, strategy=strategy)
 
@@ -547,26 +545,28 @@ class S2TCTCEncoder(FairseqEncoder):
 
         # down-sampling
         x, input_lengths = self.subsample(src_tokens, src_lengths)
-
-        if type(x) == list:
-            inner_x = x
-            # gather cosine similarity
-            if self.gather_cos_sim:
-                for x in inner_x:
-                    cos_sim_idx += 1
-                    self.add_to_dict(x, dis, cos_sim_idx)
-            x = inner_x[-1]
+        # (B, T, D) -> (T, B, D)
+        x = x.transpose(0, 1)
 
         # embedding scaling
         x = self.embed_scale * x
 
         # padding and position embedding
         encoder_padding_mask = lengths_to_padding_mask(input_lengths)
-        positions = self.embed_positions(encoder_padding_mask).transpose(0, 1)
-        if self.attn_type != "rel_selfattn":
+
+        if self.attn_type == "rel_pos":
+            positions = self.embed_positions(x)
+
+        elif self.attn_type == "rope":
+            positions = None
+
+        else:
+            positions = self.embed_positions(encoder_padding_mask).transpose(0, 1)
             x += positions
+            positions = None
+
         x = self.dropout_module(x)
-        positions = self.dropout_module(positions)
+        # positions = self.dropout_module(positions)
 
         # add emb into history
         if self.history is not None:
@@ -723,16 +723,7 @@ class CTCDecoder(object):
         ctc_logit = encoder_outs["ctc_logit"][0].transpose(0, 1)
         beam_results, beam_scores, timesteps, out_lens = self.ctc_decoder.decode(F.softmax(ctc_logit, -1), src_lengths)
 
-        # beam_results = beam_results[:, :, :out_lens.max()]
-        # for beam_idx in range(beam_size):
-        #     top_beam_tokens = beam_results[:, beam_idx, :]
-        #     top_beam_len = out_lens[:, beam_idx]
-        #     mask = torch.arange(0, top_beam_tokens.size(1)).type_as(top_beam_len). \
-        #         repeat(top_beam_len.size(0), 1).lt(top_beam_len.unsqueeze(1))
-        #     top_beam_tokens[~mask] = self.pad
-
         finalized = []
-
         for idx in range(bsz):
             hypos = []
             for beam_idx in range(beam_size):
@@ -752,8 +743,13 @@ class CTCDecoder(object):
 @register_model_architecture(model_name="s2t_ctc", arch_name="s2t_ctc")
 def base_architecture(args):
     # Convolutional subsampler
-    args.conv_kernel_sizes = getattr(args, "conv_kernel_sizes", "5,5")
-    args.conv_channels = getattr(args, "conv_channels", 1024)
+    args.subsampling_type = getattr(args, "subsampling_type", "conv1d")
+    args.subsampling_layers = getattr(args, "subsampling_layers", 2)
+    args.subsampling_filter = getattr(args, "subsampling_filter", 1024)
+    args.subsampling_kernel = getattr(args, "subsampling_kernel", 5)
+    args.subsampling_stride = getattr(args, "subsampling_stride", 2)
+    args.subsampling_norm = getattr(args, "subsampling_norm", "none")
+    args.subsampling_activation = getattr(args, "subsampling_activation", "glu")
 
     # Transformer
     args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512)

fairseq/models/speech_to_text/s2t_transformer.py

@@ -19,6 +19,7 @@ from fairseq.modules import (
     FairseqDropout,
     LayerNorm,
     PositionalEmbedding,
+    RelPositionalEncoding,
     S2TTransformerEncoderLayer,
     DynamicLinearCombination,
 )
@@ -132,6 +133,9 @@ class S2TTransformerModel(FairseqEncoderDecoderModel):
                 "reduced",
                 "rel_selfattn",
                 "relative",
+                "rel_pos",
+                "rope",
+                "abs"
             ],
             help="transformer encoder self-attention layer type"
         )
@@ -299,6 +303,13 @@ class S2TTransformerModel(FairseqEncoderDecoderModel):
 
         # Conformer setting
         parser.add_argument(
+            "--encoder-activation-fn",
+            type=str,
+            default="relu",
+            choices=utils.get_available_activation_fns(),
+            help="activation function to use",
+        )
+        parser.add_argument(
             "--macaron-style",
             default=False,
             type=bool,
@@ -369,6 +380,12 @@ class S2TTransformerModel(FairseqEncoderDecoderModel):
             type=str,
             help="type of intermedia adapter",
         )
+        parser.add_argument(
+            "--intermedia-distribution-cutoff",
+            default=-1,
+            type=int,
+            help="cutoff of the distribution",
+        )
         pass
 
     @classmethod
@@ -477,9 +494,17 @@ class S2TTransformerEncoder(FairseqEncoder):
         self.subsample = subsampling(args)
 
         self.attn_type = getattr(args, "encoder_attention_type", "selfattn")
-        self.embed_positions = PositionalEmbedding(
-            args.max_source_positions, dim, self.padding_idx
-        )
+
+        if self.attn_type == "rel_pos":
+            self.embed_positions = RelPositionalEncoding(
+                args.max_source_positions, args.encoder_embed_dim
+            )
+        elif self.attn_type == "rope":
+            self.embed_positions = None
+        else:  # Use absolute positional embedding
+            self.embed_positions = PositionalEmbedding(
+                args.max_source_positions, args.encoder_embed_dim, self.padding_idx
+            )
 
         self.layers = nn.ModuleList(
             [S2TTransformerEncoderLayer(args) for _ in range(args.encoder_layers)]
@@ -540,6 +565,8 @@ class S2TTransformerEncoder(FairseqEncoder):
             strategy = None
             if args.intermedia_adapter == "shrink":
                 strategy = getattr(args, "ctc_compress_strategy", None)
+            elif args.intermedia_adapter == "league":
+                strategy = getattr(args, "intermedia_distribution_cutoff", -1)
             self.adapter = InterAdapter(dim, args.intermedia_adapter,
                                         task.source_dictionary, strategy=strategy)
 
@@ -586,11 +613,20 @@ class S2TTransformerEncoder(FairseqEncoder):
 
         # padding and position embedding
         encoder_padding_mask = lengths_to_padding_mask(input_lengths)
-        positions = self.embed_positions(encoder_padding_mask).transpose(0, 1)
-        if self.attn_type != "rel_selfattn":
+
+        if self.attn_type == "rel_pos":
+            positions = self.embed_positions(x)
+
+        elif self.attn_type == "rope":
+            positions = None
+
+        else:
+            positions = self.embed_positions(encoder_padding_mask).transpose(0, 1)
             x += positions
+            positions = None
+
         x = self.dropout_module(x)
-        positions = self.dropout_module(positions)
+        # positions = self.dropout_module(positions)
 
         # add emb into history
         if self.history is not None:
@@ -742,8 +778,13 @@ class TransformerDecoderScriptable(TransformerDecoder):
 @register_model_architecture(model_name="s2t_transformer", arch_name="s2t_transformer")
 def base_architecture(args):
     # Convolutional subsampler
-    args.conv_kernel_sizes = getattr(args, "conv_kernel_sizes", "5,5")
-    args.conv_channels = getattr(args, "conv_channels", 1024)
+    args.subsampling_type = getattr(args, "subsampling_type", "conv1d")
+    args.subsampling_layers = getattr(args, "subsampling_layers", 2)
+    args.subsampling_filter = getattr(args, "subsampling_filter", 1024)
+    args.subsampling_kernel = getattr(args, "subsampling_kernel", 5)
+    args.subsampling_stride = getattr(args, "subsampling_stride", 2)
+    args.subsampling_norm = getattr(args, "subsampling_norm", "none")
+    args.subsampling_activation = getattr(args, "subsampling_activation", "glu")
 
     # Transformer
     args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512)
@@ -791,6 +832,7 @@ def base_architecture(args):
     args.ctc_layer = getattr(args, "ctc_layer", 0)
 
     # Conformer
+    args.encoder_activation_fn = getattr(args, "encoder_activation_fn", "relu")
     args.macaron_style = getattr(args, "macaron_style", False)
     args.use_cnn_module = getattr(args, "use_cnn_module", False)
     args.cnn_module_kernel = getattr(args, "cnn_module_kernel", 31)

fairseq/modules/__init__.py

@@ -10,7 +10,6 @@ from .adaptive_input import AdaptiveInput
 from .adaptive_softmax import AdaptiveSoftmax
 from .beamable_mm import BeamableMM
 from .character_token_embedder import CharacterTokenEmbedder
-from .convolution import ConvolutionModule
 from .downsample_convolution import DownSampleConvolutionModule
 from .conv_tbc import ConvTBC
 from .cross_entropy import cross_entropy
@@ -30,6 +29,7 @@ from .learned_positional_embedding import LearnedPositionalEmbedding
 from .lightweight_convolution import LightweightConv, LightweightConv1dTBC
 from .linearized_convolution import LinearizedConvolution
 from .local_multihead_attention import LocalMultiheadAttention
+from .location_attention import LocationAttention
 from .multihead_attention import MultiheadAttention
 from .positional_embedding import PositionalEmbedding
 from .reduced_multihead_attention import ReducedMultiheadAttention
@@ -44,6 +44,16 @@ from .transpose_last import TransposeLast
 from .unfold import unfold1d
 from .transformer_layer import TransformerDecoderLayer, TransformerEncoderLayer
 from .vggblock import VGGBlock
+from .espnet_multihead_attention import (
+    ESPNETMultiHeadedAttention,
+    RelPositionMultiHeadedAttention,
+    RotaryPositionMultiHeadedAttention,
+)
+from .rotary_positional_embedding import RotaryPositionalEmbedding
+from .positional_encoding import (
+    RelPositionalEncoding,
+)
+from .convolution import ConvolutionModule
 from .s2t_transformer_layer import S2TTransformerEncoderLayer
 from .pds_layer import PDSTransformerEncoderLayer
 
@@ -77,6 +87,7 @@ __all__ = [
     "LightweightConv",
     "LinearizedConvolution",
     "LocalMultiheadAttention",
+
     "MultiheadAttention",
     "PositionalEmbedding",
     "PDSTransformerEncoderLayer",
@@ -96,4 +107,10 @@ __all__ = [
     "TransposeLast",
     "VGGBlock",
     "unfold1d",
+    "ESPNETMultiheadedAttention",
+    "PositionalEmbedding",
+    "RelPositionMultiHeadedAttention",
+    "RelPositionalEncoding",
+    "RotaryPositionalEmbedding",
+    "RotaryPositionMultiHeadedAttention",
 ]

fairseq/modules/activations.py

 import torch
 import torch.nn as nn
+import torch.nn.functional as F
+from typing import Callable
+
+
+def get_activation_fn(activation: str) -> Callable:
+    """ Returns the activation function corresponding to `activation` """
+    from fairseq.modules import gelu, gelu_accurate
+
+    if activation == "relu":
+        return F.relu
+    elif activation == "gelu":
+        return gelu
+    elif activation == "gelu_fast":
+        return gelu_accurate
+    elif activation == "gelu_accurate":
+        return gelu_accurate
+    elif activation == "tanh":
+        return torch.tanh
+    elif activation == "linear":
+        return lambda x: x
+    elif activation == "swish":
+        return torch.nn.SiLU
+    else:
+        raise RuntimeError("--activation-fn {} not supported".format(activation))
 
 
 def get_activation_class(activation: str, dim=None):

fairseq/modules/attention.py

+# Copyright 2021, Maxime Burchi.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+###############################################################################
+# Multi-Head Attention Layers
+###############################################################################
+
+
+class MultiHeadAttention(nn.Module):
+    """Multi-Head Attention Layer
+
+    Args:
+        dim_model: model feature dimension
+        num_heads: number of attention heads
+
+    References:
+        Attention Is All You Need, Vaswani et al.
+        https://arxiv.org/abs/1706.03762
+
+    """
+
+    def __init__(self, dim_model, num_heads):
+        super(MultiHeadAttention, self).__init__()
+
+        # Attention Params
+        self.num_heads = num_heads  # H
+        self.dim_model = dim_model  # D
+        self.dim_head = dim_model // num_heads  # d
+
+        # Linear Layers
+        self.query_layer = nn.Linear(self.dim_model, self.dim_model)
+        self.key_layer = nn.Linear(self.dim_model, self.dim_model)
+        self.value_layer = nn.Linear(self.dim_model, self.dim_model)
+        self.output_layer = nn.Linear(self.dim_model, self.dim_model)
+
+    def forward(self, query, key, value, mask=None):
+
+        """Scaled Dot-Product Multi-Head Attention
+
+        Args:
+            query: Query of shape (B, T, D)
+            key: Key of shape (B, T, D)
+            value: Value of shape (B, T, D)
+            mask: Optional position mask of shape (1 or B, 1 or H, 1 or T, 1 or T)
+
+        Return:
+            out: Attention output of shape (B, T, D)
+            att_w: Attention weights of shape (B, H, T, T)
+
+        """
+
+        # Batch size B
+        batch_size = query.size(0)
+
+        # Linear Layers
+        query = self.query_layer(query)
+        key = self.key_layer(key)
+        value = self.value_layer(value)
+
+        # Reshape and Transpose (B, T, D) -> (B, H, T, d)
+        query = query.reshape(batch_size, -1, self.num_heads, self.dim_head).transpose(1, 2)
+        key = key.reshape(batch_size, -1, self.num_heads, self.dim_head).transpose(1, 2)
+        value = value.reshape(batch_size, -1, self.num_heads, self.dim_head).transpose(1, 2)
+
+        # Att scores (B, H, T, T)
+        att_scores = query.matmul(key.transpose(2, 3)) / key.shape[-1] ** 0.5
+
+        # Apply mask
+        if mask is not None:
+            att_scores += (mask * -1e9)
+
+        # Att weights (B, H, T, T)
+        att_w = att_scores.softmax(dim=-1)
+
+        # Att output (B, H, T, d)
+        out = att_w.matmul(value)
+
+        # Transpose and Reshape (B, H, T, d) -> (B, T, D)
+        out = out.transpose(1, 2).reshape(batch_size, -1, self.dim_model)
+
+        # Output linear layer
+        out = self.output_layer(out)
+
+        return out, att_w.detach()
+
+    def pad(self, query, key, value, mask, chunk_size):
+
+        # Compute Overflows
+        overflow_Q = query.size(1) % chunk_size
+        overflow_KV = key.size(1) % chunk_size
+
+        padding_Q = chunk_size - overflow_Q if overflow_Q else 0
+        padding_KV = chunk_size - overflow_KV if overflow_KV else 0
+
+        batch_size, seq_len_KV, _ = key.size()
+
+        # Input Padding (B, T, D) -> (B, T + P, D)
+        query = F.pad(query, (0, 0, 0, padding_Q), value=0)
+        key = F.pad(key, (0, 0, 0, padding_KV), value=0)
+        value = F.pad(value, (0, 0, 0, padding_KV), value=0)
+
+        # Update Padding Mask
+        if mask is not None:
+
+            # (B, 1, 1, T) -> (B, 1, 1, T + P)
+            if mask.size(2) == 1:
+                mask = F.pad(mask, pad=(0, padding_KV), value=1)
+            # (B, 1, T, T) -> (B, 1, T + P, T + P)
+            else:
+                mask = F.pad(mask, pad=(0, padding_Q, 0, padding_KV), value=1)
+
+        elif padding_KV:
+
+            # None -> (B, 1, 1, T + P)
+            mask = F.pad(query.new_zeros(batch_size, 1, 1, seq_len_KV), pad=(0, padding_KV), value=1)
+
+        return query, key, value, mask, padding_Q
+
+
+class GroupedMultiHeadAttention(MultiHeadAttention):
+    """Grouped Multi-Head Attention Layer
+
+    Grouped multi-head attention reduces attention complexity from out(T2·D) to out(T2·D/G)
+    by grouping neighbouring time elements along the feature dimension before applying
+    scaled dot-product attention.
+
+    Args:
+        dim_model: model feature dimension
+        num_heads: number of attention heads
+        group_size: attention group size
+
+    """
+
+    def __init__(self, dim_model, num_heads, group_size):
+        super(GroupedMultiHeadAttention, self).__init__(dim_model, num_heads)
+
+        # Attention Params
+        self.group_size = group_size  # G
+        self.dim_head = (self.group_size * dim_model) // self.num_heads  # d
+
+    def forward(self, query, key, value, mask=None):
+        # Batch size B
+        batch_size = query.size(0)
+
+        # Linear Layers
+        query = self.query_layer(query)
+        key = self.key_layer(key)
+        value = self.value_layer(value)
+
+        # Chunk Padding
+        query, key, value, mask, padding = self.pad(query, key, value, mask, chunk_size=self.group_size)
+
+        # Reshape and Transpose (B, T, D) -> (B, H, T//G, d)
+        query = query.reshape(batch_size, -1, self.num_heads, self.dim_head).transpose(1, 2)
+        key = key.reshape(batch_size, -1, self.num_heads, self.dim_head).transpose(1, 2)
+        value = value.reshape(batch_size, -1, self.num_heads, self.dim_head).transpose(1, 2)
+
+        # Att scores (B, H, T//G, T//G)
+        att_scores = query.matmul(key.transpose(2, 3)) / key.shape[-1] ** 0.5
+
+        # Apply mask
+        if mask is not None:
+            # Slice Mask (B, 1, T, T) -> (B, 1, T//G, T//G)
+            mask = mask[:, :, ::self.group_size, ::self.group_size]
+
+            # Apply mask
+            att_scores += (mask * -1e9)
+
+        # Att weights (B, H, T//G, T//G)
+        att_w = att_scores.softmax(dim=-1)
+
+        # Att output (B, H, T//G, d)
+        out = att_w.matmul(value)
+
+        # Transpose and Reshape (B, H, T//G, d) -> (B, T, D)
+        out = out.transpose(1, 2).reshape(batch_size, -1, self.dim_model)
+
+        # Slice Padding
+        out = out[:, :out.size(1) - padding]
+
+        # Output linear layer
+        out = self.output_layer(out)
+
+        return out, att_w.detach()
+
+
+class LocalMultiHeadAttention(MultiHeadAttention):
+    """Local Multi-Head Attention Layer
+
+    Local multi-head attention restricts the attended positions to a local neighborhood
+    around the query position. This is achieved by segmenting the hidden sequence into
+    non overlapping blocks of size key and performing scaled dot-product attention in
+    parallel for each of these blocks.
+
+    Args:
+        dim_model: model feature dimension
+        num_heads: number of attention heads
+        kernel_size: attention kernel size / window
+
+    References:
+        Image Transformer, Parmar et al.
+        https://arxiv.org/abs/1802.05751
+
+    """
+
+    def __init__(self, dim_model, num_heads, kernel_size):
+        super(LocalMultiHeadAttention, self).__init__(dim_model, num_heads)
+
+        # Attention Params
+        self.kernel_size = kernel_size  # key
+
+    def forward(self, query, key, value, mask=None):
+
+        # Batch size B
+        batch_size = query.size(0)
+
+        # Linear Layers
+        query = self.query_layer(query)
+        key = self.key_layer(key)
+        value = self.value_layer(value)
+
+        # Chunk Padding
+        query, key, value, mask, padding = self.pad(query, key, value, mask, chunk_size=self.kernel_size)
+
+        # Reshape and Transpose (B, T, D) -> (B, T//key, H, key, d)
+        query = query.reshape(batch_size, -1, self.kernel_size, self.num_heads, self.dim_head).transpose(2, 3)
+        key = key.reshape(batch_size, -1, self.kernel_size, self.num_heads, self.dim_head).transpose(2, 3)
+        value = value.reshape(batch_size, -1, self.kernel_size, self.num_heads, self.dim_head).transpose(2, 3)
+
+        # Att scores (B, T//key, H, key, key)
+        att_scores = query.matmul(key.transpose(3, 4)) / key.shape[-1] ** 0.5
+
+        # Apply mask
+        if mask is not None:
+
+            # Slice mask (B, 1, T, T) -> (B, T//key, 1, key, key)
+            masks = []
+            for m in range(mask.size(-1) // self.kernel_size):
+                masks.append(mask[:, :, m * self.kernel_size: (m + 1) * self.kernel_size,
+                             m * self.kernel_size: (m + 1) * self.kernel_size])
+            mask = torch.stack(masks, dim=1)
+
+            # Apply mask
+            att_scores = att_scores.float() - mask.float() * 1e9
+
+        # Att weights (B, T//key, H, key, key)
+        att_w = att_scores.softmax(dim=-1)
+
+        # Att output (B, T//key, H, key, d)
+        out = att_w.matmul(value)
+
+        # Transpose and Reshape (B, T//key, H, key, d) -> (B, T, D)
+        out = out.transpose(2, 3).reshape(batch_size, -1, self.dim_model)
+
+        # Slice Padding
+        out = out[:, :out.size(1) - padding]
+
+        # Output linear layer
+        out = self.output_layer(out)
+
+        return out, att_w.detach()
+
+
+class StridedMultiHeadAttention(MultiHeadAttention):
+    """Strided Multi-Head Attention Layer
+
+    Strided multi-head attention performs global sequence downsampling by striding
+    the attention query before aplying scaled dot-product attention. This results in
+    strided attention maps where query positions can attend to the entire sequence
+    context to perform downsampling.
+
+    Args:
+        dim_model: model feature dimension
+        num_heads: number of attention heads
+        stride: query stride
+
+    """
+
+    def __init__(self, dim_model, num_heads, stride):
+        super(StridedMultiHeadAttention, self).__init__(dim_model, num_heads)
+
+        # Attention Params
+        self.stride = stride  # S
+
+    def forward(self, query, key, value, mask=None):
+        # Query Subsampling (B, T, D) -> (B, T//S, D)
+        query = query[:, ::self.stride]
+
+        # Mask Subsampling (B, 1, T, T) -> (B, 1, T//S, T)
+        if mask is not None:
+            mask = mask[:, :, ::self.stride]
+
+        # Multi-Head Attention
+        return super(StridedMultiHeadAttention, self).forward(query, key, value, mask)
+
+
+class StridedLocalMultiHeadAttention(MultiHeadAttention):
+    """Strided Local Multi-Head Attention Layer
+
+    Args:
+        dim_model: model feature dimension
+        num_heads: number of attention heads
+        kernel_size: attention kernel size / window
+        stride: query stride
+
+    """
+
+    def __init__(self, dim_model, num_heads, kernel_size, stride):
+        super(StridedLocalMultiHeadAttention, self).__init__(dim_model, num_heads)
+
+        # Assert
+        assert kernel_size % stride == 0, "Attention kernel size has to be a multiple of attention stride"
+
+        # Attention Params
+        self.kernel_size = kernel_size  # key
+        self.stride = stride  # S
+
+    def forward(self, query, key, value, mask=None):
+
+        # Batch size B
+        batch_size = query.size(0)
+
+        # Query Subsampling (B, T, D) -> (B, T//S, D)
+        query = query[:, ::self.stride]
+
+        # Linear Layers
+        query = self.query_layer(query)
+        key = self.key_layer(key)
+        value = self.value_layer(value)
+
+        # Chunk Padding
+        query, key, value, mask, padding = self.pad(query, key, value, mask, chunk_size=self.kernel_size)
+
+        # Reshape and Transpose (B, T//S, D) -> (B, T//key, H, key//S, d)
+        query = query.reshape(batch_size, -1, self.kernel_size // self.stride, self.num_heads, self.dim_head).transpose(2, 3)
+        # Reshape and Transpose (B, T, D) -> (B, T//key, H, key, d)
+        key = key.reshape(batch_size, -1, self.kernel_size, self.num_heads, self.dim_head).transpose(2, 3)
+        value = value.reshape(batch_size, -1, self.kernel_size, self.num_heads, self.dim_head).transpose(2, 3)
+
+        # Att scores (B, T//key, H, key//S, key)
+        att_scores = query.matmul(key.transpose(3, 4)) / key.shape[-1] ** 0.5
+
+        # Apply mask
+        if mask is not None:
+
+            # Slice mask (B, 1, T, T) -> (B, T//key, 1, key, key)
+            masks = []
+            for m in range(mask.size(-1) // self.kernel_size):
+                masks.append(mask[:, :, m * self.kernel_size: (m + 1) * self.kernel_size,
+                             m * self.kernel_size: (m + 1) * self.kernel_size])
+            mask = torch.stack(masks, dim=1)
+
+            # Subsample mask (B, T//key, 1, key, key) -> (B, T//key, 1, key//S, key)
+            mask = mask[:, :, :, ::self.stride]
+
+            # Apply mask
+            att_scores = att_scores.float() - mask.float() * 1e9
+
+        # Att weights (B, T//key, H, key//S, key)
+        att_w = att_scores.softmax(dim=-1)
+
+        # Att output (B, T//key, H, key//S, d)
+        out = att_w.matmul(value)
+
+        # Transpose and Reshape (B, T//key, H, key//S, d) -> (B, T//S, D)
+        out = out.transpose(2, 3).reshape(batch_size, -1, self.dim_model)
+
+        # Slice Padding
+        out = out[:, :(out.size(1) - padding - 1) // self.stride + 1]
+
+        # Output linear layer
+        out = self.output_layer(out)
+
+        return out, att_w.detach()
+
+
+class MultiHeadLinearAttention(MultiHeadAttention):
+    """Multi-Head Linear Attention
+
+    Args:
+        dim_model: model feature dimension
+        num_heads: number of attention heads
+
+    References:
+        Efficient Attention: Attention with Linear Complexities, Shen et al.
+        https://arxiv.org/abs/1812.01243
+
+        Efficient conformer-based speech recognition with linear attention, Li et al.
+        https://arxiv.org/abs/2104.06865
+
+    """
+
+    def __init__(self, dim_model, num_heads):
+        super(MultiHeadLinearAttention, self).__init__(dim_model, num_heads)
+
+    def forward(self, query, key, value):
+        # Batch size B
+        batch_size = query.size(0)
+
+        # Linear Layers
+        query = self.query_layer(query)
+        key = self.key_layer(key)
+        value = self.value_layer(value)
+
+        # Reshape and Transpose (B, T, D) -> (B, N, T, d)
+        query = query.reshape(batch_size, -1, self.num_heads, self.dim_head).transpose(1, 2)
+        key = key.reshape(batch_size, -1, self.num_heads, self.dim_head).transpose(1, 2)
+        value = value.reshape(batch_size, -1, self.num_heads, self.dim_head).transpose(1, 2)
+
+        # Global Context Vector (B, N, d, d)
+        KV = (key / key.shape[-1] ** (1.0 / 4.0)).softmax(dim=-2).transpose(2, 3).matmul(value)
+
+        # Attention Output (B, N, T, d)
+        out = (query / query.shape[-1] ** (1.0 / 4.0)).softmax(dim=-1).matmul(KV)
+
+        # Transpose and Reshape (B, N, T, d) -> (B, T, D)
+        out = out.transpose(1, 2).reshape(batch_size, -1, self.dim_model)
+
+        # Output linear layer
+        out = self.output_layer(out)
+
+        return out, KV.detach()
+
+
+###############################################################################
+# Multi-Head Self-Attention Layers with Relative Sinusoidal Poditional Encodings
+###############################################################################
+
+class RelPosMultiHeadSelfAttention(MultiHeadAttention):
+    """Multi-Head Self-Attention Layer with Relative Sinusoidal Positional Encodings
+
+    Args:
+        dim_model: model feature dimension
+        num_heads: number of attention heads
+        causal: whether the attention is causal or unmasked
+        max_pos_encoding: maximum relative distance between elements
+
+    References:
+        Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context, Dai et al.
+        https://arxiv.org/abs/1901.02860
+
+    """
+
+    def __init__(self, dim_model, num_heads, causal, max_pos_encoding):
+        super(RelPosMultiHeadSelfAttention, self).__init__(dim_model, num_heads)
+
+        # Position Embedding Layer
+        self.pos_layer = nn.Linear(self.dim_model, self.dim_model)
+        self.causal = causal
+
+        # Global content and positional bias
+        self.u = nn.Parameter(torch.Tensor(self.dim_model))  # Content bias
+        self.v = nn.Parameter(torch.Tensor(self.dim_model))  # Pos bias
+        torch.nn.init.xavier_uniform_(self.u.reshape(self.num_heads, self.dim_head))  # glorot uniform
+        torch.nn.init.xavier_uniform_(self.v.reshape(self.num_heads, self.dim_head))  # glorot uniform
+
+        # Relative Sinusoidal Positional Encodings
+        self.rel_pos_enc = RelativeSinusoidalPositionalEncoding(max_pos_encoding, self.dim_model, self.causal)
+
+    def rel_to_abs(self, att_scores):
+
+        """Relative to absolute position indexing
+
+        Args:
+            att_scores: absolute-by-relative indexed attention scores of shape
+            (B, H, T, Th + 2*T-1) for full context and (B, H, T, Th + T) for causal context
+
+        Return:
+            att_scores: absolute-by-absolute indexed attention scores of shape (B, H, T, Th + T)
+
+        References:
+            causal context:
+            Music Transformer, Huang et al.
+            https://arxiv.org/abs/1809.04281
+
+            full context:
+            Attention Augmented Convolutional Networks, Bello et al.
+            https://arxiv.org/abs/1904.09925
+
+        """
+
+        # Causal Context
+        if self.causal:
+
+            # Att Scores (B, H, T, Th + T)
+            batch_size, num_heads, seq_length1, seq_length2 = att_scores.size()
+
+            # Column Padding (B, H, T, 1 + Th + T)
+            att_scores = F.pad(att_scores, pad=(1, 0), value=0)
+
+            # Flatten (B, H, T + TTh + TT)
+            att_scores = att_scores.reshape(batch_size, num_heads, -1)
+
+            # Start Padding (B, H, Th + T + TTh + TT)
+            att_scores = F.pad(att_scores, pad=(seq_length2 - seq_length1, 0), value=0)
+
+            # Reshape (B, H, 1 + T, Th + T)
+            att_scores = att_scores.reshape(batch_size, num_heads, 1 + seq_length1, seq_length2)
+
+            # Slice (B, H, T, Th + T)
+            att_scores = att_scores[:, :, 1:]
+
+        # Full Context
+        else:
+
+            # Att Scores (B, H, T, Th + 2*T-1)
+            batch_size, num_heads, seq_length1, seq_length2 = att_scores.size()
+
+            # Column Padding (B, H, T, Th + 2*T)
+            att_scores = F.pad(att_scores, pad=(0, 1), value=0)
+
+            # Flatten (B, H, TTh + 2*TT)
+            att_scores = att_scores.reshape(batch_size, num_heads, -1)
+
+            # End Padding (B, H, TTh + 2*TT + Th + T - 1)
+            att_scores = F.pad(att_scores, pad=(0, seq_length2 - seq_length1), value=0)
+
+            # Reshape (B, H, T + 1, Th + 2*T-1)
+            att_scores = att_scores.reshape(batch_size, num_heads, 1 + seq_length1, seq_length2)
+
+            # Slice (B, H, T, Th + T)
+            att_scores = att_scores[:, :, :seq_length1, seq_length1 - 1:]
+
+        return att_scores
+
+    def forward(self, query, key, value, mask=None, hidden=None):
+
+        """Scaled Dot-Product Self-Attention with relative sinusoidal position encodings
+
+        Args:
+            query: Query of shape (B, T, D)
+            key: Key of shape (B, T, D)
+            value: Value of shape (B, T, D)
+            mask: Optional position mask of shape (1 or B, 1 or H, 1 or T, 1 or T)
+            hidden: Optional Key and Value hidden states for decoding
+
+        Return:
+            out: Attention output of shape (B, T, D)
+            att_w: Attention weights of shape (B, H, T, Th + T)
+            hidden: Key and value hidden states
+
+        """
+
+        # Batch size B
+        batch_size = query.size(0)
+
+        # Linear Layers
+        query = self.query_layer(query)
+        key = self.key_layer(key)
+        value = self.value_layer(value)
+
+        # Hidden State Provided
+        if hidden:
+            key = torch.cat([hidden["key"], key], dim=1)
+            value = torch.cat([hidden["value"], value], dim=1)
+
+        # Update Hidden State
+        hidden = {"key": key.detach(), "value": value.detach()}
+
+        # Add Bias
+        Qu = query + self.u
+        Qv = query + self.v
+
+        # Relative Positional Embeddings (B, Th + 2*T-1, D) / (B, Th + T, D)
+        E = self.pos_layer(self.rel_pos_enc(batch_size, query.size(1), key.size(1) - query.size(1)))
+
+        # Reshape and Transpose (B, T, D) -> (B, H, T, d)
+        Qu = Qu.reshape(batch_size, -1, self.num_heads, self.dim_head).transpose(1, 2)
+        Qv = Qv.reshape(batch_size, -1, self.num_heads, self.dim_head).transpose(1, 2)
+        # Reshape and Transpose (B, Th + T, D) -> (B, H, Th + T, d)
+        key = key.reshape(batch_size, -1, self.num_heads, self.dim_head).transpose(1, 2)
+        value = value.reshape(batch_size, -1, self.num_heads, self.dim_head).transpose(1, 2)
+        # Reshape and Transpose (B, Th + 2*T-1, D) -> (B, H, Th + 2*T-1, d) / (B, Th + T, D) -> (B, H, Th + T, d)
+        E = E.reshape(batch_size, -1, self.num_heads, self.dim_head).transpose(1, 2)
+
+        # att_scores (B, H, T, Th + T)
+        att_scores_K = Qu.matmul(key.transpose(2, 3))
+        att_scores_E = self.rel_to_abs(Qv.matmul(E.transpose(2, 3)))
+        att_scores = (att_scores_K + att_scores_E) / key.shape[-1] ** 0.5
+
+        # Apply mask
+        if mask is not None:
+            att_scores += (mask * -1e9)
+
+        # Att weights (B, H, T, Th + T)
+        att_w = att_scores.softmax(dim=-1)
+
+        # Att output (B, H, T, d)
+        out = att_w.matmul(value)
+
+        # Transpose and Reshape (B, H, T, d) -> (B, T, D)
+        out = out.transpose(1, 2).reshape(batch_size, -1, self.dim_model)
+
+        # Output linear layer
+        out = self.output_layer(out)
+
+        return out, att_w.detach(), hidden
+
+
+class GroupedRelPosMultiHeadSelfAttention(RelPosMultiHeadSelfAttention):
+    """Grouped Multi-Head Self-Attention Layer with Relative Sinusoidal Positional Encodings
+
+    Args:
+        dim_model: model feature dimension
+        num_heads: number of attention heads
+        causal: whether the attention is causal or unmasked
+        max_pos_encoding: maximum relative distance between elements
+        group_size: attention group size
+
+    """
+
+    def __init__(self, dim_model, num_heads, causal, max_pos_encoding, group_size):
+        super(GroupedRelPosMultiHeadSelfAttention, self).__init__(dim_model, num_heads, causal, max_pos_encoding)
+
+        # Attention Params
+        self.group_size = group_size  # G
+        self.dim_head = (self.group_size * dim_model) // self.num_heads  # d
+
+        # Grouped Relative Sinusoidal Positional Encodings
+        self.rel_pos_enc = GroupedRelativeSinusoidalPositionalEncoding(max_pos_encoding, self.dim_model,
+                                                                       self.group_size, self.causal)
+
+    def forward(self, query, key, value, mask=None, hidden=None):
+
+        # Batch size B
+        batch_size = query.size(0)
+
+        # Linear Layers
+        query = self.query_layer(query)
+        key = self.key_layer(key)
+        value = self.value_layer(value)
+
+        # Hidden State Provided
+        if hidden:
+            Kh = torch.cat([hidden["key"], key], dim=1)
+            Vh = torch.cat([hidden["value"], value], dim=1)
+            key = torch.cat([hidden["key"][:, hidden["key"].size(1) % self.group_size:], key], dim=1)
+            value = torch.cat([hidden["value"][:, hidden["value"].size(1) % self.group_size:], value], dim=1)
+
+            # Update Hidden State
+            hidden = {"key": Kh.detach(), "value": Vh.detach()}
+
+        else:
+
+            # Update Hidden State
+            hidden = {"key": key.detach(), "value": value.detach()}
+
+        # Chunk Padding
+        query, key, value, mask, padding = self.pad(query, key, value, mask, chunk_size=self.group_size)
+
+        # Add Bias
+        Qu = query + self.u
+        Qv = query + self.v
+
+        # Relative Positional Embeddings (B, Th + 2*T-G, D) / (B, Th + T, D)
+        E = self.pos_layer(self.rel_pos_enc(batch_size, query.size(1), key.size(1) - query.size(1)))
+
+        # Reshape and Transpose (B, T, D) -> (B, H, T//G, d)
+        Qu = Qu.reshape(batch_size, -1, self.num_heads, self.dim_head).transpose(1, 2)
+        Qv = Qv.reshape(batch_size, -1, self.num_heads, self.dim_head).transpose(1, 2)
+        # Reshape and Transpose (B, Th + T, D) -> (B, H, Th//G + T//G, d)
+        key = key.reshape(batch_size, -1, self.num_heads, self.dim_head).transpose(1, 2)
+        value = value.reshape(batch_size, -1, self.num_heads, self.dim_head).transpose(1, 2)
+        # Reshape and Transpose (B, Th + 2*T-G, D) -> (B, H, Th//G + 2*T//G-1, d) / (B, Th + T, D) -> (B, H, Th//G + T//G, d)
+        E = E.reshape(batch_size, -1, self.num_heads, self.dim_head).transpose(1, 2)
+
+        # att_scores (B, H, T//G, Th//G + T//G)
+        att_scores_K = Qu.matmul(key.transpose(2, 3))
+        att_scores_E = self.rel_to_abs(Qv.matmul(E.transpose(2, 3)))
+        att_scores = (att_scores_K + att_scores_E) / key.shape[-1] ** 0.5
+
+        # Apply mask
+        if mask is not None:
+            # Slice Mask (B, 1, T, T) -> (B, 1, T//G, T//G)
+            mask = mask[:, :, ::self.group_size, ::self.group_size]
+
+            # Apply mask
+            att_scores += (mask * -1e9)
+
+        # Att weights (B, H, T//G, Th//G + T//G)
+        att_w = att_scores.softmax(dim=-1)
+
+        # Att output (B, H, T//G, d)
+        out = att_w.matmul(value)
+
+        # Transpose and Reshape (B, H, T//G, d) -> (B, T, D)
+        out = out.transpose(1, 2).reshape(batch_size, -1, self.dim_model)
+
+        # Slice Padding
+        out = out[:, :out.size(1) - padding]
+
+        # Output linear layer
+        out = self.output_layer(out)
+
+        return out, att_w.detach(), hidden
+
+
+class LocalRelPosMultiHeadSelfAttention(RelPosMultiHeadSelfAttention):
+    """Local Multi-Head Self-Attention with Relative Sinusoidal Positional Encodings
+
+    Args:
+        dim_model: model feature dimension
+        num_heads: number of attention heads
+        causal: whether the attention is causal or unmasked
+        kernel_size: attention kernel size / window
+
+    References:
+        Music Transformer, Huang et al.
+        https://arxiv.org/abs/1809.04281
+
+    """
+
+    def __init__(self, dim_model, num_heads, causal, kernel_size):
+        super(LocalRelPosMultiHeadSelfAttention, self).__init__(dim_model, num_heads, causal, kernel_size)
+
+        # Attention Params
+        self.kernel_size = kernel_size  # key
+
+    def rel_to_abs(self, att_scores):
+
+        """Relative to absolute position indexing
+
+        Args:
+            att_scores: absolute-by-relative indexed attention scores of shape
+            (B, N, T, 2 * key - 1) for full context and (B, H, T, key) for causal context
+
+        Return:
+            att_scores: absolute-by-absolute indexed attention scores of shape (B, T//key, H, key, key)
+
+        References:
+            Causal context:
+            Music Transformer, Huang et al.
+            https://arxiv.org/abs/1809.04281
+        """
+
+        # Causal Context
+        if self.causal:
+
+            # Att Scores (B, H, T, key)
+            batch_size, num_heads, seq_length1, seq_length2 = att_scores.size()
+
+            # Reshape (B, T//key, H, key, key)
+            att_scores = att_scores.reshape(batch_size, -1, self.num_heads, self.kernel_size, self.kernel_size)
+
+            # Column Padding (B, T//key, H, key, 1 + key)
+            att_scores = F.pad(att_scores, pad=(1, 0), value=0)
+
+            # Reshape (B, T//key, H, 1 + key, key)
+            att_scores = att_scores.reshape(batch_size, -1, self.num_heads, self.kernel_size + 1, self.kernel_size)
+
+            # Slice (B, T//key, H, key, key)
+            att_scores = att_scores[:, :, :, 1:]
+
+        # Full Context
+        else:
+
+            # Att Scores (B, H, T, 2 * key - 1)
+            batch_size, num_heads, seq_length1, seq_length2 = att_scores.size()
+
+            # Reshape (B, T//key, H, key, 2 * key - 1)
+            att_scores = att_scores.reshape(batch_size, -1, self.num_heads, self.kernel_size, seq_length2)
+
+            # Column Padding (B, T//key, H, key, 2 * key)
+            att_scores = F.pad(att_scores, pad=(0, 1), value=0)
+
+            # Flatten (B, T//key, H, key * 2 * key)
+            att_scores = att_scores.reshape(batch_size, -1, self.num_heads, 2 * self.kernel_size ** 2)
+
+            # End Padding (B, T//key, H, key * 2 * key + key - 1)
+            att_scores = F.pad(att_scores, pad=(0, self.kernel_size - 1), value=0)
+
+            # Reshape (B, T//key, H, key + 1, 2 * key - 1)
+            att_scores = att_scores.reshape(batch_size, -1, self.num_heads, self.kernel_size + 1, seq_length2)
+
+            # Slice (B, T//key, H, key, key)
+            att_scores = att_scores[:, :, :, :self.kernel_size, self.kernel_size - 1:]
+
+        return att_scores
+
+    def forward(self, query, key, value, mask=None, hidden=None):
+
+        # Batch size B
+        batch_size = query.size(0)
+
+        # Linear Layers
+        query = self.query_layer(query)
+        key = self.key_layer(key)
+        value = self.value_layer(value)
+
+        # Chunk Padding
+        query, key, value, mask, padding = self.pad(query, key, value, mask, chunk_size=self.kernel_size)
+
+        # Add Bias
+        Qu = query + self.u
+        Qv = query + self.v
+
+        # Relative Positional Embeddings (B, 2*key-1, D) / (B, key, D)
+        E = self.pos_layer(self.rel_pos_enc(batch_size))
+
+        # Reshape and Transpose (B, T, D) -> (B, H, T, d)
+        Qv = Qv.reshape(batch_size, -1, self.num_heads, self.dim_head).transpose(1, 2)
+        # Reshape and Transpose (B, T, D) -> (B, T//key, H, key, d)
+        Qu = Qu.reshape(batch_size, -1, self.kernel_size, self.num_heads, self.dim_head).transpose(2, 3)
+        key = key.reshape(batch_size, -1, self.kernel_size, self.num_heads, self.dim_head).transpose(2, 3)
+        value = value.reshape(batch_size, -1, self.kernel_size, self.num_heads, self.dim_head).transpose(2, 3)
+        # Reshape and Transpose (B, 2*key-1, D) -> (B, H, 2*key-1, d) / (B, key, D) -> (B, H, key, d)
+        E = E.reshape(batch_size, -1, self.num_heads, self.dim_head).transpose(1, 2)
+
+        # att_scores (B, T//key, H, key, key)
+        att_scores_K = Qu.matmul(key.transpose(3, 4))
+        att_scores_E = self.rel_to_abs(Qv.matmul(E.transpose(2, 3)))
+        att_scores = (att_scores_K + att_scores_E) / key.shape[-1] ** 0.5
+
+        # Mask scores
+        if mask is not None:
+
+            # Diagonal Mask (B, 1, T, T) -> (B, T//key, 1, key, key)
+            masks = []
+            for m in range(mask.size(-1) // self.kernel_size):
+                masks.append(mask[:, :, m * self.kernel_size: (m + 1) * self.kernel_size,
+                             m * self.kernel_size: (m + 1) * self.kernel_size])
+            mask = torch.stack(masks, dim=1)
+
+            # Apply Mask
+            att_scores = att_scores.float() - mask.float() * 1e9
+
+        # Attention weights (B, T//key, H, key, key)
+        att_w = att_scores.softmax(dim=-1)
+
+        # Attention output (B, T//key, H, key, d)
+        out = att_w.matmul(value)
+
+        # Transpose and Reshape (B, T//key, H, key, d) -> (B, T, D)
+        out = out.transpose(2, 3).reshape(batch_size, -1, self.dim_model)
+
+        # Slice Padding
+        out = out[:, :out.size(1) - padding]
+
+        # Output linear layer
+        out = self.output_layer(out)
+
+        return out, att_w.detach(), hidden
+
+
+class StridedRelPosMultiHeadSelfAttention(RelPosMultiHeadSelfAttention):
+    """Strided Multi-Head Self-Attention with Relative Sinusoidal Positional Encodings
+
+    Args:
+        dim_model: model feature dimension
+        num_heads: number of attention heads
+        causal: whether the attention is causal or unmasked
+        max_pos_encoding: maximum relative distance between elements
+        stride: query stride
+    """
+
+    def __init__(self, dim_model, num_heads, causal, max_pos_encoding, stride):
+        super(StridedRelPosMultiHeadSelfAttention, self).__init__(dim_model, num_heads, causal, max_pos_encoding)
+
+        # Attention Params
+        self.stride = stride  # S
+
+    def rel_to_abs(self, att_scores):
+
+        """Relative to absolute position indexing
+
+        Args:
+            att_scores: absolute-by-relative indexed attention scores of shape
+            (B, H, T//S, Th + 2 * T - 1) for full context and (B, H, T//S, Th + T) for causal context
+
+        Return:
+            att_scores: absolute-by-absolute indexed attention scores of shape (B, H, T//S,Th + T)
+
+        """
+
+        # Causal Context
+        if self.causal:
+
+            # Att Scores (B, H, T // S, Th + T)
+            batch_size, num_heads, seq_length1, seq_length2 = att_scores.size()
+
+            # Column Padding (B, H, T // S, Th + T + S)
+            att_scores = F.pad(att_scores, pad=(1, self.stride - 1), value=0)
+
+            # Flatten (B, H, TTh//S + TT//S + T)
+            att_scores = att_scores.reshape(batch_size, num_heads, -1)
+
+            # Start Padding (B, H, TTh//S + TT//S + T + Th)
+            att_scores = F.pad(att_scores, pad=(seq_length2 - self.stride * seq_length1, 0), value=0)
+
+            # Reshape (B, H, 1 + T // S, Th + T)
+            att_scores = att_scores.reshape(batch_size, num_heads, seq_length1 + 1, seq_length2)
+
+            # Slice (B, H, T // S, Th + T)
+            att_scores = att_scores[:, :, 1:]
+
+        # Full Context
+        else:
+
+            # Att Scores (B, H, T // S, Th + 2*T-1)
+            batch_size, num_heads, seq_length1, seq_length2 = att_scores.size()
+
+            # Column Padding (B, H, T // S, Th + 2*T-1 + S)
+            att_scores = F.pad(att_scores, pad=(0, self.stride), value=0)
+
+            # Flatten (B, H, TTh//S + 2*TT//S - T//S + T)
+            att_scores = att_scores.reshape(batch_size, num_heads, -1)
+
+            # End Padding (B, H, TTh//S + 2*TT//S - T//S + Th + 2T-1)
+            att_scores = F.pad(att_scores, pad=(0, seq_length2 - seq_length1 * self.stride), value=0)
+
+            # Reshape (B, H, T//S + 1, Th + 2*T-1)
+            att_scores = att_scores.reshape(batch_size, num_heads, seq_length1 + 1, seq_length2)
+
+            # Slice (B, H, T // S, Th + T)
+            att_scores = att_scores[:, :, :seq_length1, seq_length1 * self.stride - 1:]
+
+        return att_scores
+
+    def forward(self, query, key, value, mask=None, hidden=None):
+
+        # Batch size B
+        batch_size = query.size(0)
+
+        # Linear Layers
+        query = self.query_layer(query)
+        key = self.key_layer(key)
+        value = self.value_layer(value)
+
+        # Hidden State Provided
+        if hidden:
+            key = torch.cat([hidden["key"], key], dim=1)
+            value = torch.cat([hidden["value"], value], dim=1)
+
+        # Update Hidden State
+        hidden = {"key": key.detach(), "value": value.detach()}
+
+        # Chunk Padding
+        query, key, value, mask, _ = self.pad(query, key, value, mask, chunk_size=self.stride)
+
+        # Query Subsampling (B, T, D) -> (B, T//S, D)
+        query = query[:, ::self.stride]
+
+        # Add Bias
+        Qu = query + self.u
+        Qv = query + self.v
+
+        # Relative Positional Embeddings (B, Th + 2*T-1, D) / (B, Th + T, D)
+        E = self.pos_layer(self.rel_pos_enc(batch_size, self.stride * query.size(1), key.size(1) - self.stride * query.size(1)))
+
+        # Reshape and Transpose (B, T//S, D) -> (B, H, T//S, d)
+        Qu = Qu.reshape(batch_size, -1, self.num_heads, self.dim_head).transpose(1, 2)
+        Qv = Qv.reshape(batch_size, -1, self.num_heads, self.dim_head).transpose(1, 2)
+        # Reshape and Transpose (B, Th + T, D) -> (B, H, Th + T, d)
+        key = key.reshape(batch_size, -1, self.num_heads, self.dim_head).transpose(1, 2)
+        value = value.reshape(batch_size, -1, self.num_heads, self.dim_head).transpose(1, 2)
+        # Reshape and Transpose (B, Th + 2*T-1, D) -> (B, H, Th + 2*T-1, d) / (B, Th + T, D) -> (B, H, Th + T, d)
+        E = E.reshape(batch_size, -1, self.num_heads, self.dim_head).transpose(1, 2)
+
+        # att_scores (B, H, T//S, Th + T)
+        att_scores_K = Qu.matmul(key.transpose(2, 3))
+        att_scores_E = self.rel_to_abs(Qv.matmul(E.transpose(2, 3)))
+        att_scores = (att_scores_K + att_scores_E) / key.shape[-1] ** 0.5
+
+        # Apply mask
+        if mask is not None:
+
+            # Mask Subsampling (B, 1, T, T) -> (B, 1, T//S, T)
+            if mask is not None:
+                mask = mask[:, :, ::self.stride]
+
+            # Apply mask
+            att_scores += (mask * -1e9)
+
+        # Att weights (B, H, T//S, Th + T)
+        att_w = att_scores.softmax(dim=-1)
+
+        # Att output (B, H, T//S, d)
+        out = att_w.matmul(value)
+
+        # Transpose and Reshape (B, H, T//S, d) -> (B, T//S, D)
+        out = out.transpose(1, 2).reshape(batch_size, -1, self.dim_model)
+
+        # Output linear layer
+        out = self.output_layer(out)
+
+        return out, att_w.detach(), hidden
+
+
+class StridedLocalRelPosMultiHeadSelfAttention(RelPosMultiHeadSelfAttention):
+    """Strided Local Multi-Head Self-Attention with Relative Sinusoidal Positional Encodings
+
+    Args:
+        dim_model: model feature dimension
+        num_heads: number of attention heads
+        causal: whether the attention is causal or unmasked
+        kernel_size: attention kernel size / window
+        stride: query stride
+    """
+
+    def __init__(self, dim_model, num_heads, causal, kernel_size, stride):
+        super(StridedLocalRelPosMultiHeadSelfAttention, self).__init__(dim_model, num_heads, causal, kernel_size)
+
+        # Assert
+        assert kernel_size % stride == 0, "Attention kernel size has to be a multiple of attention stride"
+
+        # Attention Params
+        self.kernel_size = kernel_size  # key
+        self.stride = stride  # S
+
+    def rel_to_abs(self, att_scores):
+
+        """Relative to absolute position indexing
+
+        Args:
+            att_scores: absolute-by-relative indexed attention scores of shape
+            (B, H, T//S, 2 * key - 1) for full context and (B, H, T//S, key) for causal context
+
+        Return:
+            att_scores: absolute-by-absolute indexed attention scores of shape (B, T//key, H, key//S, key)
+        """
+
+        # Causal Context
+        if self.causal:
+
+            # Att Scores (B, H, T//S, key)
+            batch_size, num_heads, seq_length1, seq_length2 = att_scores.size()
+
+            # Reshape (B, T//key, H, key//S, key)
+            att_scores = att_scores.reshape(batch_size, -1, self.num_heads, self.kernel_size // self.stride,
+                                            self.kernel_size)
+
+            # Column Padding (B, T//key, H, key//S, key + S)
+            att_scores = F.pad(att_scores, pad=(1, self.stride - 1), value=0)
+
+            # Reshape (B, T//key, H, 1 + key//S, key)
+            att_scores = att_scores.reshape(batch_size, -1, self.num_heads, self.kernel_size // self.stride + 1,
+                                            self.kernel_size)
+
+            # Slice (B, T//key, H, key//S, key)
+            att_scores = att_scores[:, :, :, 1:]
+
+        # Full Context
+        else:
+
+            # Att Scores (B, H, T//S, 2*key-1)
+            batch_size, num_heads, seq_length1, seq_length2 = att_scores.size()
+
+            # Reshape (B, T//key, H, key//S, 2*key-1)
+            att_scores = att_scores.reshape(batch_size, -1, self.num_heads, self.kernel_size // self.stride,
+                                            seq_length2)
+
+            # Column Padding (B, T//key, H, key//S, 2*key-1 + S)
+            att_scores = F.pad(att_scores, pad=(0, self.stride), value=0)
+
+            # Flatten (B, T//key, H, 2KK//S - key//S + key)
+            att_scores = att_scores.reshape(batch_size, -1, self.num_heads,
+                                            self.kernel_size // self.stride * (2 * self.kernel_size - 1 + self.stride))
+
+            # End Padding (B, T//key, H, 2KK//S - key//S + 2K-1)
+            att_scores = F.pad(att_scores, pad=(0, self.kernel_size - 1), value=0)
+
+            # Reshape (B, T//key, H, key//S + 1, 2*key-1)
+            att_scores = att_scores.reshape(batch_size, -1, self.num_heads, self.kernel_size // self.stride + 1,
+                                            seq_length2)
+
+            # Slice (B, T//key, H, key//S, key)
+            att_scores = att_scores[:, :, :, :self.kernel_size // self.stride, self.kernel_size - 1:]
+
+        return att_scores
+
+    def forward(self, query, key, value, mask=None, hidden=None):
+
+        # Batch size B
+        batch_size = query.size(0)
+
+        # Chunk Padding
+        query, key, value, mask, padding = self.pad(query, key, value, mask, chunk_size=self.kernel_size)
+
+        # Query Subsampling (B, T, D) -> (B, T//S, D)
+        query = query[:, ::self.stride]
+
+        # Linear Layers
+        query = self.query_layer(query)
+        key = self.key_layer(key)
+        value = self.value_layer(value)
+
+        # Add Bias
+        Qu = query + self.u
+        Qv = query + self.v
+
+        # Relative Positional Embeddings (B, 2*key-1, D) / (B, key, D)
+        E = self.pos_layer(self.rel_pos_enc(batch_size))
+
+        # Reshape and Transpose (B, T//S, D) -> (B, H, T//S, d)
+        Qv = Qu.reshape(batch_size, -1, self.num_heads, self.dim_head).transpose(1, 2)
+        # Reshape and Transpose (B, T//S, D) -> (B, T//key, H, key//S, d)
+        Qu = Qv.reshape(batch_size, -1, self.kernel_size // self.stride, self.num_heads, self.dim_head).transpose(2, 3)
+        # Reshape and Transpose (B, T, D) -> (B, T//key, H, key, d)
+        key = key.reshape(batch_size, -1, self.kernel_size, self.num_heads, self.dim_head).transpose(2, 3)
+        value = value.reshape(batch_size, -1, self.kernel_size, self.num_heads, self.dim_head).transpose(2, 3)
+        # Reshape and Transpose (B, 2*key-1, D) -> (B, H, 2*key-1, d) / (B, key, D) -> (B, H, key, d)
+        E = E.reshape(batch_size, -1, self.num_heads, self.dim_head).transpose(1, 2)
+
+        # att_scores (B, T//key, H, key//S, key)
+        att_scores_K = Qu.matmul(key.transpose(3, 4))
+        att_scores_E = self.rel_to_abs(Qv.matmul(E.transpose(2, 3)))
+        att_scores = (att_scores_K + att_scores_E) / key.shape[-1] ** 0.5
+
+        # Mask scores
+        if mask is not None:
+
+            # Diagonal Mask (B, 1, T, T) -> (B, T//key, 1, key, key)
+            masks = []
+            for m in range(mask.size(-1) // self.kernel_size):
+                masks.append(mask[:, :, m * self.kernel_size: (m + 1) * self.kernel_size,
+                             m * self.kernel_size: (m + 1) * self.kernel_size])
+            mask = torch.stack(masks, dim=1)
+
+            # Stride Mask (B, T//key, 1, key, key) -> (B, T//key, 1, key//S, key)
+            mask = mask[:, :, :, ::self.stride]
+
+            # Apply Mask
+            att_scores = att_scores.float() - mask.float() * 1e9
+
+        # Attention weights (B, T//key, H, key//S, key)
+        att_w = att_scores.softmax(dim=-1)
+
+        # Attention output (B, T//key, H, key//S, d)
+        out = att_w.matmul(value)
+
+        # Transpose and Reshape (B, T//key, H, key//S, d) -> (B, T//S, D)
+        out = out.transpose(2, 3).reshape(batch_size, -1, self.dim_model)
+
+        # Slice Padding
+        out = out[:, :(self.stride * out.size(1) - padding - 1) // self.stride + 1]
+
+        # Output linear layer
+        out = self.output_layer(out)
+
+        return out, att_w.detach(), hidden
+
+
+###############################################################################
+# Positional Encodings
+###############################################################################
+
+class SinusoidalPositionalEncoding(nn.Module):
+    """
+
+    Sinusoidal Positional Encoding
+
+    Reference: "Attention Is All You Need" by Vaswani et al.
+    https://arxiv.org/abs/1706.03762
+
+    """
+
+    def __init__(self, max_len, dim_model):
+        super(SinusoidalPositionalEncoding, self).__init__()
+
+        pos_encoding = torch.zeros(max_len, dim_model)
+        pos = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
+        i = torch.arange(0, dim_model // 2, dtype=torch.float).unsqueeze(0)
+        angles = pos / 10000 ** (2 * i / dim_model)
+
+        pos_encoding[:, 0::2] = angles.sin()
+        pos_encoding[:, 1::2] = angles.cos()
+        pos_encoding = pos_encoding.unsqueeze(0)
+
+        self.register_buffer('pos_encoding', pos_encoding, persistent=False)
+
+    def forward(self, batch_size=1, seq_len=None):
+
+        # (B, T, D)
+        if seq_len is not None:
+            P = self.pos_encoding[:, :seq_len]
+
+        # (B, Tmax, D)
+        else:
+            P = self.pos_encoding
+
+        return P.repeat(batch_size, 1, 1)
+
+
+class RelativeSinusoidalPositionalEncoding(nn.Module):
+    """
+        Relative Sinusoidal Positional Encoding
+
+        Positional encoding for left context (sin) and right context (cos)
+        Total context = 2 * max_len - 1
+    """
+
+    def __init__(self, max_len, dim_model, causal=False):
+        super(RelativeSinusoidalPositionalEncoding, self).__init__()
+
+        # PE
+        pos_encoding = torch.zeros(2 * max_len - 1, dim_model)
+
+        # Positions (max_len - 1, ..., max_len - 1)
+        pos_left = torch.arange(start=max_len - 1, end=0, step=-1, dtype=torch.float)
+        pos_right = torch.arange(start=0, end=-max_len, step=-1, dtype=torch.float)
+        pos = torch.cat([pos_left, pos_right], dim=0).unsqueeze(1)
+
+        # Angles
+        angles = pos / 10000 ** (2 * torch.arange(0, dim_model // 2, dtype=torch.float).unsqueeze(0) / dim_model)
+
+        # Rel Sinusoidal PE
+        pos_encoding[:, 0::2] = angles.sin()
+        pos_encoding[:, 1::2] = angles.cos()
+
+        pos_encoding = pos_encoding.unsqueeze(0)
+
+        self.register_buffer('pos_encoding', pos_encoding, persistent=False)
+        self.max_len = max_len
+        self.causal = causal
+
+    def forward(self, batch_size=1, seq_len=None, hidden_len=0):
+
+        # Causal Context
+        if self.causal:
+
+            # (B, Th + T, D)
+            if seq_len is not None:
+                R = self.pos_encoding[:, self.max_len - seq_len - hidden_len: self.max_len]
+
+            # (B, Tmax, D)
+            else:
+                R = self.pos_encoding[:, :self.max_len]
+
+        # Full Context
+        else:
+
+            # (B, Th + 2*T-1, D)
+            if seq_len is not None:
+                R = self.pos_encoding[:, self.max_len - seq_len - hidden_len: self.max_len - 1 + seq_len]
+
+            # (B, 2*Tmax-1, D)
+            else:
+                R = self.pos_encoding
+
+        return R.repeat(batch_size, 1, 1)
+
+
+class GroupedRelativeSinusoidalPositionalEncoding(nn.Module):
+    """
+        Relative Sinusoidal Positional Encoding for grouped multi-head attention
+
+        Positional encoding for left context (sin) and right context (cos)
+        Total context = 2 * max_len - group_size
+    """
+
+    def __init__(self, max_len, dim_model, group_size=1, causal=False):
+        super(GroupedRelativeSinusoidalPositionalEncoding, self).__init__()
+
+        # PE
+        pos_encoding = torch.zeros(2 * max_len - group_size % 2, dim_model)
+
+        # Positions (max_len - 1, ..., max_len - 1)
+        pos_left = torch.arange(start=max_len - 1, end=group_size % 2 - 1, step=-1, dtype=torch.float)
+        pos_right = torch.arange(start=0, end=-max_len, step=-1, dtype=torch.float)
+        pos = torch.cat([pos_left, pos_right], dim=0).unsqueeze(1)
+
+        # Angles
+        angles = pos / 10000 ** (2 * torch.arange(0, dim_model // 2, dtype=torch.float).unsqueeze(0) / dim_model)
+
+        # Rel Sinusoidal PE
+        pos_encoding[:, 0::2] = angles.sin()
+        pos_encoding[:, 1::2] = angles.cos()
+
+        pos_encoding = pos_encoding.unsqueeze(0)
+
+        self.register_buffer('pos_encoding', pos_encoding, persistent=False)
+        self.max_len = max_len
+        self.causal = causal
+        self.group_size = group_size
+
+    def forward(self, batch_size=1, seq_len=None, hidden_len=0):
+
+        # Causal Context
+        if self.causal:
+
+            # (B, Th + T, D)
+            if seq_len is not None:
+                R = self.pos_encoding[:, self.max_len - seq_len - hidden_len: self.max_len]
+
+            # (B, Tmax, D)
+            else:
+                R = self.pos_encoding[:, :self.max_len]
+        else:
+
+            # (B, Th + 2*T-G, D)
+            if seq_len is not None:
+                R = self.pos_encoding[:,
+                    self.max_len - seq_len + self.group_size // 2 - hidden_len: self.max_len - self.group_size % 2 + seq_len - self.group_size // 2]
+
+            # (B, 2*Tmax-G, D)
+            else:
+                R = self.pos_encoding
+
+        return R.repeat(batch_size, 1, 1)
+
+
+class MultiHeadSelfAttentionModule(nn.Module):
+    """Multi-Head Self-Attention Module
+
+    Args:
+        dim_model: model feature dimension
+        num_heads: number of attention heads
+        Pdrop: residual dropout probability
+        max_pos_encoding: maximum position
+        relative_pos_enc: whether to use relative postion embedding
+        causal: True for causal attention with masked future context
+        group_size: Attention group size
+        kernel_size: Attention kernel size
+        stride: Query stride
+        linear_att: whether to use multi-head linear self-attention
+
+    """
+
+    def __init__(self,
+                 dim_model,
+                 num_heads,
+                 Pdrop,
+                 max_pos_encoding,
+                 relative_pos_enc,
+                 causal,
+                 group_size,
+                 kernel_size,
+                 stride,
+                 linear_att):
+        super(MultiHeadSelfAttentionModule, self).__init__()
+
+        # Assert
+        assert not (group_size > 1 and kernel_size is not None), "Local grouped attention not implemented"
+        assert not (group_size > 1 and stride > 1 is not None), "Strided grouped attention not implemented"
+        assert not (linear_att and relative_pos_enc), "Linear attention requires absolute positional encodings"
+
+        # Pre Norm
+        self.norm = nn.LayerNorm(dim_model, eps=1e-6)
+
+        # Multi-Head Linear Attention
+        if linear_att:
+            self.mhsa = MultiHeadLinearAttention(dim_model, num_heads)
+
+        # Grouped Multi-Head Self-Attention
+        elif group_size > 1:
+            if relative_pos_enc:
+                self.mhsa = GroupedRelPosMultiHeadSelfAttention(dim_model, num_heads, causal, max_pos_encoding,
+                                                                group_size)
+            else:
+                self.mhsa = GroupedMultiHeadAttention(dim_model, num_heads, group_size)
+
+        # Local Multi-Head Self-Attention
+        elif kernel_size is not None and stride == 1:
+            if relative_pos_enc:
+                self.mhsa = LocalRelPosMultiHeadSelfAttention(dim_model, num_heads, causal, kernel_size)
+            else:
+                self.mhsa = LocalMultiHeadAttention(dim_model, num_heads, kernel_size)
+
+        # Strided Multi-Head Self-Attention
+        elif kernel_size is None and stride > 1:
+            if relative_pos_enc:
+                self.mhsa = StridedRelPosMultiHeadSelfAttention(dim_model, num_heads, causal, max_pos_encoding, stride)
+            else:
+                self.mhsa = StridedMultiHeadAttention(dim_model, num_heads, stride)
+
+        # Strided Local Multi-Head Self-Attention
+        elif stride > 1 and kernel_size is not None:
+            if relative_pos_enc:
+                self.mhsa = StridedLocalRelPosMultiHeadSelfAttention(dim_model, num_heads, causal, kernel_size, stride)
+            else:
+                self.mhsa = StridedLocalMultiHeadAttention(dim_model, num_heads, kernel_size, stride)
+
+        # Multi-Head Self-Attention
+        else:
+            if relative_pos_enc:
+                self.mhsa = RelPosMultiHeadSelfAttention(dim_model, num_heads, causal, max_pos_encoding)
+            else:
+                self.mhsa = MultiHeadAttention(dim_model, num_heads)
+
+        # Dropout
+        self.dropout = nn.Dropout(Pdrop)
+
+        # Module Params
+        self.rel_pos_enc = relative_pos_enc
+        self.linear_att = linear_att
+
+    def forward(self, x, mask=None, hidden=None):
+
+        # Pre Norm
+        x = self.norm(x)
+
+        # Multi-Head Self-Attention
+        if self.linear_att:
+            x, attention = self.mhsa(x, x, x)
+        elif self.rel_pos_enc:
+            x, attention, hidden = self.mhsa(x, x, x, mask, hidden)
+        else:
+            x, attention = self.mhsa(x, x, x, mask)
+
+        # Dropout
+        x = self.dropout(x)
+
+        return x, attention, hidden

fairseq/modules/convolution.py

-#!/usr/bin/env python3
-# -*- coding: utf-8 -*-
-
-# Copyright 2021 Mobvoi Inc. All Rights Reserved.
-# Author: di.wu@mobvoi.com (DI WU)
-"""ConvolutionModule definition."""
-
-from typing import Optional, Tuple
-
 import torch
 from torch import nn
 
-from fairseq.modules.layer_norm import LayerNorm
 from fairseq.modules.activations import get_activation_class
 
 
 class ConvolutionModule(nn.Module):
-    """ConvolutionModule in Conformer model."""
-    def __init__(self,
-                 channels: int,
-                 kernel_size: int = 15,
-                 norm: str = "batch_norm",
-                 bias: bool = True):
-        """Construct an ConvolutionModule object.
+    """Convolution block used in the conformer block"""
+
+    def __init__(
+        self,
+        embed_dim,
+        channels,
+        depthwise_kernel_size,
+        dropout,
+        activation_fn="swish",
+        bias=False,
+        export=False,
+    ):
+        """
         Args:
-            channels (int): The number of channels of conv layers.
-            kernel_size (int): Kernel size of conv layers.
-            causal (int): Whether use causal convolution or not
+            embed_dim: Embedding dimension
+            channels: Number of channels in depthwise conv layers
+            depthwise_kernel_size: Depthwise conv layer kernel size
+            dropout: dropout value
+            activation_fn: Activation function to use after depthwise convolution kernel
+            bias: If bias should be added to conv layers
+            export: If layernorm should be exported to jit
         """
-        super().__init__()
-
-        self.pointwise_conv1 = nn.Conv1d(
-            channels,
+        super(ConvolutionModule, self).__init__()
+        assert (
+            depthwise_kernel_size - 1
+        ) % 2 == 0, "kernel_size should be a odd number for 'SAME' padding"
+        self.pointwise_conv1 = torch.nn.Conv1d(
+            embed_dim,
             2 * channels,
             kernel_size=1,
             stride=1,
             padding=0,
             bias=bias,
         )
-
-        # kernel_size should be an odd number for none causal convolution
-        assert (kernel_size - 1) % 2 == 0
-        padding = (kernel_size - 1) // 2
-
-        self.depthwise_conv = nn.Conv1d(
+        self.glu = torch.nn.GLU(dim=1)
+        self.depthwise_conv = torch.nn.Conv1d(
             channels,
             channels,
-            kernel_size,
+            depthwise_kernel_size,
             stride=1,
-            padding=padding,
+            padding=(depthwise_kernel_size - 1) // 2,
             groups=channels,
             bias=bias,
         )
-
-        assert norm in ['batch_norm', 'layer_norm']
-        if norm == "batch_norm":
-            self.use_layer_norm = False
-            self.norm = nn.BatchNorm1d(channels)
-        else:
-            self.use_layer_norm = True
-            self.norm = LayerNorm(channels)
-
-        self.pointwise_conv2 = nn.Conv1d(
-            channels,
+        self.batch_norm = nn.BatchNorm1d(channels)
+        self.activation = get_activation_class(activation_fn)
+        self.pointwise_conv2 = torch.nn.Conv1d(
             channels,
+            embed_dim,
             kernel_size=1,
             stride=1,
             padding=0,
             bias=bias,
         )
-        self.activation = get_activation_class("swish")
+        self.dropout = torch.nn.Dropout(dropout)
 
-    def forward(
-        self,
-        x: torch.Tensor,
-        mask_pad: Optional[torch.Tensor] = None,
-    ) -> Tuple[torch.Tensor, torch.Tensor]:
-        """Compute convolution module.
+    def forward(self, x):
+        """
         Args:
-            x (torch.Tensor): Input tensor (#batch, time, channels).
-            mask_pad (torch.Tensor): used for batch padding
+            x: Input of shape B X T X C
         Returns:
-            torch.Tensor: Output tensor (#batch, time, channels).
+          Tensor of shape B X T X C
         """
         # exchange the temporal dimension and the feature dimension
         x = x.transpose(1, 2)
 
-        zero_mask_pad = mask_pad.unsqueeze(1).repeat(1, x.size(1), 1)
-        # mask batch padding
-        if mask_pad is not None:
-            x.masked_fill_(zero_mask_pad, 0.0)
-
         # GLU mechanism
-        x = self.pointwise_conv1(x)  # (batch, 2*channel, time)
-        x = nn.functional.glu(x, dim=1)  # (batch, channel, time)
+        x = self.pointwise_conv1(x)  # (batch, 2*channel, dim)
+        x = self.glu(x)  # (batch, channel, dim)
 
         # 1D Depthwise Conv
         x = self.depthwise_conv(x)
-        if self.use_layer_norm:
-            x = x.transpose(1, 2)
-        x = self.activation(self.norm(x))
-        if self.use_layer_norm:
-            x = x.transpose(1, 2)
-        x = self.pointwise_conv2(x)
-        # mask batch padding
-        if zero_mask_pad is not None:
-            x.masked_fill_(zero_mask_pad, 0.0)
+        x = self.batch_norm(x)
+        x = self.activation(x)
 
+        x = self.pointwise_conv2(x)
+        x = self.dropout(x)
         return x.transpose(1, 2)
+
+#
+# class ConvolutionModule(nn.Module):
+#     """ConvolutionModule in Conformer model."""
+#     def __init__(self,
+#                  channels: int,
+#                  kernel_size: int = 15,
+#                  norm: str = "batch_norm",
+#                  bias: bool = True):
+#         """Construct an ConvolutionModule object.
+#         Args:
+#             channels (int): The number of channels of conv layers.
+#             kernel_size (int): Kernel size of conv layers.
+#             causal (int): Whether use causal convolution or not
+#         """
+#         super().__init__()
+#
+#         self.pointwise_conv1 = nn.Conv1d(
+#             channels,
+#             2 * channels,
+#             kernel_size=1,
+#             stride=1,
+#             padding=0,
+#             bias=bias,
+#         )
+#
+#         # kernel_size should be an odd number for none causal convolution
+#         assert (kernel_size - 1) % 2 == 0
+#         padding = (kernel_size - 1) // 2
+#
+#         self.depthwise_conv = nn.Conv1d(
+#             channels,
+#             channels,
+#             kernel_size,
+#             stride=1,
+#             padding=padding,
+#             groups=channels,
+#             bias=bias,
+#         )
+#
+#         assert norm in ['batch_norm', 'layer_norm']
+#         if norm == "batch_norm":
+#             self.use_layer_norm = False
+#             self.norm = nn.BatchNorm1d(channels)
+#         else:
+#             self.use_layer_norm = True
+#             self.norm = LayerNorm(channels)
+#
+#         self.pointwise_conv2 = nn.Conv1d(
+#             channels,
+#             channels,
+#             kernel_size=1,
+#             stride=1,
+#             padding=0,
+#             bias=bias,
+#         )
+#         self.activation = get_activation_class("swish")
+#
+#     def forward(
+#         self,
+#         x: torch.Tensor,
+#         mask_pad: Optional[torch.Tensor] = None,
+#     ) -> Tuple[torch.Tensor, torch.Tensor]:
+#         """Compute convolution module.
+#         Args:
+#             x (torch.Tensor): Input tensor (#batch, time, channels).
+#             mask_pad (torch.Tensor): used for batch padding
+#         Returns:
+#             torch.Tensor: Output tensor (#batch, time, channels).
+#         """
+#         # exchange the temporal dimension and the feature dimension
+#         x = x.transpose(1, 2)
+#
+#         # zero_mask_pad = mask_pad.unsqueeze(1).repeat(1, x.size(1), 1)
+#         # # mask batch padding
+#         # if mask_pad is not None:
+#         #     x.masked_fill_(zero_mask_pad, 0.0)
+#
+#         # GLU mechanism
+#         x = self.pointwise_conv1(x)  # (batch, 2*channel, time)
+#         x = nn.functional.glu(x, dim=1)  # (batch, channel, time)
+#
+#         # 1D Depthwise Conv
+#         x = self.depthwise_conv(x)
+#         if self.use_layer_norm:
+#             x = x.transpose(1, 2)
+#         x = self.activation(self.norm(x))
+#         if self.use_layer_norm:
+#             x = x.transpose(1, 2)
+#         x = self.pointwise_conv2(x)
+#
+#         # # mask batch padding
+#         # if zero_mask_pad is not None:
+#         #     x.masked_fill_(zero_mask_pad, 0.0)
+#
+#         return x.transpose(1, 2)

fairseq/modules/espnet_multihead_attention.py

+#!/usr/bin/env python3
+# -*- coding: utf-8 -*-
+
+# Copyright 2019 Shigeki Karita
+#  Apache 2.0  (http://www.apache.org/licenses/LICENSE-2.0)
+
+"""Multi-Head Attention layer definition."""
+
+import math
+import torch
+from torch import nn
+from fairseq.modules.rotary_positional_embedding import (
+    RotaryPositionalEmbedding,
+    apply_rotary_pos_emb,
+)
+
+
+class ESPNETMultiHeadedAttention(nn.Module):
+    """Multi-Head Attention layer.
+    Args:
+        n_head: The number of heads.
+        n_feat: The number of features.
+        dropout: Dropout rate.
+    """
+
+    def __init__(self, n_feat, n_head, dropout):
+        """Construct an MultiHeadedAttention object."""
+        super(ESPNETMultiHeadedAttention, self).__init__()
+        assert n_feat % n_head == 0
+        # We assume d_v always equals d_k
+        self.d_k = n_feat // n_head
+        self.h = n_head
+        self.linear_q = nn.Linear(n_feat, n_feat)
+        self.linear_k = nn.Linear(n_feat, n_feat)
+        self.linear_v = nn.Linear(n_feat, n_feat)
+        self.linear_out = nn.Linear(n_feat, n_feat)
+        self.attn = None
+        self.dropout = nn.Dropout(p=dropout)
+
+    def forward_qkv(self, query, key, value, **kwargs):
+        """Transform query, key and value.
+        Args:
+            query: Query tensor  B X T1 X C
+            key: Key tensor B X T2 X C
+            value: Value tensor  B X T2 X C
+        Returns:
+            torch.Tensor: Transformed query tensor  B X n_head X T1 X d_k
+            torch.Tensor: Transformed key tensor B X n_head X T2 X d_k
+            torch.Tensor: Transformed value tensor  B X n_head X T2 X d_k
+        """
+        n_batch = query.size(0)
+        q = self.linear_q(query).view(n_batch, -1, self.h, self.d_k)
+        k = self.linear_k(key).view(n_batch, -1, self.h, self.d_k)
+        v = self.linear_v(value).view(n_batch, -1, self.h, self.d_k)
+        q = q.transpose(1, 2)  # (batch, head, time1, d_k)
+        k = k.transpose(1, 2)  # (batch, head, time2, d_k)
+        v = v.transpose(1, 2)  # (batch, head, time2, d_k)
+        return q, k, v
+
+    def forward_attention(self, value, scores, mask):
+        """Compute attention context vector.
+        Args:
+            value: Transformed value B X n_head X T2 X d_k.
+            scores: Attention score  B X n_head X T1 X T2
+            mask: Mask  T2 X B
+        Returns:
+            torch.Tensor: Transformed value  B X T1 X d_model
+                weighted by the attention score  B X T1 X T2
+        """
+        n_batch = value.size(0)
+        if mask is not None:
+            scores = scores.masked_fill(
+                mask.unsqueeze(1).unsqueeze(2).to(bool),
+                float("-inf"),  # (batch, head, time1, time2)
+            )
+            self.attn = torch.softmax(scores, dim=-1)  # (batch, head, time1, time2)
+
+        else:
+            self.attn = torch.softmax(scores, dim=-1)  # (batch, head, time1, time2)
+        p_attn = self.dropout(self.attn)
+        x = torch.matmul(p_attn, value)  # (batch, head, time1, d_k)
+        x = (
+            x.transpose(1, 2).contiguous().view(n_batch, -1, self.h * self.d_k)
+        )  # (batch, time1, d_model)
+
+        return self.linear_out(x)  # (batch, time1, d_model)
+
+    def forward(self, query, key, value, key_padding_mask=None, **kwargs):
+        """Compute scaled dot product attention.
+        Args:
+            query (torch.Tensor): Query tensor T X B X C
+            key (torch.Tensor): Key tensor T X B X C
+            value (torch.Tensor): Value tensor T X B X C
+            mask (torch.Tensor): Mask tensor T X B
+        Returns:
+            torch.Tensor: Output tensor T X B X D.
+        """
+        query = query.transpose(0, 1)
+        key = key.transpose(0, 1)
+        value = value.transpose(0, 1)
+
+        q, k, v = self.forward_qkv(query, key, value)
+        scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.d_k)
+        scores = self.forward_attention(v, scores, key_padding_mask)
+        scores = scores.transpose(0, 1)
+        return scores, None
+
+
+class RelPositionMultiHeadedAttention(ESPNETMultiHeadedAttention):
+    """Multi-Head Attention layer with relative position encoding.
+    Paper: https://arxiv.org/abs/1901.02860
+    Args:
+        n_head: The number of heads.
+        n_feat: The number of features.
+        dropout: Dropout rate.
+        zero_triu: Whether to zero the upper triangular part of attention matrix.
+    """
+
+    def __init__(self, n_feat, n_head, dropout, zero_triu=False):
+        """Construct an RelPositionMultiHeadedAttention object."""
+        super().__init__(n_feat, n_head, dropout)
+        self.zero_triu = zero_triu
+        # linear transformation for positional encoding
+        self.linear_pos = nn.Linear(n_feat, n_feat, bias=False)
+        # these two learnable bias are used in matrix c and matrix d
+        # as described in https://arxiv.org/abs/1901.02860 Section 3.3
+        self.pos_bias_u = nn.Parameter(torch.Tensor(self.h, self.d_k))
+        self.pos_bias_v = nn.Parameter(torch.Tensor(self.h, self.d_k))
+        torch.nn.init.xavier_uniform_(self.pos_bias_u)
+        torch.nn.init.xavier_uniform_(self.pos_bias_v)
+
+    def rel_shift(self, x):
+        """Compute relative positional encoding.
+        Args:
+            x: Input tensor B X n_head X T X 2T-1
+        Returns:
+            torch.Tensor: Output tensor.
+        """
+        zero_pad = torch.zeros((*x.size()[:3], 1), device=x.device, dtype=x.dtype)
+        x_padded = torch.cat([zero_pad, x], dim=-1)
+
+        x_padded = x_padded.view(*x.size()[:2], x.size(3) + 1, x.size(2))
+        x = x_padded[:, :, 1:].view_as(x)[
+            :, :, :, : x.size(-1) // 2 + 1
+        ]  # only keep the positions from 0 to time2
+
+        if self.zero_triu:
+            ones = torch.ones((x.size(2), x.size(3)), device=x.device)
+            x = x * torch.tril(ones, x.size(3) - x.size(2))[None, None, :, :]
+
+        return x
+
+    def forward(self, query, key, value, pos_emb, key_padding_mask=None, **kwargs):
+        """Compute scaled dot product attention.
+        Args:
+            query: Query tensor T X B X C
+            key: Key tensor T X B X C
+            value: Value tensor T X B X C
+            pos_emb: Positional embedding tensor B X 2T-1 X C
+            key_padding_mask: Mask tensor T X B
+        Returns:
+            torch.Tensor: Output tensor T X B X C.
+        """
+        query = query.transpose(0, 1)
+        key = key.transpose(0, 1)
+        value = value.transpose(0, 1)
+        pos_emb = pos_emb.transpose(0, 1)
+        q, k, v = self.forward_qkv(query, key, value)
+        q = q.transpose(1, 2)  # (batch, time1, head, d_k)
+        n_batch_pos = pos_emb.size(0)
+        p = self.linear_pos(pos_emb).view(n_batch_pos, -1, self.h, self.d_k)
+        p = p.transpose(1, 2)  # (batch, head, 2*time1-1, d_k)
+
+        # (batch, head, time1, d_k)
+        q_with_bias_u = (q + self.pos_bias_u).transpose(1, 2)
+        # (batch, head, time1, d_k)
+        q_with_bias_v = (q + self.pos_bias_v).transpose(1, 2)
+
+        # compute attention score
+        # first compute matrix a and matrix c
+        # as described in https://arxiv.org/abs/1901.02860 Section 3.3
+        # (batch, head, time1, time2)
+        matrix_ac = torch.matmul(q_with_bias_u, k.transpose(-2, -1))
+
+        # compute matrix b and matrix d
+        # (batch, head, time1, 2*time1-1)
+        matrix_bd = torch.matmul(q_with_bias_v, p.transpose(-2, -1))
+        matrix_bd = self.rel_shift(matrix_bd)
+
+        scores = (matrix_ac + matrix_bd) / math.sqrt(
+            self.d_k
+        )  # (batch, head, time1, time2)
+
+        scores = self.forward_attention(v, scores, key_padding_mask)
+        scores = scores.transpose(0, 1)
+        return scores, None
+
+
+class RotaryPositionMultiHeadedAttention(ESPNETMultiHeadedAttention):
+    def __init__(
+        self,
+        n_feat,
+        n_head,
+        dropout,
+        precision,
+        rotary_emd_base=10000,
+    ):
+        """Construct an RotaryPositionMultiHeadedAttention object."""
+        super().__init__(n_feat, n_head, dropout)
+        precision = torch.float
+        self.rotary_ndims = self.d_k  # also try self.d_k//2
+        if precision == "fp16":
+            precision = torch.half
+
+        self.rotary_emb = RotaryPositionalEmbedding(
+            self.rotary_ndims, base=rotary_emd_base, precision=precision
+        )
+
+    def forward(self, query, key, value, key_padding_mask=None, **kwargs):
+        """Compute rotary position attention.
+        Args:
+            query: Query tensor T X B X C
+            key: Key tensor T X B X C
+            value: Value tensor T X B X C
+            key_padding_mask: Mask tensor T X B
+        Returns:
+            torch.Tensor: Output tensor T X B X D.
+        Notes:
+            Assumes self attn
+        """
+
+        T, B, C = value.size()
+        query = query.view(T, B, self.h, self.d_k)
+        key = key.view(T, B, self.h, self.d_k)
+        value = value.view(T, B, self.h, self.d_k)
+        cos, sin = self.rotary_emb(value, seq_len=T)
+        query, key = apply_rotary_pos_emb(
+            query, key, cos, sin, offset=0
+        )  # offset is based on layer_past
+
+        query = query.view(T, B, self.h * self.d_k)
+        key = key.view(T, B, self.h * self.d_k)
+        value = value.view(T, B, self.h * self.d_k)
+
+        # TBD to BTD
+        query = query.transpose(0, 1)
+        key = key.transpose(0, 1)
+        value = value.transpose(0, 1)
+
+        q, k, v = self.forward_qkv(query, key, value)
+        scores = torch.matmul(q, k.transpose(-2, -1)) / math.sqrt(self.d_k)
+        scores = self.forward_attention(v, scores, key_padding_mask)
+        scores = scores.transpose(0, 1)
+        return scores, None

fairseq/modules/location_attention.py

+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch.nn as nn
+import torch
+import torch.nn.functional as F
+
+
+class LocationAttention(nn.Module):
+    """
+    Attention-Based Models for Speech Recognition
+    https://arxiv.org/pdf/1506.07503.pdf
+
+    :param int encoder_dim: # projection-units of encoder
+    :param int decoder_dim: # units of decoder
+    :param int attn_dim: attention dimension
+    :param int conv_dim: # channels of attention convolution
+    :param int conv_kernel_size: filter size of attention convolution
+    """
+
+    def __init__(
+        self,
+        attn_dim,
+        encoder_dim,
+        decoder_dim,
+        attn_state_kernel_size,
+        conv_dim,
+        conv_kernel_size,
+        scaling=2.0,
+    ):
+        super(LocationAttention, self).__init__()
+        self.attn_dim = attn_dim
+        self.decoder_dim = decoder_dim
+        self.scaling = scaling
+        self.proj_enc = nn.Linear(encoder_dim, attn_dim)
+        self.proj_dec = nn.Linear(decoder_dim, attn_dim, bias=False)
+        self.proj_attn = nn.Linear(conv_dim, attn_dim, bias=False)
+        self.conv = nn.Conv1d(
+            attn_state_kernel_size,
+            conv_dim,
+            2 * conv_kernel_size + 1,
+            padding=conv_kernel_size,
+            bias=False,
+        )
+        self.proj_out = nn.Sequential(nn.Tanh(), nn.Linear(attn_dim, 1))
+
+        self.proj_enc_out = None  # cache
+
+    def clear_cache(self):
+        self.proj_enc_out = None
+
+    def forward(self, encoder_out, encoder_padding_mask, decoder_h, attn_state):
+        """
+        :param torch.Tensor encoder_out: padded encoder hidden state B x T x D
+        :param torch.Tensor encoder_padding_mask: encoder padding mask
+        :param torch.Tensor decoder_h: decoder hidden state B x D
+        :param torch.Tensor attn_prev: previous attention weight B x K x T
+        :return: attention weighted encoder state (B, D)
+        :rtype: torch.Tensor
+        :return: previous attention weights (B x T)
+        :rtype: torch.Tensor
+        """
+        bsz, seq_len, _ = encoder_out.size()
+        if self.proj_enc_out is None:
+            self.proj_enc_out = self.proj_enc(encoder_out)
+
+        # B x K x T -> B x C x T
+        attn = self.conv(attn_state)
+        # B x C x T -> B x T x C -> B x T x D
+        attn = self.proj_attn(attn.transpose(1, 2))
+
+        if decoder_h is None:
+            decoder_h = encoder_out.new_zeros(bsz, self.decoder_dim)
+        dec_h = self.proj_dec(decoder_h).view(bsz, 1, self.attn_dim)
+
+        out = self.proj_out(attn + self.proj_enc_out + dec_h).squeeze(2)
+        out.masked_fill_(encoder_padding_mask, -float("inf"))
+
+        w = F.softmax(self.scaling * out, dim=1)
+        c = torch.sum(encoder_out * w.view(bsz, seq_len, 1), dim=1)
+        return c, w

fairseq/modules/positional_encoding.py

+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch.nn as nn
+import math
+import torch
+
+
+class PositionalEncoding(nn.Module):
+    """Positional encoding.
+
+    Args:
+        d_model: Embedding dimension.
+        dropout_rate: Dropout rate.
+        max_len: Maximum input length.
+        reverse: Whether to reverse the input position.
+    """
+
+    def __init__(self, d_model, dropout_rate, max_len=5000, reverse=False):
+        """Construct an PositionalEncoding object."""
+        super(PositionalEncoding, self).__init__()
+        self.d_model = d_model
+        self.reverse = reverse
+        self.xscale = math.sqrt(self.d_model)
+        self.dropout = nn.Dropout(p=dropout_rate)
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(1, max_len))
+
+    def extend_pe(self, x):
+        """Reset the positional encodings."""
+        if self.pe is not None:
+            if self.pe.size(1) >= x.size(1):
+                if self.pe.dtype != x.dtype or self.pe.device != x.device:
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        pe = torch.zeros(x.size(1), self.d_model)
+        if self.reverse:
+            position = torch.arange(
+                x.size(1) - 1, -1, -1.0, dtype=torch.float32
+            ).unsqueeze(1)
+        else:
+            position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe[:, 0::2] = torch.sin(position * div_term)
+        pe[:, 1::2] = torch.cos(position * div_term)
+        pe = pe.unsqueeze(0)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor):
+        """Add positional encoding.
+        Args:
+            x (torch.Tensor): Input tensor B X T X C
+        Returns:
+            torch.Tensor: Encoded tensor B X T X C
+        """
+        self.extend_pe(x)
+        x = x * self.xscale + self.pe[:, : x.size(1)]
+        return self.dropout(x)
+
+
+class RelPositionalEncoding(nn.Module):
+    """Relative positional encoding module (new implementation).
+
+    Args:
+        d_model: Embedding dimension.
+        dropout_rate: Dropout rate.
+        max_len: Maximum input length.
+    """
+
+    def __init__(self, max_len, d_model):
+        """Construct an PositionalEncoding object."""
+        super(RelPositionalEncoding, self).__init__()
+        self.d_model = d_model
+        self.pe = None
+        self.extend_pe(torch.tensor(0.0).expand(1, max_len))
+
+    def extend_pe(self, x):
+        """Reset the positional encodings."""
+        if self.pe is not None:
+            # self.pe contains both positive and negative parts
+            # the length of self.pe is 2 * input_len - 1
+            if self.pe.size(1) >= x.size(1) * 2 - 1:
+                if self.pe.dtype != x.dtype or self.pe.device != x.device:
+                    self.pe = self.pe.to(dtype=x.dtype, device=x.device)
+                return
+        # Suppose `i` means to the position of query vecotr and `j` means the
+        # position of key vector. We use position relative positions when keys
+        # are to the left (i>j) and negative relative positions otherwise (i<j).
+        pe_positive = torch.zeros(x.size(1), self.d_model)
+        pe_negative = torch.zeros(x.size(1), self.d_model)
+        position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1)
+        div_term = torch.exp(
+            torch.arange(0, self.d_model, 2, dtype=torch.float32)
+            * -(math.log(10000.0) / self.d_model)
+        )
+        pe_positive[:, 0::2] = torch.sin(position * div_term)
+        pe_positive[:, 1::2] = torch.cos(position * div_term)
+        pe_negative[:, 0::2] = torch.sin(-1 * position * div_term)
+        pe_negative[:, 1::2] = torch.cos(-1 * position * div_term)
+
+        # Reserve the order of positive indices and concat both positive and
+        # negative indices. This is used to support the shifting trick
+        # as in https://arxiv.org/abs/1901.02860
+        pe_positive = torch.flip(pe_positive, [0]).unsqueeze(0)
+        pe_negative = pe_negative[1:].unsqueeze(0)
+        pe = torch.cat([pe_positive, pe_negative], dim=1)
+        self.pe = pe.to(device=x.device, dtype=x.dtype)
+
+    def forward(self, x: torch.Tensor):
+        """Add positional encoding.
+        Args:
+            x : Input tensor T X B X C.
+        Returns:
+            torch.Tensor: Encoded tensor T X B X C.
+
+        """
+        x = x.transpose(0, 1)  # Change TBC to BTC
+        self.extend_pe(x)
+        pos_emb = self.pe[
+            :,
+            self.pe.size(1) // 2 - x.size(1) + 1 : self.pe.size(1) // 2 + x.size(1),
+        ]
+        pos_emb = pos_emb.transpose(0, 1)  # change to TBC
+        return pos_emb

fairseq/modules/rotary_positional_embedding.py

+import torch
+
+
+class RotaryPositionalEmbedding(torch.nn.Module):
+    def __init__(self, dim, base=10000, precision=torch.half):
+        """Rotary positional embedding
+        Reference : https://blog.eleuther.ai/rotary-embeddings/
+        Paper: https://arxiv.org/pdf/2104.09864.pdf
+        Args:
+            dim: Dimension of embedding
+            base: Base value for exponential
+            precision: precision to use for numerical values
+        """
+        super().__init__()
+        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
+        self.register_buffer("inv_freq", inv_freq)
+        self.seq_len_cached = None
+        self.cos_cached = None
+        self.sin_cached = None
+        self.precision = precision
+
+    def forward(self, x, seq_len=None):
+        """
+        Args:
+            x: Input x with T X B X C
+            seq_len: Sequence length of input x
+        """
+        if seq_len != self.seq_len_cached:
+            self.seq_len_cached = seq_len
+            t = torch.arange(seq_len, device=x.device).type_as(self.inv_freq)
+            freqs = torch.einsum("i,j->ij", t, self.inv_freq)
+            emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
+            self.cos_cached = emb.cos()[:, None, None, :]
+            self.sin_cached = emb.sin()[:, None, None, :]
+        return self.cos_cached, self.sin_cached
+
+
+# rotary pos emb helpers:
+def rotate_half(x):
+    x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :]
+    return torch.cat(
+        (-x2, x1), dim=x1.ndim - 1
+    )  # dim=-1 triggers a bug in earlier torch versions
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, offset: int = 0):
+    cos, sin = (
+        cos[offset : q.shape[0] + offset, ...],
+        sin[offset : q.shape[0] + offset, ...],
+    )
+    return (q * cos) + (rotate_half(q) * sin), (k * cos) + (rotate_half(k) * sin)

fairseq/modules/s2t_transformer_layer.py

@@ -7,17 +7,62 @@ from typing import Optional
 
 import torch
 import torch.nn as nn
-from fairseq import utils
 from fairseq.modules import (
     LayerNorm,
     MultiheadAttention,
     RelPositionMultiheadAttention,
     RelativeMultiheadAttention,
-    ConvolutionModule
+    ConvolutionModule,
+    ESPNETMultiHeadedAttention,
+    RelPositionMultiHeadedAttention,
+    RotaryPositionMultiHeadedAttention,
 )
 from fairseq.modules.fairseq_dropout import FairseqDropout
-from fairseq.modules.quant_noise import quant_noise
 from torch import Tensor
+from fairseq.modules.activations import get_activation_fn, get_activation_class
+
+
+class FeedForwardModule(torch.nn.Module):
+    """Positionwise feed forward layer used in conformer"""
+
+    def __init__(
+        self,
+        input_feat,
+        hidden_units,
+        dropout1,
+        dropout2,
+        activation_fn="relu",
+        bias=True,
+    ):
+        """
+        Args:
+            input_feat: Input feature dimension
+            hidden_units: Hidden unit dimension
+            dropout1: dropout value for layer1
+            dropout2: dropout value for layer2
+            activation_fn: Name of activation function
+            bias: If linear layers should have bias
+        """
+
+        super(FeedForwardModule, self).__init__()
+        self.w_1 = torch.nn.Linear(input_feat, hidden_units, bias=bias)
+        self.w_2 = torch.nn.Linear(hidden_units, input_feat, bias=bias)
+        self.dropout1 = torch.nn.Dropout(dropout1)
+        self.dropout2 = torch.nn.Dropout(dropout2)
+        self.activation = get_activation_class(activation_fn)
+
+    def forward(self, x):
+        """
+        Args:
+            x: Input Tensor of shape  T X B X C
+        Returns:
+            Tensor of shape T X B X C
+        """
+        x = self.w_1(x)
+        x = self.activation(x)
+        x = self.dropout1(x)
+        x = self.w_2(x)
+        return self.dropout2(x)
 
 
 class S2TTransformerEncoderLayer(nn.Module):
@@ -38,6 +83,10 @@ class S2TTransformerEncoderLayer(nn.Module):
     def __init__(self, args):
         super().__init__()
         self.args = args
+
+        embed_dim = args.encoder_embed_dim
+        ffn_dim = args.encoder_ffn_embed_dim
+        dropout = args.dropout
         self.embed_dim = args.encoder_embed_dim
         self.quant_noise = getattr(args, 'quant_noise_pq', 0)
         self.quant_noise_block_size = getattr(args, 'quant_noise_pq_block_size', 8) or 8
@@ -45,77 +94,53 @@ class S2TTransformerEncoderLayer(nn.Module):
         self.self_attn = self.build_self_attention(self.embed_dim, args)
         self.self_attn_layer_norm = LayerNorm(self.embed_dim)
         self.dropout_module = FairseqDropout(
-            args.dropout, module_name=self.__class__.__name__
-        )
-        self.activation_fn = utils.get_activation_fn(
-            activation=getattr(args, 'activation_fn', 'relu') or "relu"
-        )
-        activation_dropout_p = getattr(args, "activation_dropout", 0) or 0
-        if activation_dropout_p == 0:
-            # for backwards compatibility with models that use args.relu_dropout
-            activation_dropout_p = getattr(args, "relu_dropout", 0) or 0
-        self.activation_dropout_module = FairseqDropout(
-            float(activation_dropout_p), module_name=self.__class__.__name__
+            dropout, module_name=self.__class__.__name__
         )
+        self.normalize_before = args.encoder_normalize_before
+        activation = getattr(args, 'encoder_activation_fn', 'relu')
+
         if args.macaron_style:
-            self.macaron_fc1 = self.build_fc1(
-                self.embed_dim,
-                args.encoder_ffn_embed_dim,
-                self.quant_noise,
-                self.quant_noise_block_size,
+            self.macaron_ffn = FeedForwardModule(
+                embed_dim,
+                ffn_dim,
+                dropout,
+                dropout,
+                activation
             )
-            self.macaron_fc2 = self.build_fc2(
-                args.encoder_ffn_embed_dim,
-                self.embed_dim,
-                self.quant_noise,
-                self.quant_noise_block_size,
-            )
-            self.macaron_norm = LayerNorm(self.embed_dim)
+            self.macaron_norm = LayerNorm(embed_dim)
             self.ffn_scale = 0.5
         else:
-            self.macaron_fc1 = None
-            self.macaron_fc2 = None
+            self.macaron_ffn = None
             self.macaron_norm = None
             self.ffn_scale = 1.0
 
         if args.use_cnn_module:
-            self.conv_norm = LayerNorm(self.embed_dim)
+            self.conv_norm = LayerNorm(embed_dim)
             self.conv_module = ConvolutionModule(
                 self.embed_dim,
-                args.cnn_module_kernel)
-            self.final_norm = LayerNorm(self.embed_dim)
+                self.embed_dim,
+                depthwise_kernel_size=args.cnn_module_kernel,
+                dropout=args.dropout,
+                activation_fn=getattr(args, 'activation_fn', 'swish'))
+            self.final_norm = LayerNorm(embed_dim)
         else:
             self.conv_norm = None
             self.conv_module = None
             self.final_norm = None
 
-        self.normalize_before = args.encoder_normalize_before
-        self.fc1 = self.build_fc1(
-            self.embed_dim,
-            args.encoder_ffn_embed_dim,
-            self.quant_noise,
-            self.quant_noise_block_size,
-        )
-        self.fc2 = self.build_fc2(
-            args.encoder_ffn_embed_dim,
-            self.embed_dim,
-            self.quant_noise,
-            self.quant_noise_block_size,
-        )
-
+        self.ffn = FeedForwardModule(
+                embed_dim,
+                ffn_dim,
+                dropout,
+                dropout,
+                activation
+            )
         self.ffn_norm = LayerNorm(self.embed_dim)
 
-    def build_fc1(self, input_dim, output_dim, q_noise, qn_block_size):
-        return quant_noise(
-            nn.Linear(input_dim, output_dim), p=q_noise, block_size=qn_block_size
-        )
-
-    def build_fc2(self, input_dim, output_dim, q_noise, qn_block_size):
-        return quant_noise(
-            nn.Linear(input_dim, output_dim), p=q_noise, block_size=qn_block_size
-        )
-
     def build_self_attention(self, embed_dim, args):
+        attention_heads = args.encoder_attention_heads
+        dropout = args.dropout
+
         if self.attn_type == "selfattn":
             attn_func = MultiheadAttention
         elif self.attn_type == "rel_selfattn":
@@ -125,8 +150,8 @@ class S2TTransformerEncoderLayer(nn.Module):
             if max_relative_length != -1:
                 return RelativeMultiheadAttention(
                     embed_dim,
-                    args.encoder_attention_heads,
-                    dropout=args.attention_dropout,
+                    attention_heads,
+                    dropout=dropout,
                     self_attention=True,
                     q_noise=self.quant_noise,
                     qn_block_size=self.quant_noise_block_size,
@@ -135,6 +160,25 @@ class S2TTransformerEncoderLayer(nn.Module):
             else:
                 print("The maximum encoder relative length %d can not be -1!" % max_relative_length)
                 exit(1)
+        elif self.attn_type == "rel_pos":
+            return RelPositionMultiHeadedAttention(
+                embed_dim,
+                attention_heads,
+                dropout=dropout,
+            )
+        elif self.attn_type == "rope":
+            return RotaryPositionMultiHeadedAttention(
+                embed_dim,
+                attention_heads,
+                dropout=dropout,
+                precision=args.fp16
+            )
+        elif self.attn_type == "abs":
+            return ESPNETMultiHeadedAttention(
+                embed_dim,
+                attention_heads,
+                dropout=dropout,
+            )
         else:
             attn_func = MultiheadAttention
             print("The attention type %s is not supported!" % self.attn_type)
@@ -191,23 +235,23 @@ class S2TTransformerEncoderLayer(nn.Module):
         # Note that we cannot use -inf here, because at some edge cases,
         # the attention weight (before softmax) for some padded element in query
         # will become -inf, which results in NaN in model parameters
-        if attn_mask is not None:
-            attn_mask = attn_mask.masked_fill(attn_mask.to(torch.bool), -1e8)
+        # if attn_mask is not None:
+        #     attn_mask = attn_mask.masked_fill(attn_mask.to(torch.bool), -1e8)
 
         # whether to use macaron style
         if self.macaron_norm is not None:
             residual = x
             if self.normalize_before:
                 x = self.macaron_norm(x)
-            x = self.macaron_fc2(self.activation_dropout_module(self.activation_fn(self.macaron_fc1(x))))
-            x = residual + self.ffn_scale * self.dropout_module(x)
+            x = self.macaron_ffn(x)
+            x = residual + self.ffn_scale * x
             if not self.normalize_before:
                 x = self.macaron_norm(x)
 
         residual = x
         if self.normalize_before:
             x = self.self_attn_layer_norm(x)
-        if self.attn_type == "rel_selfattn":
+        if self.attn_type == "rel_selfattn" or self.attn_type == "rel_pos":
             assert pos_emb is not None, "Positions is necessary for RPE!"
             x, _ = self.self_attn(
                 query=x,
@@ -234,22 +278,23 @@ class S2TTransformerEncoderLayer(nn.Module):
 
         # convolution module
         if self.conv_module is not None:
-            x = x.transpose(0, 1)
             residual = x
+
+            x = x.transpose(0, 1)
             if self.normalize_before:
                 x = self.conv_norm(x)
-            x = residual + self.dropout_module(self.conv_module(x, encoder_padding_mask))
+
+            x = self.conv_module(x)
+            x = x.transpose(0, 1)
+            x = residual + x
+
             if not self.normalize_before:
                 x = self.conv_norm(x)
-            x = x.transpose(0, 1)
 
         residual = x
         if self.normalize_before:
             x = self.ffn_norm(x)
-        x = self.activation_fn(self.fc1(x))
-        x = self.activation_dropout_module(x)
-        x = self.fc2(x)
-        x = self.dropout_module(x)
+        x = self.ffn(x)
         x = self.residual_connection(self.ffn_scale * x, residual)
         if not self.normalize_before:
             x = self.ffn_norm(x)

fairseq/modules/speech_to_text/subsampling.py

 import torch
 import torch.nn as nn
 
+from typing import List
+
 from fairseq.modules.activations import Swish
 from fairseq.modules.layer_norm import LayerNorm
 
@@ -46,6 +48,61 @@ def get_norm(norm_type, size, transpose=False):
         raise RuntimeError("normalization type {} not supported".format(norm_type))
 
 
+class Conv1dSubsampler(nn.Module):
+    """Convolutional subsampler: a stack of 1D convolution (along temporal
+    dimension) followed by non-linear activation via gated linear units
+    (https://arxiv.org/abs/1911.08460)
+
+    Args:
+        in_channels (int): the number of input channels
+        mid_channels (int): the number of intermediate channels
+        out_channels (int): the number of output channels
+        kernel_sizes (List[int]): the kernel size for each convolutional layer
+    """
+
+    def __init__(
+            self,
+            in_channels: int,
+            mid_channels: int,
+            out_channels: int,
+            kernel_sizes: List[int] = (3, 3),
+    ):
+        super(Conv1dSubsampler, self).__init__()
+        self.n_layers = len(kernel_sizes)
+        self.conv_layers = nn.ModuleList(
+            nn.Conv1d(
+                in_channels if i == 0 else mid_channels // 2,
+                mid_channels if i < self.n_layers - 1 else out_channels * 2,
+                k,
+                stride=2,
+                padding=k // 2,
+            )
+            for i, k in enumerate(kernel_sizes)
+        )
+
+    def get_out_seq_lens_tensor(self, in_seq_lens_tensor):
+        out = in_seq_lens_tensor.clone()
+        for _ in range(self.n_layers):
+            out = ((out.float() - 1) / 2 + 1).floor().long()
+        return out
+
+    def forward(self, src_tokens, src_lengths):
+        bsz, in_seq_len, _ = src_tokens.size()  # B x T x (C x D)
+        x = src_tokens.transpose(1, 2).contiguous()  # -> B x (C x D) x T
+        inner_x = []
+        for conv in self.conv_layers:
+            x = conv(x)
+            x = nn.functional.glu(x, dim=1)
+            inner_x.append(x)
+        _, _, out_seq_len = x.size()
+        # x = x.transpose(1, 2).transpose(0, 1).contiguous()  # -> T x B x (C x D)
+        out_inner_x = []
+        for x in inner_x:
+            out_inner_x.append(x.transpose(1, 2).transpose(0, 1).contiguous())
+        return out_inner_x, self.get_out_seq_lens_tensor(src_lengths)
+
+
+# fairseq style
 class Conv1dSubsampling(nn.Module):
     """Conv1d Subsampling Block
 
@@ -74,12 +131,14 @@ class Conv1dSubsampling(nn.Module):
 
         # Layers
         self.layers = nn.ModuleList([nn.Sequential(
-            nn.Conv1d(in_dim if layer_id == 0 else filters[layer_id - 1],
-                      filters[layer_id] * 2 if act == "glu" else filters[layer_id],
+            nn.Conv1d(in_dim if layer_id == 0 else filters[layer_id - 1] // 2 if act == "glu" else filters[layer_id - 1],
+                      filters[layer_id] * 2 if act == "glu" and layer_id == num_layers - 1 else filters[layer_id],
                       kernel_size,
                       stride=stride,
                       padding=(kernel_size - 1) // 2),
-            get_norm(norm, filters[layer_id], transpose=True if norm == "layer" else False),
+            get_norm(norm,
+                     filters[layer_id] * 2 if act == "glu" and layer_id == num_layers - 1 else filters[layer_id],
+                     transpose=True if norm == "layer" else False),
             get_activation_class(act, dim=1)
         ) for layer_id in range(num_layers)])
 
@@ -126,12 +185,14 @@ class Conv2dSubsampling(nn.Module):
         # Conv 2D Subsampling Layers
 
         self.layers = nn.ModuleList([nn.Sequential(
-            nn.Conv2d(1 if layer_id == 0 else filters[layer_id - 1],
-                      filters[layer_id] * 2 if act =="glu" else filters[layer_id],
+            nn.Conv2d(1 if layer_id == 0 else filters[layer_id - 1] // 2 if act == "glu" else filters[layer_id - 1],
+                      filters[layer_id] * 2 if act == "glu" and layer_id == num_layers - 1 else filters[layer_id],
                       kernel_size,
                       stride=stride,
                       padding=(kernel_size - 1) // 2),
-            get_norm(norm, filters[layer_id], transpose=True if norm == "layer" else False),
+            get_norm(norm,
+                     filters[layer_id] * 2 if act == "glu" and layer_id == num_layers - 1 else filters[layer_id],
+                     transpose=True if norm == "layer" else False),
             get_activation_class(act, dim=1)
         ) for layer_id in range(num_layers)])
         self.linear = nn.Linear(filters[-1] * in_dim // 2 ** num_layers, filters[-1])
@@ -139,7 +200,7 @@ class Conv2dSubsampling(nn.Module):
     def forward(self, x, x_len):
 
         # (B, T, D) -> (B, D, T) -> (B, 1, D, T)
-        x = x.tranpose(1, 2).unsqueeze(dim=1)
+        x = x.transpose(1, 2).unsqueeze(dim=1)
 
         # Layers
         for layer in self.layers:

fairseq/utils.py

@@ -17,7 +17,9 @@ from typing import Callable, Dict, List, Optional
 
 import torch
 import torch.nn.functional as F
+
 from fairseq.modules.multihead_attention import MultiheadAttention
+
 from torch import Tensor
 
 
@@ -514,6 +516,8 @@ def get_activation_fn(activation: str) -> Callable:
         return torch.tanh
     elif activation == "linear":
         return lambda x: x
+    elif activation == "swish":
+        return torch.nn.SiLU
     else:
         raise RuntimeError("--activation-fn {} not supported".format(activation))
 
@@ -526,6 +530,7 @@ def get_available_activation_fns() -> List:
         "gelu_accurate",
         "tanh",
         "linear",
+        "swish",
     ]