tinybert hub

model_zoo/official/nlp/bert/mindspore_hub_conf.py

@@ -59,11 +59,11 @@ def create_network(name, *args, **kwargs):
     if name == 'bert_base':
         if "seq_length" in kwargs:
             bert_net_cfg_base.seq_length = kwargs["seq_length"]
-        is_training = kwargs.get("is_training", default=False)
+        is_training = kwargs.get("is_training", False)
         return BertModel(bert_net_cfg_base, is_training, *args)
     if name == 'bert_nezha':
         if "seq_length" in kwargs:
             bert_net_cfg_nezha.seq_length = kwargs["seq_length"]
-        is_training = kwargs.get("is_training", default=False)
+        is_training = kwargs.get("is_training", False)
         return BertModel(bert_net_cfg_nezha, is_training, *args)
     raise NotImplementedError(f"{name} is not implemented in the repo")

model_zoo/official/nlp/tinybert/README.md

@@ -207,6 +207,7 @@ options:
 `gd_config.py` and `td_config.py` contain parameters of BERT model and options for optimizer and lossscale.
 ### Options:
 ```
+batch_size                          batch size of input dataset: N, default is 16
 Parameters for lossscale:
     loss_scale_value                initial value of loss scale: N, default is 2^8
     scale_factor                    factor used to update loss scale: N, default is 2
@@ -223,7 +224,6 @@ Parameters for optimizer:
 ### Parameters:
 ```
 Parameters for bert network:
-    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 30522
     hidden_size                     size of bert encoder layers: N
@@ -239,8 +239,6 @@ Parameters for bert network:
     type_vocab_size                 size of token type vocab: N, default is 2
     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/tinybert/mindspore_hub_conf.py

@@ -0,0 +1,49 @@
+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# 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.
+# ============================================================================
+'''
+Bert hub interface for bert base and bert nezha
+'''
+from src.tinybert_model import TinyBertModel
+from src.tinybert_model import BertConfig
+import mindspore.common.dtype as mstype
+
+tinybert_student_net_cfg = BertConfig(
+    seq_length=128,
+    vocab_size=30522,
+    hidden_size=384,
+    num_hidden_layers=4,
+    num_attention_heads=12,
+    intermediate_size=1536,
+    hidden_act="gelu",
+    hidden_dropout_prob=0.1,
+    attention_probs_dropout_prob=0.1,
+    max_position_embeddings=512,
+    type_vocab_size=2,
+    initializer_range=0.02,
+    use_relative_positions=False,
+    dtype=mstype.float32,
+    compute_type=mstype.float16
+)
+
+def create_network(name, *args, **kwargs):
+    '''
+    Create tinybert network.
+    '''
+    if name == "tinybert":
+        if "seq_length" in kwargs:
+            tinybert_student_net_cfg.seq_length = kwargs["seq_length"]
+        is_training = kwargs.get("is_training", False)
+        return TinyBertModel(tinybert_student_net_cfg, is_training, *args)
+    raise NotImplementedError(f"{name} is not implemented in the repo")

model_zoo/official/nlp/tinybert/run_general_distill.py

@@ -110,7 +110,7 @@ def run_general_distill():
         dataset_type = DataType.MINDRECORD
     else:
         raise Exception("dataset format is not supported yet")
-    dataset = create_tinybert_dataset('gd', bert_teacher_net_cfg.batch_size, device_num, rank,
+    dataset = create_tinybert_dataset('gd', common_cfg.batch_size, device_num, rank,
                                       args_opt.do_shuffle, args_opt.data_dir, args_opt.schema_dir,
                                       data_type=dataset_type)
     dataset_size = dataset.get_dataset_size()

model_zoo/official/nlp/tinybert/run_task_distill.py

@@ -29,7 +29,7 @@ from mindspore import log as logger
 from src.dataset import create_tinybert_dataset, DataType
 from src.utils import LossCallBack, ModelSaveCkpt, EvalCallBack, BertLearningRate
 from src.assessment_method import Accuracy
-from src.td_config import phase1_cfg, phase2_cfg, td_teacher_net_cfg, td_student_net_cfg
+from src.td_config import phase1_cfg, phase2_cfg, eval_cfg, td_teacher_net_cfg, td_student_net_cfg
 from src.tinybert_for_gd_td import BertEvaluationWithLossScaleCell, BertNetworkWithLoss_td, BertEvaluationCell
 from src.tinybert_model import BertModelCLS
 
@@ -130,7 +130,7 @@ def run_predistill():
         dataset_type = DataType.MINDRECORD
     else:
         raise Exception("dataset format is not supported yet")
-    dataset = create_tinybert_dataset('td', td_teacher_net_cfg.batch_size,
+    dataset = create_tinybert_dataset('td', cfg.batch_size,
                                       device_num, rank, args_opt.do_shuffle,
                                       args_opt.train_data_dir, args_opt.schema_dir,
                                       data_type=dataset_type)
@@ -194,7 +194,7 @@ def run_task_distill(ckpt_file):
 
     rank = 0
     device_num = 1
-    train_dataset = create_tinybert_dataset('td', td_teacher_net_cfg.batch_size,
+    train_dataset = create_tinybert_dataset('td', cfg.batch_size,
                                             device_num, rank, args_opt.do_shuffle,
                                             args_opt.train_data_dir, args_opt.schema_dir)
 
@@ -224,7 +224,7 @@ def run_task_distill(ckpt_file):
 
     optimizer = AdamWeightDecay(group_params, learning_rate=lr_schedule, eps=optimizer_cfg.AdamWeightDecay.eps)
 
-    eval_dataset = create_tinybert_dataset('td', td_teacher_net_cfg.batch_size,
+    eval_dataset = create_tinybert_dataset('td', eval_cfg.batch_size,
                                            device_num, rank, args_opt.do_shuffle,
                                            args_opt.eval_data_dir, args_opt.schema_dir)
     print('td2 eval dataset size: ', eval_dataset.get_dataset_size())
@@ -269,7 +269,7 @@ def do_eval_standalone():
     load_param_into_net(eval_model, new_param_dict)
     eval_model.set_train(False)
 
-    eval_dataset = create_tinybert_dataset('td', batch_size=td_student_net_cfg.batch_size,
+    eval_dataset = create_tinybert_dataset('td', batch_size=eval_cfg.batch_size,
                                            device_num=1, rank=0, do_shuffle="false",
                                            data_dir=args_opt.eval_data_dir,
                                            schema_dir=args_opt.schema_dir)

model_zoo/official/nlp/tinybert/src/gd_config.py

@@ -20,6 +20,7 @@ from easydict import EasyDict as edict
 from .tinybert_model import BertConfig
 
 common_cfg = edict({
+    'batch_size': 32,
     'loss_scale_value': 2 ** 16,
     'scale_factor': 2,
     'scale_window': 1000,
@@ -38,7 +39,6 @@ teacher network: The BERT-base network.
 student network: The network which is inherited from teacher network.
 '''
 bert_teacher_net_cfg = BertConfig(
-    batch_size=32,
     seq_length=128,
     vocab_size=30522,
     hidden_size=768,
@@ -52,13 +52,10 @@ bert_teacher_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
 )
 bert_student_net_cfg = BertConfig(
-    batch_size=32,
     seq_length=128,
     vocab_size=30522,
     hidden_size=384,
@@ -72,8 +69,6 @@ bert_student_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/tinybert/src/td_config.py

@@ -20,6 +20,7 @@ from easydict import EasyDict as edict
 from .tinybert_model import BertConfig
 
 phase1_cfg = edict({
+    'batch_size': 32,
     'loss_scale_value': 2 ** 8,
     'scale_factor': 2,
     'scale_window': 50,
@@ -36,6 +37,7 @@ phase1_cfg = edict({
 })
 
 phase2_cfg = edict({
+    'batch_size': 32,
     'loss_scale_value': 2 ** 16,
     'scale_factor': 2,
     'scale_window': 50,
@@ -51,13 +53,16 @@ phase2_cfg = edict({
     }),
 })
 
+eval_cfg = edict({
+    'batch_size': 32,
+})
+
 '''
 Including two kinds of network: \
 teacher network: The BERT-base network with finetune.
 student network: The model which is producted by GD phase.
 '''
 td_teacher_net_cfg = BertConfig(
-    batch_size=32,
     seq_length=128,
     vocab_size=30522,
     hidden_size=768,
@@ -71,13 +76,10 @@ td_teacher_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
 )
 td_student_net_cfg = BertConfig(
-    batch_size=32,
     seq_length=128,
     vocab_size=30522,
     hidden_size=384,
@@ -91,8 +93,6 @@ td_student_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
 )