def main():
# parser config
config_file = "./config.ini"
cp = ConfigParser()
cp.read(config_file)
# default config
output_fold = cp["TRAIN"].get("output_fold")
epochs = cp["TRAIN"].getint("epochs")
batch_size = cp["TRAIN"].getint("batch_size")
generator_workers = cp["TRAIN"].getint("generator_workers")
output_weights_name = cp["TRAIN"].get("output_weights_name")
sequence_length_max = cp["TRAIN"].getint("sequence_length_max")
output_model_name = cp["TRAIN"].get("output_model_name")
save_weights_only = cp["TRAIN"].getboolean("save_weights_only")
cyclicLR_mode = cp["TRAIN"].get("cyclicLR_mode")
base_lr = cp["TRAIN"].getfloat("base_lr")
max_lr = cp["TRAIN"].getfloat("max_lr")
file_train = cp["TRAIN"].get("file_train")
file_valid = cp["TRAIN"].get("file_valid")
file_test = cp["TRAIN"].get("file_test")
today = datetime.date.today()
formatted_today = today.strftime('%y%m%d')
output_dir = os.path.join('experiments', formatted_today, output_fold)
if not os.path.isdir(output_dir):
os.makedirs(output_dir)
output_dir_src = os.path.join(output_dir, 'src')
if not os.path.isdir(output_dir_src):
os.makedirs(output_dir_src)
print(f"backup config file to {output_dir_src}")
shutil.copy(config_file, os.path.join(output_dir_src, os.path.split(config_file)[1]))
train_file = os.path.basename(__file__)
shutil.copy(train_file, os.path.join(output_dir_src, train_file))
logging.basicConfig(level='DEBUG')
bert_path = get_file('bert_sample_model',
"http://s3.bmio.net/kashgari/bert_sample_model.tar.bz2",
cache_dir=DATA_PATH,
untar=True)
train_x, train_y = preprocess(file_train)
validate_x, validate_y = preprocess(file_valid)
test_x, test_y = preprocess(file_test)
#'bert-large-cased'
embedding = BERTEmbedding(bert_path,
sequence_length=sequence_length_max,
task=kashgari.CLASSIFICATION,
trainable=True,
layer_nums=4)
#embedding = BERTEmbedding('/home/ys1/pretrained_models/BERT/Japanese_L-12_H-768_A-12_E-30_BPE/', sequence_length=sequence_length_max, task=kashgari.CLASSIFICATION)
# 还可以选择 CNNModel CNNLSTMModel
# model = BiGRU_Model(embedding)
hyper = BiLSTM_Model.get_default_hyper_parameters()
print(f'hyper parameters is:{hyper}')
#hyper parameters is:{'layer_bi_lstm': {'units': 128, 'return_sequences': False}, 'layer_dense': {'activation': 'softmax'}}
# hyper['layer_bi_lstm']['units'] = 32
model = BiLSTM_Model(embedding, hyper_parameters=hyper)
# model.build_model(train_x, train_y)
# model.build_multi_gpu_model(gpus=2)
# print(model.summary())
if save_weights_only:
model_weights = os.path.join(output_dir, output_weights_name)
else:
model_weights = os.path.join(output_dir, output_model_name)
checkpoint = keras.callbacks.ModelCheckpoint(
model_weights,
save_weights_only=save_weights_only,
save_best_only=True,
verbose=1,
)
earlystop = keras.callbacks.EarlyStopping(monitor='val_loss', min_delta=0, patience=20, verbose=0, mode='min')
csv_logger = keras.callbacks.CSVLogger(os.path.join(output_dir, 'training.csv'))
batch_size_cycliclr = ceil(len(train_x)/batch_size)
if cyclicLR_mode == 'exp_range':
gamma = 0.99994
else:
gamma = 1.
clr = CyclicLR(mode=cyclicLR_mode, step_size=batch_size_cycliclr, base_lr=base_lr, max_lr=max_lr, gamma=gamma)
save_min_loss = SaveMinLoss(filepath=output_dir)
tb = keras.callbacks.TensorBoard(log_dir=os.path.join(output_dir, "logs"), batch_size=batch_size, update_freq=1000)
# 这是 Kashgari 内置回调函数，会在训练过程计算精确度，召回率和 F1
eval_callback = EvalCallBack(kash_model=model,
valid_x=validate_x,
valid_y=validate_y,
step=5)
callbacks = [
eval_callback,
checkpoint,
tb,
csv_logger,
# clr,
save_min_loss,
earlystop,
]
print("** start training **")
model.fit(train_x,
train_y,
x_validate=validate_x,
y_validate=validate_y,
epochs=epochs,
batch_size=batch_size,
callbacks=callbacks,
fit_kwargs={
'workers': generator_workers,
'use_multiprocessing': True,
'class_weight': 'auto',
}
)
model_path = os.path.join(output_dir, 'model')
model.save(model_path)
report_evaluate = model.evaluate(test_x, test_y, debug_info=True)
with open(os.path.join(output_dir, 'report_evaluate.log'), 'w') as f:
f.write(f"The evaluate report is : \n{str(report_evaluate)}")
# for i in range(0, len(test_y)):
# test_y[i] = str(test_y[i])
if os.path.exists(args.bertpath):
print('Using BERT: ' + args.bertpath)
bert_embed = BERTEmbedding(
args.bertpath,
task=kashgari.CLASSIFICATION,
# trainable=True,
sequence_length=30)
else:
raise ValueError("Invalid BertPath! " + args.bertpath)
if str(args.model).lower() == 'bilstm':
print('Use BiLSTM Architecture')
MODELFILENAME = 'BiLSTM'
hyper = BiLSTM_Model.get_default_hyper_parameters()
hyper['layer_bi_lstm']['units'] = 64
model = BiLSTM_Model(bert_embed, hyper_parameters=hyper)
elif str(args.model).lower() == 'custom':
print('Use BiLSTM Architecture')
MODELFILENAME = 'CUSTOM'
hyper = DoubleBLSTMModel.get_default_hyper_parameters()
#started working directly in model definition
model = DoubleBLSTMModel(bert_embed, hyper_parameters=hyper)
else:
raise ValueError(
"So far only bilstm or custom model available for BERT embeddings!"
)