move batch_size from bert_cfg_cfg to cfg

model_zoo/official/nlp/bert/README.md

@@ -312,6 +312,7 @@ Parameters for training and evaluation can be set in file `config.py` and `finet
 ```
 config for lossscale and etc.
     bert_network                    version of BERT model: base | nezha, default is base
+    batch_size                      batch size of input dataset: N, default is 16
     loss_scale_value                initial value of loss scale: N, default is 2^32
     scale_factor                    factor used to update loss scale: N, default is 2
     scale_window                    steps for once updatation of loss scale: N, default is 1000   
@@ -321,7 +322,6 @@ config for lossscale and etc.
 ### Parameters:
 ```
 Parameters for dataset and network (Pre-Training/Fine-Tuning/Evaluation):
-    batch_size                      batch size of input dataset: N, default is 16
     seq_length                      length of input sequence: N, default is 128
     vocab_size                      size of each embedding vector: N, must be consistant with the dataset you use. Default is 21136
     hidden_size                     size of bert encoder layers: N, default is 768
@@ -335,8 +335,6 @@ Parameters for dataset and network (Pre-Training/Fine-Tuning/Evaluation):
     type_vocab_size                 size of token type vocab: N, default is 16
     initializer_range               initialization value of TruncatedNormal: Q, default is 0.02
     use_relative_positions          use relative positions or not: True | False, default is False
-    input_mask_from_dataset         use the input mask loaded form dataset or not: True | False, default is True
-    token_type_ids_from_dataset     use the token type ids loaded from dataset or not: True | False, default is True
     dtype                           data type of input: mstype.float16 | mstype.float32, default is mstype.float32
     compute_type                    compute type in BertTransformer: mstype.float16 | mstype.float32, default is mstype.float16
 

model_zoo/official/nlp/bert/mindspore_hub_conf.py

@@ -19,7 +19,6 @@ from src.bert_model import BertModel
 from src.bert_model import BertConfig
 import mindspore.common.dtype as mstype
 bert_net_cfg_base = BertConfig(
-    batch_size=32,
     seq_length=128,
     vocab_size=21128,
     hidden_size=768,
@@ -33,13 +32,10 @@ bert_net_cfg_base = BertConfig(
     type_vocab_size=2,
     initializer_range=0.02,
     use_relative_positions=False,
-    input_mask_from_dataset=True,
-    token_type_ids_from_dataset=True,
     dtype=mstype.float32,
     compute_type=mstype.float16
 )
 bert_net_cfg_nezha = BertConfig(
-    batch_size=32,
     seq_length=128,
     vocab_size=21128,
     hidden_size=1024,
@@ -53,8 +49,6 @@ bert_net_cfg_nezha = BertConfig(
     type_vocab_size=2,
     initializer_range=0.02,
     use_relative_positions=True,
-    input_mask_from_dataset=True,
-    token_type_ids_from_dataset=True,
     dtype=mstype.float32,
     compute_type=mstype.float16
 )
@@ -63,15 +57,11 @@ def create_network(name, *args, **kwargs):
     Create bert network for base and nezha.
     '''
     if name == 'bert_base':
-        if "batch_size" in kwargs:
-            bert_net_cfg_base.batch_size = kwargs["batch_size"]
         if "seq_length" in kwargs:
             bert_net_cfg_base.seq_length = kwargs["seq_length"]
         is_training = kwargs.get("is_training", default=False)
         return BertModel(bert_net_cfg_base, is_training, *args)
     if name == 'bert_nezha':
-        if "batch_size" in kwargs:
-            bert_net_cfg_nezha.batch_size = kwargs["batch_size"]
         if "seq_length" in kwargs:
             bert_net_cfg_nezha.seq_length = kwargs["seq_length"]
         is_training = kwargs.get("is_training", default=False)

model_zoo/official/nlp/bert/pretrain_eval.py

@@ -131,7 +131,7 @@ def bert_predict():
     '''
     devid = int(os.getenv('DEVICE_ID'))
     context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", device_id=devid)
-    dataset = get_enwiki_512_dataset(bert_net_cfg.batch_size, 1)
+    dataset = get_enwiki_512_dataset(cfg.batch_size, 1)
     net_for_pretraining = BertPretrainEva(bert_net_cfg)
     net_for_pretraining.set_train(False)
     param_dict = load_checkpoint(cfg.finetune_ckpt)

model_zoo/official/nlp/bert/run_classifier.py

@@ -188,7 +188,7 @@ def run_classifier():
                           assessment_method=assessment_method)
 
     if args_opt.do_train.lower() == "true":
-        ds = create_classification_dataset(batch_size=bert_net_cfg.batch_size, repeat_count=1,
+        ds = create_classification_dataset(batch_size=optimizer_cfg.batch_size, repeat_count=1,
                                            assessment_method=assessment_method,
                                            data_file_path=args_opt.train_data_file_path,
                                            schema_file_path=args_opt.schema_file_path,
@@ -204,7 +204,7 @@ def run_classifier():
                                                            ds.get_dataset_size(), epoch_num, "classifier")
 
     if args_opt.do_eval.lower() == "true":
-        ds = create_classification_dataset(batch_size=bert_net_cfg.batch_size, repeat_count=1,
+        ds = create_classification_dataset(batch_size=optimizer_cfg.batch_size, repeat_count=1,
                                            assessment_method=assessment_method,
                                            data_file_path=args_opt.eval_data_file_path,
                                            schema_file_path=args_opt.schema_file_path,

model_zoo/official/nlp/bert/run_ner.py

@@ -104,9 +104,9 @@ def do_eval(dataset=None, network=None, use_crf="", num_class=2, assessment_meth
     if load_checkpoint_path == "":
         raise ValueError("Finetune model missed, evaluation task must load finetune model!")
     if assessment_method == "clue_benchmark":
-        bert_net_cfg.batch_size = 1
-    net_for_pretraining = network(bert_net_cfg, False, num_class, use_crf=(use_crf.lower() == "true"),
-                                  tag_to_index=tag_to_index)
+        optimizer_cfg.batch_size = 1
+    net_for_pretraining = network(bert_net_cfg, optimizer_cfg.batch_size, False, num_class,
+                                  use_crf=(use_crf.lower() == "true"), tag_to_index=tag_to_index)
     net_for_pretraining.set_train(False)
     param_dict = load_checkpoint(load_checkpoint_path)
     load_param_into_net(net_for_pretraining, param_dict)
@@ -211,11 +211,11 @@ def run_ner():
         number_labels = len(tag_to_index)
     else:
         number_labels = args_opt.num_class
-    netwithloss = BertNER(bert_net_cfg, True, num_labels=number_labels,
+    netwithloss = BertNER(bert_net_cfg, optimizer_cfg.batch_size, True, num_labels=number_labels,
                           use_crf=(args_opt.use_crf.lower() == "true"),
                           tag_to_index=tag_to_index, dropout_prob=0.1)
     if args_opt.do_train.lower() == "true":
-        ds = create_ner_dataset(batch_size=bert_net_cfg.batch_size, repeat_count=1,
+        ds = create_ner_dataset(batch_size=optimizer_cfg.batch_size, repeat_count=1,
                                 assessment_method=assessment_method, data_file_path=args_opt.train_data_file_path,
                                 schema_file_path=args_opt.schema_file_path,
                                 do_shuffle=(args_opt.train_data_shuffle.lower() == "true"))

model_zoo/official/nlp/bert/run_pretrain.py

@@ -107,7 +107,7 @@ def run_pretrain():
 
     if args_opt.accumulation_steps > 1:
         logger.info("accumulation steps: {}".format(args_opt.accumulation_steps))
-        logger.info("global batch size: {}".format(bert_net_cfg.batch_size * args_opt.accumulation_steps))
+        logger.info("global batch size: {}".format(cfg.batch_size * args_opt.accumulation_steps))
         if args_opt.enable_data_sink == "true":
             args_opt.data_sink_steps *= args_opt.accumulation_steps
             logger.info("data sink steps: {}".format(args_opt.data_sink_steps))

model_zoo/official/nlp/bert/run_squad.py

@@ -123,7 +123,7 @@ def do_eval(dataset=None, vocab_file="", eval_json="", load_checkpoint_path="",
         start = logits[1].asnumpy()
         end = logits[2].asnumpy()
 
-        for i in range(bert_net_cfg.batch_size):
+        for i in range(optimizer_cfg.batch_size):
             unique_id = int(ids[i])
             start_logits = [float(x) for x in start[i].flat]
             end_logits = [float(x) for x in end[i].flat]
@@ -193,7 +193,7 @@ def run_squad():
     netwithloss = BertSquad(bert_net_cfg, True, 2, dropout_prob=0.1)
 
     if args_opt.do_train.lower() == "true":
-        ds = create_squad_dataset(batch_size=bert_net_cfg.batch_size, repeat_count=1,
+        ds = create_squad_dataset(batch_size=optimizer_cfg.batch_size, repeat_count=1,
                                   data_file_path=args_opt.train_data_file_path,
                                   schema_file_path=args_opt.schema_file_path,
                                   do_shuffle=(args_opt.train_data_shuffle.lower() == "true"))
@@ -207,7 +207,7 @@ def run_squad():
                                                            ds.get_dataset_size(), epoch_num, "squad")
 
     if args_opt.do_eval.lower() == "true":
-        ds = create_squad_dataset(batch_size=bert_net_cfg.batch_size, repeat_count=1,
+        ds = create_squad_dataset(batch_size=optimizer_cfg.batch_size, repeat_count=1,
                                   data_file_path=args_opt.eval_data_file_path,
                                   schema_file_path=args_opt.schema_file_path, is_training=False,
                                   do_shuffle=(args_opt.eval_data_shuffle.lower() == "true"))

model_zoo/official/nlp/bert/src/bert_for_finetune.py

@@ -274,15 +274,15 @@ class BertNER(nn.Cell):
     """
     Train interface for sequence labeling finetuning task.
     """
-    def __init__(self, config, is_training, num_labels=11, use_crf=False, tag_to_index=None, dropout_prob=0.0,
-                 use_one_hot_embeddings=False):
+    def __init__(self, config, batch_size, is_training, num_labels=11, use_crf=False,
+                 tag_to_index=None, dropout_prob=0.0, use_one_hot_embeddings=False):
         super(BertNER, self).__init__()
         self.bert = BertNERModel(config, is_training, num_labels, use_crf, dropout_prob, use_one_hot_embeddings)
         if use_crf:
             if not tag_to_index:
                 raise Exception("The dict for tag-index mapping should be provided for CRF.")
             from src.CRF import CRF
-            self.loss = CRF(tag_to_index, config.batch_size, config.seq_length, is_training)
+            self.loss = CRF(tag_to_index, batch_size, config.seq_length, is_training)
         else:
             self.loss = CrossEntropyCalculation(is_training)
         self.num_labels = num_labels

model_zoo/official/nlp/bert/src/bert_for_pre_training.py

@@ -92,9 +92,8 @@ class GetMaskedLMOutput(nn.Cell):
         self.matmul = P.MatMul(transpose_b=True)
         self.log_softmax = nn.LogSoftmax(axis=-1)
         self.shape_flat_offsets = (-1, 1)
-        self.rng = Tensor(np.array(range(0, config.batch_size)).astype(np.int32))
         self.last_idx = (-1,)
-        self.shape_flat_sequence_tensor = (config.batch_size * config.seq_length, self.width)
+        self.shape_flat_sequence_tensor = (-1, self.width)
         self.seq_length_tensor = Tensor(np.array((config.seq_length,)).astype(np.int32))
         self.cast = P.Cast()
         self.compute_type = config.compute_type
@@ -105,8 +104,8 @@ class GetMaskedLMOutput(nn.Cell):
                   output_weights,
                   positions):
         """Get output log_probs"""
-        flat_offsets = self.reshape(
-            self.rng * self.seq_length_tensor, self.shape_flat_offsets)
+        rng = F.tuple_to_array(F.make_range(P.Shape()(input_tensor)[0]))
+        flat_offsets = self.reshape(rng * self.seq_length_tensor, self.shape_flat_offsets)
         flat_position = self.reshape(positions + flat_offsets, self.last_idx)
         flat_sequence_tensor = self.reshape(input_tensor, self.shape_flat_sequence_tensor)
         input_tensor = self.gather(flat_sequence_tensor, flat_position, 0)

model_zoo/official/nlp/bert/src/config.py

@@ -19,6 +19,7 @@ from easydict import EasyDict as edict
 import mindspore.common.dtype as mstype
 from .bert_model import BertConfig
 cfg = edict({
+    'batch_size': 32,
     'bert_network': 'base',
     'loss_scale_value': 65536,
     'scale_factor': 2,
@@ -57,7 +58,6 @@ large: BERT-NEZHA(a Chinese pretrained language model developed by Huawei, which
 '''
 if cfg.bert_network == 'base':
     bert_net_cfg = BertConfig(
-        batch_size=64,
         seq_length=128,
         vocab_size=21128,
         hidden_size=768,
@@ -71,14 +71,11 @@ if cfg.bert_network == 'base':
         type_vocab_size=2,
         initializer_range=0.02,
         use_relative_positions=False,
-        input_mask_from_dataset=True,
-        token_type_ids_from_dataset=True,
         dtype=mstype.float32,
         compute_type=mstype.float16
     )
 if cfg.bert_network == 'nezha':
     bert_net_cfg = BertConfig(
-        batch_size=96,
         seq_length=128,
         vocab_size=21128,
         hidden_size=1024,
@@ -92,14 +89,11 @@ if cfg.bert_network == 'nezha':
         type_vocab_size=2,
         initializer_range=0.02,
         use_relative_positions=True,
-        input_mask_from_dataset=True,
-        token_type_ids_from_dataset=True,
         dtype=mstype.float32,
         compute_type=mstype.float16
     )
 if cfg.bert_network == 'large':
     bert_net_cfg = BertConfig(
-        batch_size=24,
         seq_length=512,
         vocab_size=30522,
         hidden_size=1024,
@@ -113,8 +107,6 @@ if cfg.bert_network == 'large':
         type_vocab_size=2,
         initializer_range=0.02,
         use_relative_positions=False,
-        input_mask_from_dataset=True,
-        token_type_ids_from_dataset=True,
         dtype=mstype.float32,
         compute_type=mstype.float16
     )

model_zoo/official/nlp/bert/src/dataset.py

@@ -20,7 +20,7 @@ import mindspore.common.dtype as mstype
 import mindspore.dataset.engine.datasets as de
 import mindspore.dataset.transforms.c_transforms as C
 from mindspore import log as logger
-from .config import bert_net_cfg
+from .config import cfg
 
 
 def create_bert_dataset(device_num=1, rank=0, do_shuffle="true", data_dir=None, schema_dir=None):
@@ -46,7 +46,7 @@ def create_bert_dataset(device_num=1, rank=0, do_shuffle="true", data_dir=None,
     ds = ds.map(operations=type_cast_op, input_columns="input_mask")
     ds = ds.map(operations=type_cast_op, input_columns="input_ids")
     # apply batch operations
-    ds = ds.batch(bert_net_cfg.batch_size, drop_remainder=True)
+    ds = ds.batch(cfg.batch_size, drop_remainder=True)
     logger.info("data size: {}".format(ds.get_dataset_size()))
     logger.info("repeat count: {}".format(ds.get_repeat_count()))
     return ds

model_zoo/official/nlp/bert/src/finetune_eval_config.py

@@ -22,6 +22,7 @@ import mindspore.common.dtype as mstype
 from .bert_model import BertConfig
 
 optimizer_cfg = edict({
+    'batch_size': 16,
     'optimizer': 'Lamb',
     'AdamWeightDecay': edict({
         'learning_rate': 2e-5,
@@ -45,7 +46,6 @@ optimizer_cfg = edict({
 })
 
 bert_net_cfg = BertConfig(
-    batch_size=16,
     seq_length=128,
     vocab_size=21128,
     hidden_size=768,
@@ -59,8 +59,6 @@ bert_net_cfg = BertConfig(
     type_vocab_size=2,
     initializer_range=0.02,
     use_relative_positions=False,
-    input_mask_from_dataset=True,
-    token_type_ids_from_dataset=True,
     dtype=mstype.float32,
     compute_type=mstype.float16,
 )

model_zoo/official/nlp/bert/src/finetune_eval_model.py

@@ -107,7 +107,7 @@ class BertNERModel(nn.Cell):
         self.reshape = P.Reshape()
         self.shape = (-1, config.hidden_size)
         self.use_crf = use_crf
-        self.origin_shape = (config.batch_size, config.seq_length, self.num_labels)
+        self.origin_shape = (-1, config.seq_length, self.num_labels)
 
     def construct(self, input_ids, input_mask, token_type_id):
         """Return the final logits as the results of log_softmax."""

tests/st/networks/models/bert/bert_performance/test_bert_tdt_lossscale.py

@@ -41,11 +41,10 @@ DATA_DIR = ["/home/workspace/mindspore_dataset/bert/example/examples.tfrecord"]
 SCHEMA_DIR = "/home/workspace/mindspore_dataset/bert/example/datasetSchema.json"
 
 
-def get_config(version='base', batch_size=1):
+def get_config(version='base'):
     """get config"""
     if version == 'base':
         bert_config = BertConfig(
-            batch_size=batch_size,
             seq_length=128,
             vocab_size=21136,
             hidden_size=768,
@@ -59,13 +58,10 @@ def get_config(version='base', batch_size=1):
             type_vocab_size=2,
             initializer_range=0.02,
             use_relative_positions=True,
-            input_mask_from_dataset=True,
-            token_type_ids_from_dataset=True,
             dtype=mstype.float32,
             compute_type=mstype.float32)
     elif version == 'large':
         bert_config = BertConfig(
-            batch_size=batch_size,
             seq_length=128,
             vocab_size=21136,
             hidden_size=1024,
@@ -79,12 +75,10 @@ def get_config(version='base', batch_size=1):
             type_vocab_size=2,
             initializer_range=0.02,
             use_relative_positions=False,
-            input_mask_from_dataset=True,
-            token_type_ids_from_dataset=True,
             dtype=mstype.float32,
             compute_type=mstype.float16)
     else:
-        bert_config = BertConfig(batch_size=batch_size)
+        bert_config = BertConfig()
     return bert_config
 
 
@@ -186,8 +180,7 @@ def test_bert_performance():
     context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", reserve_class_name_in_scope=False)
     ds, new_repeat_count, sink_size = me_de_train_dataset(sink_mode=True)
     version = os.getenv('VERSION', 'large')
-    batch_size = 16
-    config = get_config(version=version, batch_size=batch_size)
+    config = get_config(version=version)
     netwithloss = BertNetworkWithLoss(config, True)
 
     lr = BertLearningRate(decay_steps=sink_size * new_repeat_count,

tests/st/networks/models/bert/bert_precision/test_bert_tdt_lossscale.py

@@ -41,11 +41,10 @@ DATA_DIR = ["/home/workspace/mindspore_dataset/bert/example/examples.tfrecord"]
 SCHEMA_DIR = "/home/workspace/mindspore_dataset/bert/example/datasetSchema.json"
 
 
-def get_config(version='base', batch_size=1):
+def get_config(version='base'):
     """get config"""
     if version == 'base':
         bert_config = BertConfig(
-            batch_size=batch_size,
             seq_length=128,
             vocab_size=21136,
             hidden_size=768,
@@ -59,13 +58,10 @@ def get_config(version='base', batch_size=1):
             type_vocab_size=2,
             initializer_range=0.02,
             use_relative_positions=True,
-            input_mask_from_dataset=True,
-            token_type_ids_from_dataset=True,
             dtype=mstype.float32,
             compute_type=mstype.float32)
     elif version == 'large':
         bert_config = BertConfig(
-            batch_size=batch_size,
             seq_length=128,
             vocab_size=21136,
             hidden_size=1024,
@@ -79,12 +75,10 @@ def get_config(version='base', batch_size=1):
             type_vocab_size=2,
             initializer_range=0.02,
             use_relative_positions=False,
-            input_mask_from_dataset=True,
-            token_type_ids_from_dataset=True,
             dtype=mstype.float32,
             compute_type=mstype.float16)
     else:
-        bert_config = BertConfig(batch_size=batch_size)
+        bert_config = BertConfig()
     return bert_config
 
 
@@ -185,8 +179,7 @@ def test_bert_percision():
     context.set_context(mode=context.GRAPH_MODE, device_target="Ascend", reserve_class_name_in_scope=False)
     ds, new_repeat_count, _ = me_de_train_dataset()
     version = os.getenv('VERSION', 'large')
-    batch_size = 16
-    config = get_config(version=version, batch_size=batch_size)
+    config = get_config(version=version)
     netwithloss = BertNetworkWithLoss(config, True)
     lr = BertLearningRate(decay_steps=ds.get_dataset_size()*new_repeat_count,
                           learning_rate=5e-5, end_learning_rate=1e-9,