fix some bugs in trans_dict.py

model_zoo/research/nlp/gpt2/convert_tf_ckpt/trans_dict.py

@@ -446,8 +446,8 @@ trans_dict_tf = {
     'h35.mlp.c_fc.w': 'gpt2_decoder.layers.35.feedforward.c_fc.weight',
     'h35.mlp.c_proj.b': 'gpt2_decoder.layers.35.feedforward.c_proj.bias',
     'h35.mlp.c_proj.w': 'gpt2_decoder.layers.35.feedforward.c_proj.weight',
-    'ln_f.b': 'layer_norm.layer_norm.gamma',
-    'ln_f.g': 'layer_norm.layer_norm.beta',
+    'ln_f.b': 'layer_norm.layer_norm.beta',
+    'ln_f.g': 'layer_norm.layer_norm.gamma',
     'wpe': 'gpt2_embedding_postprocess.position_embedding_table',
     'wte': 'gpt2_embedding_lookup.embedding_table'
 }  # transfer dictionary

model_zoo/research/nlp/gpt2/create_cbt_data.py

@@ -89,7 +89,7 @@ def main():
     parser.add_argument("--output_file", type=str, required=True, default="", help='Output MindRecord file. ')
     parser.add_argument("--num_splits", type=int, default=1,
                         help='The MindRecord file will be split into the number of partition. ')
-    parser.add_argument("--max_seq_length", type=int, required=True, help='Maximum sequence length. ')
+    parser.add_argument("--max_length", type=int, required=True, help='Maximum sequence length. ')
     parser.add_argument("--num_choice", type=int, required=True, help='Number of choices. ')
     parser.add_argument("--vocab_file", type=str, required=True, default='', help='url of gpt2-vocab.json ')
     parser.add_argument("--merge_file", type=str, required=True, default='', help='url of gpt2-merges.txt ')
@@ -127,7 +127,7 @@ def main():
             if total_read % 500 == 0:
                 logging.info("%d ...", total_read)
 
-            output = create_instance(tokenizer, line, args.max_seq_length, num_choice)
+            output = create_instance(tokenizer, line, args.max_length, num_choice)
             features = write_instance_to_file(writer, instance=output)
             total_written += 1
 

model_zoo/research/nlp/gpt2/create_lambada_data.py

@@ -84,7 +84,7 @@ def main():
     parser.add_argument("--output_file", type=str, required=True, help='Output MindRecord file. ')
     parser.add_argument("--num_splits", type=int, default=1,
                         help='The MindRecord file will be split into the number of partition. ')
-    parser.add_argument("--max_seq_length", type=int, required=True, help='Maximum sequence length. ')
+    parser.add_argument("--max_length", type=int, required=True, help='Maximum sequence length. ')
     parser.add_argument("--vocab_file", type=str, required=True, default='', help='url of gpt2-vocab.json ')
     parser.add_argument("--merge_file", type=str, required=True, default='', help='url of gpt2-merges.txt ')
     args = parser.parse_args()
@@ -119,7 +119,7 @@ def main():
             if total_read % 500 == 0:
                 logging.info("%d ...", total_read)
 
-            output = create_instance(tokenizer, line, args.max_seq_length)
+            output = create_instance(tokenizer, line, args.max_length)
             features = write_instance_to_file(writer, instance=output)
             total_written += 1
 

model_zoo/research/nlp/gpt2/create_lm_data.py

@@ -70,7 +70,7 @@ def main():
     parser.add_argument("--output_file", type=str, required=True, help='Output MindRecord file. ')
     parser.add_argument("--num_splits", type=int, default=1,
                         help='The MindRecord file will be split into the number of partition. ')
-    parser.add_argument("--max_seq_length", type=int, required=True, help='Maximum sequence length. ')
+    parser.add_argument("--max_length", type=int, required=True, help='Maximum sequence length. ')
     parser.add_argument("--vocab_file", type=str, required=True, default='', help='url of gpt2-vocab.json ')
     parser.add_argument("--merge_file", type=str, required=True, default='', help='url of gpt2-merges.txt ')
     args = parser.parse_args()
@@ -105,7 +105,7 @@ def main():
             if total_read % 500 == 0:
                 logging.info("%d ...", total_read)
 
-            output = create_instance(tokenizer, line, args.max_seq_length)
+            output = create_instance(tokenizer, line, args.max_length)
             features = write_instance_to_file(writer, instance=output)
             total_written += 1
 

model_zoo/research/nlp/gpt2/create_summary_data.py

@@ -74,7 +74,7 @@ def main():
     parser.add_argument("--output_file", type=str, required=True, help='Output MindRecord file.')
     parser.add_argument("--num_splits", type=int, default=1,
                         help='The MindRecord file will be split into the number of partition. ')
-    parser.add_argument("--max_seq_length", type=int, required=True, help='Maximum sequence length.')
+    parser.add_argument("--max_length", type=int, required=True, help='Maximum sequence length.')
     parser.add_argument("--vocab_file", type=str, required=True, default='', help='url of gpt2-vocab.json ')
     parser.add_argument("--merge_file", type=str, required=True, default='', help='url of gpt2-merges.txt ')
     parser.add_argument("--mode", type=str, required=True, default='cnn_dailymail', help='mode of dataset creation')
@@ -109,7 +109,7 @@ def main():
             if total_read % 500 == 0:
                 logging.info("%d ...", total_read)
 
-            output = create_instance(tokenizer, line, args.max_seq_length)
+            output = create_instance(tokenizer, line, args.max_length)
             features = write_instance_to_file(writer, instance=output)
             total_written += 1
 

model_zoo/research/nlp/gpt2/run_CBT_task.py

@@ -32,6 +32,7 @@ from src.utils.metric_method import Accuracy
 from src.dataset import create_cbt_dataset, create_language_model_dataset
 from src.utils.lr_schedule import GPT2LearningRate
 from src.utils.task_utils import calculate_choice_prob_for_cbt
+from src.utils.get_config_setting import get_train_setting, get_model_setting
 
 
 def do_train(dataset=None, network=None, load_checkpoint_path="", save_checkpoint_path="", epoch_num=1):
@@ -248,6 +249,8 @@ def run_cbt_task():
                         use_one_hot_embeddings=False)
 
     if args_opt.do_train.lower() == "true":
+        get_train_setting(cfg)
+        get_model_setting(cfg, gpt2_net_cfg)
         print("==============    Start Loading Train Dataset   ============")
         print(" | Train Dataset: {}".format(args_opt.train_data_file_path))
         print(" | Checkpoint: {}".format(args_opt.load_pretrain_ckpt_path))
@@ -256,6 +259,7 @@ def run_cbt_task():
         do_train(train_dataset, gpt2_loss, load_pretrain_ckpt_path, save_finetune_ckpt_path, epoch_num)
 
     if args_opt.do_eval.lower() == "true":
+        get_model_setting(cfg, gpt2_net_cfg)
         print("============== Start Loading Evaluation Dataset ============")
         print(" | Eval Dataset: {}".format(args_opt.eval_data_file_path))
         print(" | Checkpoint: {}".format(args_opt.load_finetune_ckpt_path))

model_zoo/research/nlp/gpt2/run_ReadComprehension.py

@@ -33,6 +33,7 @@ from src.dataset import create_language_model_dataset
 from src.utils.lr_schedule import GPT2LearningRate
 from src.utils.tokenization import Tokenizer
 from src.GPT2_generation import GenerateForReadComprehension
+from src.utils.get_config_setting import get_train_setting, get_model_setting
 
 
 def do_train(dataset=None, network=None, load_checkpoint_path="", save_checkpoint_path="", epoch_num=1):
@@ -121,7 +122,7 @@ def do_eval(dataset=None, network=None, metric=None, load_checkpoint_path="", ev
     if load_checkpoint_path == "":
         raise ValueError("Finetune model missed, evaluation task must load finetune model!")
     if metric.lower() == "f1":
-        print("Prepare to calculate the BLEU score ...")
+        print("Prepare to calculate the F1 score ...")
 
         gpt2_rc = network(config=gpt2_net_cfg,
                           is_training=False,
@@ -269,6 +270,8 @@ def run_Readcomprehension():
                          use_one_hot_embeddings=False)
 
     if args_opt.do_train.lower() == "true":
+        get_train_setting(cfg)
+        get_model_setting(cfg, gpt2_net_cfg)
         print("==============   Start Loading Translation Train Dataset   ==============")
         print(" | Train Dataset: {}".format(args_opt.train_data_file_path))
         print(" | Checkpoint: {}".format(args_opt.load_pretrain_ckpt_path))
@@ -277,6 +280,7 @@ def run_Readcomprehension():
         do_train(train_dataset, gpt2_loss, load_pretrain_ckpt_path, save_finetune_ckpt_path, epoch_num)
 
     if args_opt.do_eval.lower() == "true":
+        get_model_setting(cfg, gpt2_net_cfg)
         print("============   Start Loading Translation Evaluation Dataset  ============")
         print(" | Eval Dataset: {}".format(args_opt.eval_data_file_path))
         print(" | Checkpoint: {}".format(args_opt.load_finetune_ckpt_path))

model_zoo/research/nlp/gpt2/run_summarization.py

@@ -36,6 +36,7 @@ from src.utils.lr_schedule import GPT2LearningRate
 from src.utils.tokenization import Tokenizer
 from src.utils.task_utils import clean_hypo, modify_paramdict
 from src.GPT2_generation import GenerateForSummarization
+from src.utils.get_config_setting import get_train_setting, get_model_setting
 
 
 def do_train(dataset=None, network=None, load_checkpoint_path="", save_checkpoint_path="", epoch_num=1):
@@ -272,6 +273,8 @@ def run_summarization():
         raise Exception("Device target error, Ascend is supported.")
 
     if args_opt.do_train.lower() == "true":
+        get_train_setting(cfg)
+        get_model_setting(cfg, gpt2_net_cfg)
         train_data_file_path = args_opt.train_data_file_path
         gpt2_loss = GPT2Summarization(config=gpt2_net_cfg,
                                       is_training=True,
@@ -282,6 +285,7 @@ def run_summarization():
         do_train(train_dataset, gpt2_loss, load_pretrain_ckpt_path, save_finetune_ckpt_path, epoch_num)
 
     if args_opt.do_eval.lower() == "true":
+        get_model_setting(cfg, gpt2_net_cfg)
         eval_dataset_file_path = args_opt.eval_data_file_path
         print("============== Start Loading Evaluation Dataset ============")
         eval_dataset = create_language_model_dataset(do_shuffle=(args_opt.train_data_shuffle.lower() == "true"),

model_zoo/research/nlp/gpt2/run_translation.py

@@ -33,6 +33,7 @@ from src.utils.lr_schedule import GPT2LearningRate
 from src.utils.tokenization import Tokenizer
 from src.utils.metric_method import BLEU
 from src.GPT2_generation import GenerateForTranslation
+from src.utils.get_config_setting import get_train_setting, get_model_setting
 
 
 def do_train(dataset=None, network=None, load_checkpoint_path="", save_checkpoint_path="", epoch_num=1):
@@ -274,6 +275,8 @@ def run_translation():
                                 use_one_hot_embeddings=False)
 
     if args_opt.do_train.lower() == "true":
+        get_train_setting(cfg)
+        get_model_setting(cfg, gpt2_net_cfg)
         print("==============   Start Loading Translation Train Dataset   ==============")
         print(" | Train Dataset: {}".format(args_opt.train_data_file_path))
         print(" | Checkpoint: {}".format(args_opt.load_pretrain_ckpt_path))
@@ -282,6 +285,7 @@ def run_translation():
         do_train(train_dataset, gpt2_loss, load_pretrain_ckpt_path, save_finetune_ckpt_path, epoch_num)
 
     if args_opt.do_eval.lower() == "true":
+        get_model_setting(cfg, gpt2_net_cfg)
         print("============   Start Loading Translation Evaluation Dataset  ============")
         print(" | Eval Dataset: {}".format(args_opt.eval_data_file_path))
         print(" | Checkpoint: {}".format(args_opt.load_finetune_ckpt_path))

model_zoo/research/nlp/gpt2/src/utils/tensor_manipulations.py

@@ -56,7 +56,7 @@ def extract_string_from_tensor(input_ids, mode="single", config=None, tokenizer=
 
         for batch_idx in range(batch_size):
             sentence_tensor = input_ids_np[batch_idx]
-            sentence_list = sentence_tensor.asnumpy().tolist()[1:]
+            sentence_list = sentence_tensor.tolist()[1:]
 
             sentence = tokenizer.decode(sentence_list)
             prompt_start = 0
@@ -73,7 +73,7 @@ def extract_string_from_tensor(input_ids, mode="single", config=None, tokenizer=
     if mode == "single":
         for batch_idx in range(batch_size):
             sentence_tensor = input_ids_np[batch_idx]
-            sentence_list = sentence_tensor.asnumpy().tolist()[1:]
+            sentence_list = sentence_tensor.tolist()[1:]
 
             sentence = tokenizer.decode(sentence_list)
             prompt_start = 0