add model hparams instruction for transformer and transformer_dla

tensor2tensor/models/transformer.py

@@ -807,6 +807,7 @@ def transformer_base_v2():
   hparams.relu_dropout = 0.1
   return hparams
 
+
 @registry.register_hparams
 def transformer_base_rpr_dropout1():
   hparams = transformer_base()
@@ -834,14 +835,35 @@ def transformer_big_multistep2():
     hparams.optimizer = "MultistepAdam"
     hparams.optimizer_multistep_accumulate_steps = 2
     hparams.batch_size = 2048
-    #hparams.attention_dropout = 0.1
+    hparams.attention_dropout = 0.1
     #hparams.relu_dropout = 0.1
     return hparams
 
 @registry.register_hparams
-def transformer_before_test():
+def transformer_before_shared():
     # new model use optimizer MultistepAdam
     hparams = transformer_before()
-    hparams.shared_decoder_input_and_softmax_weights = int(False)
+    hparams.shared_decoder_input_and_softmax_weights = int(True)
+    return hparams
 
+
+@registry.register_hparams
+def transformer_before_shared25():
+    # new model use optimizer MultistepAdam to train Deep Transformer-pre-norm
+    # we found that the model can achieve better performance when we set the true update batch to 8192 and revise
+    # the learning rate to 0.4. Then the max learning rate you get is nearly to 0.002, it is worth noting that the real learning
+    # rate is varying according to the global step , affected by "learning_rate" and "learning_rate_warmup_steps".
+    hparams = transformer_before_shared()
+    hparams.learning_rate = 0.4
+    hparams.learning_rate_warmup_steps = 8000
+    hparams.optimizer = "MultistepAdam"
+    hparams.optimizer_multistep_accumulate_steps = 4
+    # it's likely to oom when you train deep transformer-pre-norm within 4096 batch_size
+    hparams.batch_size = 2048
     return hparams
+
+
+# you can define your own hparams like above to fix the target task
+# @registry.register_hparams
+# def transformer_new****():
+# return hparams
\ No newline at end of file

tensor2tensor/models/transformer_dla.py

@@ -437,6 +437,9 @@ def transformer_dla():
 
 @registry.register_hparams
 def transformer_dla_base():
+  # we implement the Dynamic linear combination of layers in Transformer, we choose the "learnable dense" schema and
+  # we reviese the learning rate to 0.4 , warmup step to 8000 based Transformer-Pre-Norm . Due to the much more representation
+  # than baseline's, we reset the batch size to 2048, but four steps per update instead.
   hparams = transformer_dla()
   hparams.encoder_layers = 6
   hparams.decoder_layers = 6
@@ -452,9 +455,9 @@ def transformer_dla_base():
   return hparams
 
 
-
 @registry.register_hparams
 def transformer_dla_base25_shared():
+  # share the decoder input and the softmax embedding
   hparams = transformer_dla_base()
   hparams.shared_decoder_input_and_softmax_weights = int(True)
   hparams.encoder_layers = 25
@@ -462,7 +465,33 @@ def transformer_dla_base25_shared():
 
 
 @registry.register_hparams
+def transformer_dla_base30_shared():
+  hparams = transformer_dla_base()
+  hparams.shared_decoder_input_and_softmax_weights = int(True)
+  hparams.encoder_layers = 30
+  return hparams
+
+
+@registry.register_hparams
+def transformer_dla_base25_shared_filter4096():
+  # only enlarge the filter size
+  hparams = transformer_dla_base25_shared()
+  hparams.filter_size = 4096
+  return hparams
+
+
+@registry.register_hparams
+def transformer_dla_base25_shared_big():
+  # enlarge the hidden size and filter size, keep the relation of filter_size = 4 * hidden_size
+  hparams = transformer_dla_base25_shared()
+  hparams.hidden_size = 768
+  hparams.filter_size = 3072
+  return hparams
+
+
+@registry.register_hparams
 def transformer_dla_base_v2():
+  # in our paper, we found the larger warmup can even performer better in En-De task
   hparams = transformer_dla_base()
   hparams.learning_rate = 0.4 * (2**0.5)
   hparams.learning_rate_warmup_steps = 16000