support share_all_embedding and share_decoder_input_and_softmax

scripts/convert_t2t_to_fairseq.py

@@ -30,7 +30,8 @@ def find_useful_param(model_file):
   global max_relative_length
   global normalize_before
   global use_dense
-
+  global share_all_embedding
+  global share_decoder_input_and_softmax
   trainable_param = dict()
   try:
     reader = pywrap_tensorflow.NewCheckpointReader(model_file)
@@ -85,6 +86,16 @@ def find_useful_param(model_file):
         if match and use_dense is False:
           use_dense = True
 
+        match = re.match('symbol_modality_(\d+)_(\d+)/shared_target',key)
+        if match and share_decoder_input_and_softmax is False:
+          share_decoder_input_and_softmax = True
+          tgt_vocab_size = int(match.group(1))
+
+        match = re.match('symbol_modality_(\d+)_(\d+)/shared/$', key)
+        if match and share_all_embedding is False:
+          share_all_embedding = True
+          tgt_vocab_size = int(match.group(1))
+
   except Exception as e:  # pylint: disable=broad-except
     print(str(e))
   assert len(trainable_param) > 0, "not found any trainable parameters"
@@ -99,9 +110,9 @@ def find_useful_param(model_file):
   if use_relative_position_representation is None:
     use_relative_position_representation = 0
 
-  print('find src-vocab:{} tgt-vocab:{} emb-size:{} src-layer:{} tgt-layer:{} activation:{} relative:{} normalize_before:{} use_dense:{}'.
+  print('find src-vocab:{} tgt-vocab:{} emb-size:{} src-layer:{} tgt-layer:{} activation:{} relative:{} normalize_before:{} use_dense:{} share_decoder_input_and_softmax:{} share_all_embedding:{}'.
         format(src_vocab_size, tgt_vocab_size, emb_size, src_layer_num, tgt_layer_num,
-               activation_function, use_relative_position_representation, normalize_before, use_dense))
+               activation_function, use_relative_position_representation, normalize_before, use_dense, share_decoder_input_and_softmax, share_all_embedding))
 
 
 
@@ -127,6 +138,10 @@ def load_param():
     return tensor
 
   # source embedding, shape is src_vocab * emb_size
+  if share_all_embedding:
+    src_emb_tensor = _get_tensor('symbol_modality_{}_{}/shared/weights'.format(src_vocab_size, emb_size),
+                                 shard=shard)
+  else:
     src_emb_tensor = _get_tensor('symbol_modality_{}_{}/input_emb/weights'.format(src_vocab_size, emb_size), shard=shard)
   param_dict['src_emb'] = src_emb_tensor
 
@@ -231,8 +246,12 @@ def load_param():
 
   # decoder
   # target embedding, shape is tgt_vocab * emb_size
-  tgt_emb_tensor = _get_tensor('symbol_modality_{}_{}/target_emb/weights'.format(tgt_vocab_size, emb_size),
-                               shard=shard)
+  if share_all_embedding:
+    tgt_emb_tensor = _get_tensor('symbol_modality_{}_{}/shared/weights'.format(tgt_vocab_size, emb_size),shard=shard)
+  elif share_decoder_input_and_softmax:
+    tgt_emb_tensor = _get_tensor('symbol_modality_{}_{}/shared_target/weights'.format(tgt_vocab_size, emb_size),shard=shard)
+  else:
+    tgt_emb_tensor = _get_tensor('symbol_modality_{}_{}/target_emb/weights'.format(tgt_vocab_size, emb_size),shard=shard)
   param_dict['tgt_emb'] = tgt_emb_tensor
   param_dict['tgt_pos_emb'] = get_pos_emb(tgt_max_length, emb_size=emb_size)
 
@@ -333,6 +352,14 @@ def load_param():
     layer_weight = _get_tensor('body/decoder/layer_history/layer_weight')
     param_dict['dec_layer_weight'] = layer_weight
 
+
+  if share_all_embedding:
+    softmax_w_tensor = _get_tensor('symbol_modality_{}_{}/shared/weights'.format(tgt_vocab_size, emb_size),
+                                   shard=shard)
+  if share_decoder_input_and_softmax:
+    softmax_w_tensor = _get_tensor('symbol_modality_{}_{}/shared_target/weights'.format(tgt_vocab_size, emb_size),
+                                   shard=shard)
+  else:
     softmax_w_tensor = _get_tensor('symbol_modality_{}_{}/softmax/weights'.format(tgt_vocab_size, emb_size),
                                  shard=shard)
   # note: we transpose the matrix, from (V,H) to (H,V)
@@ -421,6 +448,11 @@ def convert_settings(settings):
     args['decoder_normalize_before'] = True
     args['attention_dropout'] = 0.1
     args['relu_dropout'] = 0.1
+  assert share_all_embedding ^ share_decoder_input_and_softmax
+  if share_all_embedding:
+    args['share_all_embeddings'] = True
+  if share_decoder_input_and_softmax:
+    args['share_decoder_input_output_embed'] = True
 
   return argparse.Namespace(**args)
 
@@ -690,6 +722,8 @@ if __name__ == '__main__':
   max_relative_length = -1
   normalize_before = False
   use_dense = False
+  share_all_embedding=False
+  share_decoder_input_and_softmax=False
 
   start = time.time()
 

train.py

@@ -52,6 +52,8 @@ def main(args):
 
     # Build model and criterion
     model = task.build_model(args)
+    # for k,v in model.named_parameters():
+    #     print("k:%s"%k)
     print(model)
     criterion = task.build_criterion(args)
     print('| model {}, criterion {}'.format(args.arch, criterion.__class__.__name__))