def main(config_path):
logger.info('------------Match-LSTM Train--------------')
logger.info('loading config file...')
global_config = read_config(config_path)
# set random seed
seed = global_config['model']['global']['random_seed']
torch.manual_seed(seed)
enable_cuda = global_config['train']['enable_cuda']
device = torch.device("cuda" if enable_cuda else "cpu")
if torch.cuda.is_available() and not enable_cuda:
logger.warning("CUDA is avaliable, you can enable CUDA in config file")
elif not torch.cuda.is_available() and enable_cuda:
raise ValueError("CUDA is not abaliable, please unable CUDA in config file")
logger.info('reading squad dataset...')
dataset = SquadDataset(global_config)
logger.info('constructing model...')
model = MatchLSTMModel(global_config).to(device)
criterion = MyNLLLoss()
# optimizer
optimizer_choose = global_config['train']['optimizer']
optimizer_lr = global_config['train']['learning_rate']
optimizer_param = filter(lambda p: p.requires_grad, model.parameters())
if optimizer_choose == 'adamax':
optimizer = optim.Adamax(optimizer_param)
elif optimizer_choose == 'adadelta':
optimizer = optim.Adadelta(optimizer_param)
elif optimizer_choose == 'adam':
optimizer = optim.Adam(optimizer_param)
elif optimizer_choose == 'sgd':
optimizer = optim.SGD(optimizer_param,
lr=optimizer_lr)
else:
raise ValueError('optimizer "%s" in config file not recoginized' % optimizer_choose)
# check if exist model weight
weight_path = global_config['data']['model_path']
if os.path.exists(weight_path):
logger.info('loading existing weight...')
weight = torch.load(weight_path, map_location=lambda storage, loc: storage)
if enable_cuda:
weight = torch.load(weight_path, map_location=lambda storage, loc: storage.cuda())
# weight = pop_dict_keys(weight, ['pointer', 'init_ptr_hidden']) # partial initial weight
model.load_state_dict(weight, strict=False)
# training arguments
logger.info('start training...')
train_batch_size = global_config['train']['batch_size']
valid_batch_size = global_config['train']['valid_batch_size']
# batch_train_data = dataset.get_dataloader_train(train_batch_size)
# batch_dev_data = dataset.get_dataloader_dev(valid_batch_size)
batch_train_data = list(dataset.get_batch_train(train_batch_size))
batch_dev_data = list(dataset.get_batch_dev(valid_batch_size))
clip_grad_max = global_config['train']['clip_grad_norm']
enable_char = global_config['model']['encoder']['enable_char']
best_valid_f1 = None
# every epoch
for epoch in range(global_config['train']['epoch']):
# train
model.train() # set training = True, make sure right dropout
sum_loss = train_on_model(model=model,
criterion=criterion,
optimizer=optimizer,
batch_data=batch_train_data,
epoch=epoch,
clip_grad_max=clip_grad_max,
device=device,
enable_char=enable_char,
batch_char_func=dataset.gen_batch_with_char)
logger.info('epoch=%d, sum_loss=%.5f' % (epoch, sum_loss))
# evaluate
with torch.no_grad():
model.eval() # let training = False, make sure right dropout
valid_score_em, valid_score_f1, valid_loss = eval_on_model(model=model,
criterion=criterion,
batch_data=batch_dev_data,
epoch=epoch,
device=device,
enable_char=enable_char,
batch_char_func=dataset.gen_batch_with_char)
logger.info("epoch=%d, ave_score_em=%.2f, ave_score_f1=%.2f, sum_loss=%.5f" %
(epoch, valid_score_em, valid_score_f1, valid_loss))
# save model when best f1 score
if best_valid_f1 is None or valid_score_f1 > best_valid_f1:
save_model(model,
epoch=epoch,
model_weight_path=global_config['data']['model_path'],
checkpoint_path=global_config['data']['checkpoint_path'])
logger.info("saving model weight on epoch=%d" % epoch)
best_valid_f1 = valid_score_f1
logger.info('finished.')
def train(config_path):
logger.info('------------MODEL TRAIN--------------')
logger.info('loading config file...')
global_config = read_config(config_path)
# set random seed
seed = global_config['global']['random_seed']
torch.manual_seed(seed)
enable_cuda = global_config['train']['enable_cuda']
device = torch.device("cuda" if enable_cuda else "cpu")
if torch.cuda.is_available() and not enable_cuda:
logger.warning("CUDA is avaliable, you can enable CUDA in config file")
elif not torch.cuda.is_available() and enable_cuda:
raise ValueError("CUDA is not abaliable, please unable CUDA in config file")
logger.info('reading dataset...')
dataset = Dataset(global_config)
logger.info('constructing model...')
dataset_h5_path = global_config['data']['dataset_h5']
model = MatchLSTMPlus(dataset_h5_path)
model = model.to(device)
criterion = MyNLLLoss()
optimizer_param = filter(lambda p: p.requires_grad, model.parameters())
model_rerank = None
rank_k = global_config['global']['rank_k']
if global_config['global']['enable_rerank']:
model_rerank = ReRanker(dataset_h5_path)
model_rerank = model_rerank.to(device)
criterion = torch.nn.NLLLoss()
optimizer_param = filter(lambda p: p.requires_grad, model_rerank.parameters())
# optimizer
optimizer_choose = global_config['train']['optimizer']
optimizer_lr = global_config['train']['learning_rate']
if optimizer_choose == 'adamax':
optimizer = optim.Adamax(optimizer_param)
elif optimizer_choose == 'adadelta':
optimizer = optim.Adadelta(optimizer_param)
elif optimizer_choose == 'adam':
optimizer = optim.Adam(optimizer_param)
elif optimizer_choose == 'sgd':
optimizer = optim.SGD(optimizer_param,
lr=optimizer_lr)
else:
raise ValueError('optimizer "%s" in config file not recoginized' % optimizer_choose)
# check if exist model weight
weight_path = global_config['data']['model_path']
if os.path.exists(weight_path):
logger.info('loading existing weight...')
weight = torch.load(weight_path, map_location=lambda storage, loc: storage)
if enable_cuda:
weight = torch.load(weight_path, map_location=lambda storage, loc: storage.cuda())
# weight = pop_dict_keys(weight, ['pointer', 'init_ptr_hidden']) # partial initial weight
model.load_state_dict(weight, strict=False)
rerank_weight_path = global_config['data']['rerank_model_path']
if global_config['global']['enable_rerank'] and os.path.exists(rerank_weight_path):
logger.info('loading existing rerank weight...')
weight = torch.load(rerank_weight_path, map_location=lambda storage, loc: storage)
if enable_cuda:
weight = torch.load(rerank_weight_path, map_location=lambda storage, loc: storage.cuda())
model_rerank.load_state_dict(weight, strict=False)
# training arguments
logger.info('start training...')
train_batch_size = global_config['train']['batch_size']
valid_batch_size = global_config['train']['valid_batch_size']
num_workers = global_config['global']['num_data_workers']
batch_train_data = dataset.get_dataloader_train(train_batch_size, num_workers)
batch_dev_data = dataset.get_dataloader_dev(valid_batch_size, num_workers)
clip_grad_max = global_config['train']['clip_grad_norm']
best_avg = 0.
# every epoch
for epoch in range(global_config['train']['epoch']):
# train
model.train() # set training = True, make sure right dropout
if global_config['global']['enable_rerank']:
model_rerank.train()
sum_loss = train_on_model(model=model,
criterion=criterion,
optimizer=optimizer,
batch_data=batch_train_data,
epoch=epoch,
clip_grad_max=clip_grad_max,
device=device,
model_rerank=model_rerank,
rank_k=rank_k)
logger.info('epoch=%d, sum_loss=%.5f' % (epoch, sum_loss))
# evaluate
with torch.no_grad():
model.eval() # let training = False, make sure right dropout
if global_config['global']['enable_rerank']:
model_rerank.eval()
valid_score_em, valid_score_f1, valid_loss = eval_on_model(model=model,
criterion=criterion,
batch_data=batch_dev_data,
epoch=epoch,
device=device,
model_rerank=model_rerank,
rank_k=rank_k)
valid_avg = (valid_score_em + valid_score_f1) / 2
logger.info("epoch=%d, ave_score_em=%.2f, ave_score_f1=%.2f, sum_loss=%.5f" %
(epoch, valid_score_em, valid_score_f1, valid_loss))
# save model when best avg score
if valid_avg > best_avg:
if model_rerank is not None:
save_model(model_rerank,
epoch=epoch,
model_weight_path=global_config['data']['rerank_model_path'],
checkpoint_path=global_config['data']['checkpoint_path'])
logging.info("saving rerank model weight on epoch=%d" % epoch)
else:
save_model(model,
epoch=epoch,
model_weight_path=global_config['data']['model_path'],
checkpoint_path=global_config['data']['checkpoint_path'])
logger.info("saving model weight on epoch=%d" % epoch)
best_avg = valid_avg
logger.info('finished.')
def test(config_path, out_path):
logger.info('------------MODEL PREDICT--------------')
logger.info('loading config file...')
global_config = read_config(config_path)
# set random seed
seed = global_config['global']['random_seed']
torch.manual_seed(seed)
#set default gpu
os.environ["CUDA_VISIBLE_DEVICES"] = str(global_config['train']["gpu_id"])
enable_cuda = global_config['test']['enable_cuda']
device = torch.device("cuda" if enable_cuda else "cpu")
if torch.cuda.is_available() and not enable_cuda:
logger.warning("CUDA is avaliable, you can enable CUDA in config file")
elif not torch.cuda.is_available() and enable_cuda:
raise ValueError(
"CUDA is not abaliable, please unable CUDA in config file")
torch.set_grad_enabled(
False) # make sure all tensors below have require_grad=False,
logger.info('reading squad dataset...')
dataset = SquadDataset(global_config)
logger.info('constructing model...')
model_choose = global_config['global']['model']
dataset_h5_path = global_config['data']['dataset_h5']
if model_choose == 'base':
model_config = read_config('config/base_model.yaml')
model = BaseModel(dataset_h5_path, model_config)
elif model_choose == 'match-lstm':
model = MatchLSTM(dataset_h5_path)
elif model_choose == 'match-lstm+':
model = MatchLSTMPlus(dataset_h5_path,
global_config['preprocess']['use_domain_tag'])
elif model_choose == 'r-net':
model = RNet(dataset_h5_path)
elif model_choose == 'm-reader':
model = MReader(dataset_h5_path)
else:
raise ValueError('model "%s" in config file not recoginized' %
model_choose)
model = model.to(device)
model.eval() # let training = False, make sure right dropout
# load model weight
logger.info('loading model weight...')
model_weight_path = global_config['data']['model_path']
assert os.path.exists(
model_weight_path
), "not found model weight file on '%s'" % model_weight_path
weight = torch.load(model_weight_path,
map_location=lambda storage, loc: storage)
if enable_cuda:
weight = torch.load(model_weight_path,
map_location=lambda storage, loc: storage.cuda())
model.load_state_dict(weight, strict=False)
# forward
logger.info('forwarding...')
batch_size = global_config['test']['batch_size']
num_workers = global_config['global']['num_data_workers']
if 'test_path' not in global_config['data']['dataset']:
batch_test_data = dataset.get_dataloader_dev(batch_size, num_workers)
else:
batch_test_data = dataset.get_dataloader_test(batch_size, num_workers)
# to just evaluate score or write answer to file
if out_path is None:
criterion = MyNLLLoss()
score_em, score_f1, sum_loss = eval_on_model(
model=model,
criterion=criterion,
batch_data=batch_test_data,
epoch=None,
device=device)
logger.info(
"test: ave_score_em=%.2f, ave_score_f1=%.2f, sum_loss=%.5f" %
(score_em, score_f1, sum_loss))
else:
#context_right_space = dataset.get_all_ct_right_space_dev()
context_right_space = dataset.get_all_ct_right_space_test()
predict_ans = predict_on_model(
model=model,
batch_data=batch_test_data,
device=device,
id_to_word_func=dataset.sentence_id2word,
right_space=context_right_space)
#samples_id = dataset.get_all_samples_id_dev()
samples_id = dataset.get_all_samples_id_test()
ans_with_id = dict(zip(samples_id, predict_ans))
logging.info('writing predict answer to file %s' % out_path)
with open(out_path, 'w') as f:
json.dump(ans_with_id, f)
logging.info('finished.')
def test(config_path, out_path):
logger.info('------------MODEL PREDICT--------------')
logger.info('loading config file...')
global_config = read_config(config_path)
# set random seed
seed = global_config['global']['random_seed']
torch.manual_seed(seed)
enable_cuda = global_config['test']['enable_cuda']
device = torch.device("cuda" if enable_cuda else "cpu")
if torch.cuda.is_available() and not enable_cuda:
logger.warning("CUDA is avaliable, you can enable CUDA in config file")
elif not torch.cuda.is_available() and enable_cuda:
raise ValueError(
"CUDA is not abaliable, please unable CUDA in config file")
torch.set_grad_enabled(
False) # make sure all tensors below have require_grad=False,
logger.info('reading dataset...')
dataset = Dataset(global_config)
logger.info('constructing model...')
dataset_h5_path = global_config['data']['dataset_h5']
model = MatchLSTMPlus(dataset_h5_path)
model = model.to(device)
model.eval() # let training = False, make sure right dropout
criterion = MyNLLLoss()
model_rerank = None
rank_k = global_config['global']['rank_k']
if global_config['global']['enable_rerank']:
model_rerank = ReRanker(dataset_h5_path)
model_rerank = model_rerank.to(device)
model_rerank.eval()
criterion = torch.nn.NLLLoss()
# load model weight
logger.info('loading model weight...')
model_weight_path = global_config['data']['model_path']
assert os.path.exists(
model_weight_path
), "not found model weight file on '%s'" % model_weight_path
weight = torch.load(model_weight_path,
map_location=lambda storage, loc: storage)
if enable_cuda:
weight = torch.load(model_weight_path,
map_location=lambda storage, loc: storage.cuda())
model.load_state_dict(weight, strict=False)
if global_config['global']['enable_rerank']:
rerank_weight_path = global_config['data']['rerank_model_path']
assert os.path.exists(
rerank_weight_path
), "not found rerank model weight file on '%s'" % rerank_weight_path
logger.info('loading rerank model weight...')
weight = torch.load(rerank_weight_path,
map_location=lambda storage, loc: storage)
if enable_cuda:
weight = torch.load(
rerank_weight_path,
map_location=lambda storage, loc: storage.cuda())
model_rerank.load_state_dict(weight, strict=False)
# forward
logger.info('forwarding...')
batch_size = global_config['test']['batch_size']
num_workers = global_config['global']['num_data_workers']
batch_dev_data = dataset.get_dataloader_dev(batch_size, num_workers)
# to just evaluate score or write answer to file
if out_path is None:
score_em, score_f1, sum_loss = eval_on_model(model=model,
criterion=criterion,
batch_data=batch_dev_data,
epoch=None,
device=device,
model_rerank=model_rerank,
rank_k=rank_k)
logger.info(
"test: ave_score_em=%.2f, ave_score_f1=%.2f, sum_loss=%.5f" %
(score_em, score_f1, sum_loss))
else:
context_right_space = dataset.get_all_ct_right_space_dev()
predict_ans = predict_on_model(
model=model,
batch_data=batch_dev_data,
device=device,
id_to_word_func=dataset.sentence_id2word,
right_space=context_right_space,
model_rerank=model_rerank,
rank_k=rank_k)
samples_id = dataset.get_all_samples_id_dev()
ans_with_id = dict(zip(samples_id, predict_ans))
logging.info('writing predict answer to file %s' % out_path)
with open(out_path, 'w') as f:
json.dump(ans_with_id, f)
logging.info('finished.')
def train(config_path):
logger.info('------------MODEL TRAIN--------------')
logger.info('loading config file...')
global_config = read_config(config_path)
# set random seed
seed = global_config['global']['random_seed']
torch.manual_seed(seed)
#set default gpu
os.environ["CUDA_VISIBLE_DEVICES"] = str(global_config['train']["gpu_id"])
enable_cuda = global_config['train']['enable_cuda']
device = torch.device("cuda" if enable_cuda else "cpu")
if torch.cuda.is_available() and not enable_cuda:
logger.warning("CUDA is avaliable, you can enable CUDA in config file")
elif not torch.cuda.is_available() and enable_cuda:
raise ValueError(
"CUDA is not abaliable, please unable CUDA in config file")
logger.info('reading squad dataset...')
dataset = SquadDataset(global_config)
logger.info('constructing model...')
model_choose = global_config['global']['model']
dataset_h5_path = global_config['data']['dataset_h5']
if model_choose == 'base':
model_config = read_config('config/base_model.yaml')
model = BaseModel(dataset_h5_path, model_config)
elif model_choose == 'match-lstm':
model = MatchLSTM(dataset_h5_path)
elif model_choose == 'match-lstm+':
model = MatchLSTMPlus(dataset_h5_path,
global_config['preprocess']['use_domain_tag'])
elif model_choose == 'r-net':
model = RNet(dataset_h5_path)
elif model_choose == 'm-reader':
model = MReader(dataset_h5_path)
else:
raise ValueError('model "%s" in config file not recoginized' %
model_choose)
model = model.to(device)
criterion = MyNLLLoss()
# optimizer
optimizer_choose = global_config['train']['optimizer']
optimizer_lr = global_config['train']['learning_rate']
optimizer_param = filter(lambda p: p.requires_grad, model.parameters())
if optimizer_choose == 'adamax':
optimizer = optim.Adamax(optimizer_param)
elif optimizer_choose == 'adadelta':
optimizer = optim.Adadelta(optimizer_param)
elif optimizer_choose == 'adam':
optimizer = optim.Adam(optimizer_param)
elif optimizer_choose == 'sgd':
optimizer = optim.SGD(optimizer_param, lr=optimizer_lr)
else:
raise ValueError('optimizer "%s" in config file not recoginized' %
optimizer_choose)
# check if exist model weight
weight_path = global_config['data']['model_path']
if os.path.exists(weight_path):
logger.info('loading existing weight...')
weight = torch.load(weight_path,
map_location=lambda storage, loc: storage)
if enable_cuda:
weight = torch.load(
weight_path, map_location=lambda storage, loc: storage.cuda())
# weight = pop_dict_keys(weight, ['pointer', 'init_ptr_hidden']) # partial initial weight
model.load_state_dict(weight, strict=False)
# training arguments
logger.info('start training...')
train_batch_size = global_config['train']['batch_size']
valid_batch_size = global_config['train']['valid_batch_size']
num_workers = global_config['global']['num_data_workers']
batch_train_data = dataset.get_dataloader_train(train_batch_size,
num_workers)
batch_dev_data = dataset.get_dataloader_dev(valid_batch_size, num_workers)
clip_grad_max = global_config['train']['clip_grad_norm']
best_avg = 0.
# every epoch
for epoch in range(global_config['train']['epoch']):
# train
model.train() # set training = True, make sure right dropout
sum_loss = train_on_model(model=model,
criterion=criterion,
optimizer=optimizer,
batch_data=batch_train_data,
epoch=epoch,
clip_grad_max=clip_grad_max,
device=device)
logger.info('epoch=%d, sum_loss=%.5f' % (epoch, sum_loss))
# evaluate
with torch.no_grad():
model.eval() # let training = False, make sure right dropout
valid_score_em, valid_score_f1, valid_loss = eval_on_model(
model=model,
criterion=criterion,
batch_data=batch_dev_data,
epoch=epoch,
device=device)
valid_avg = (valid_score_em + valid_score_f1) / 2
logger.info(
"epoch=%d, ave_score_em=%.2f, ave_score_f1=%.2f, sum_loss=%.5f" %
(epoch, valid_score_em, valid_score_f1, valid_loss))
# save model when best avg score
if valid_avg > best_avg:
save_model(
model,
epoch=epoch,
model_weight_path=global_config['data']['model_path'],
checkpoint_path=global_config['data']['checkpoint_path'])
logger.info("saving model weight on epoch=%d" % epoch)
best_avg = valid_avg
logger.info('pretraining finished.')
if global_config['global']['finetune']:
batch_train_data = dataset.get_dataloader_train2(
train_batch_size, num_workers)
batch_dev_data = dataset.get_dataloader_dev2(valid_batch_size,
num_workers)
for epoch in range(global_config['train']['finetune_epoch']):
# train
model.train() # set training = True, make sure right dropout
sum_loss = train_on_model(model=model,
criterion=criterion,
optimizer=optimizer,
batch_data=batch_train_data,
epoch=epoch,
clip_grad_max=clip_grad_max,
device=device)
logger.info('finetune epoch=%d, sum_loss=%.5f' % (epoch, sum_loss))
# evaluate
with torch.no_grad():
model.eval() # let training = False, make sure right dropout
valid_score_em, valid_score_f1, valid_loss = eval_on_model(
model=model,
criterion=criterion,
batch_data=batch_dev_data,
epoch=epoch,
device=device)
valid_avg = (valid_score_em + valid_score_f1) / 2
logger.info(
"finetune epoch=%d, ave_score_em=%.2f, ave_score_f1=%.2f, sum_loss=%.5f"
% (epoch, valid_score_em, valid_score_f1, valid_loss))
# save model when best avg score
if valid_avg > best_avg:
save_model(
model,
epoch=epoch,
model_weight_path=global_config['data']['model_path'],
checkpoint_path=global_config['data']['checkpoint_path'])
logger.info("saving model weight on epoch=%d" % epoch)
best_avg = valid_avg
if global_config['global']['finetune2']:
batch_train_data = dataset.get_dataloader_train3(
train_batch_size, num_workers)
batch_dev_data = dataset.get_dataloader_dev3(valid_batch_size,
num_workers)
for epoch in range(global_config['train']['finetune_epoch2']):
# train
model.train() # set training = True, make sure right dropout
sum_loss = train_on_model(model=model,
criterion=criterion,
optimizer=optimizer,
batch_data=batch_train_data,
epoch=epoch,
clip_grad_max=clip_grad_max,
device=device)
logger.info('finetune2 epoch=%d, sum_loss=%.5f' %
(epoch, sum_loss))
# evaluate
with torch.no_grad():
model.eval() # let training = False, make sure right dropout
valid_score_em, valid_score_f1, valid_loss = eval_on_model(
model=model,
criterion=criterion,
batch_data=batch_dev_data,
epoch=epoch,
device=device)
valid_avg = (valid_score_em + valid_score_f1) / 2
logger.info(
"finetune2 epoch=%d, ave_score_em=%.2f, ave_score_f1=%.2f, sum_loss=%.5f"
% (epoch, valid_score_em, valid_score_f1, valid_loss))
# save model when best avg score
if valid_avg > best_avg:
save_model(
model,
epoch=epoch,
model_weight_path=global_config['data']['model_path'],
checkpoint_path=global_config['data']['checkpoint_path'])
logger.info("saving model weight on epoch=%d" % epoch)
best_avg = valid_avg
logger.info('finished.')
def main(config_path, out_path):
logger.info('------------Match-LSTM Evaluate--------------')
logger.info('loading config file...')
global_config = read_config(config_path)
# set random seed
seed = global_config['model']['global']['random_seed']
torch.manual_seed(seed)
enable_cuda = global_config['test']['enable_cuda']
device = torch.device("cuda" if enable_cuda else "cpu")
if torch.cuda.is_available() and not enable_cuda:
logger.warning("CUDA is avaliable, you can enable CUDA in config file")
elif not torch.cuda.is_available() and enable_cuda:
raise ValueError("CUDA is not abaliable, please unable CUDA in config file")
torch.no_grad() # make sure all tensors below have require_grad=False
logger.info('reading squad dataset...')
dataset = SquadDataset(global_config)
logger.info('constructing model...')
model = MatchLSTMModel(global_config).to(device)
model.eval() # let training = False, make sure right dropout
# load model weight
logger.info('loading model weight...')
model_weight_path = global_config['data']['model_path']
assert os.path.exists(model_weight_path), "not found model weight file on '%s'" % model_weight_path
weight = torch.load(model_weight_path, map_location=lambda storage, loc: storage)
if enable_cuda:
weight = torch.load(model_weight_path, map_location=lambda storage, loc: storage.cuda())
model.load_state_dict(weight, strict=False)
# forward
logger.info('forwarding...')
enable_char = global_config['model']['encoder']['enable_char']
batch_size = global_config['test']['batch_size']
# batch_dev_data = dataset.get_dataloader_dev(batch_size)
batch_dev_data = list(dataset.get_batch_dev(batch_size))
# to just evaluate score or write answer to file
if out_path is None:
criterion = MyNLLLoss()
score_em, score_f1, sum_loss = eval_on_model(model=model,
criterion=criterion,
batch_data=batch_dev_data,
epoch=None,
device=device,
enable_char=enable_char,
batch_char_func=dataset.gen_batch_with_char)
logger.info("test: ave_score_em=%.2f, ave_score_f1=%.2f, sum_loss=%.5f" % (score_em, score_f1, sum_loss))
else:
predict_ans = predict_on_model(model=model,
batch_data=batch_dev_data,
device=device,
enable_char=enable_char,
batch_char_func=dataset.gen_batch_with_char,
id_to_word_func=dataset.sentence_id2word)
samples_id = dataset.get_all_samples_id_dev()
ans_with_id = dict(zip(samples_id, predict_ans))
logging.info('writing predict answer to file %s' % out_path)
with open(out_path, 'w') as f:
json.dump(ans_with_id, f)
logging.info('finished.')