def main():
logger.info('------------Analysis SQuAD dataset--------------')
logger.info('loading config file...')
global_config = read_config()
logger.info('reading squad dataset...')
dataset = SquadDataset(global_config)
train_context_len_cnt, train_context_len = dataset.gather_context_seq_len('train')
dev_context_len_cnt, dev_context_len = dataset.gather_context_seq_len('dev')
logging.info('train context length cnt: ' + str(train_context_len_cnt))
logging.info('dev context length cnt: ' + str(dev_context_len_cnt))
train_context_len.plot.scatter('length', 'cnt', title='train')
dev_context_len.plot.scatter('length', 'cnt', title='dev')
plt.show()
def main():
logger.info('------------Analysis SQuAD dataset--------------')
logger.info('loading config file...')
global_config = read_config()
logger.info('reading squad dataset...')
dataset = SquadDataset(global_config)
train_context_len_cnt, train_context_len = dataset.gather_context_seq_len('train')
dev_context_len_cnt, dev_context_len = dataset.gather_context_seq_len('dev')
train_answer_len = dataset.gather_answer_seq_len('train', max_len=9)
dev_answer_len = dataset.gather_answer_seq_len('dev', max_len=9)
logging.info('train context length cnt: ' + str(train_context_len_cnt))
logging.info('dev context length cnt: ' + str(dev_context_len_cnt))
sns.set()
train_context_len.plot.scatter('length', 'cnt', title='train')
plt.xlabel('passage length')
dev_context_len.plot.scatter('length', 'cnt', title='dev')
plt.xlabel('passage length')
train_answer_len.plot.line('length', 'cnt', marker='o', title='train')
plt.xticks(range(len(train_answer_len['length'])), train_answer_len['length'])
plt.xlabel('answer length')
dev_answer_len.plot.line('length', 'cnt', marker='o', title='dev')
plt.xticks(range(len(dev_answer_len['length'])), dev_answer_len['length'])
plt.xlabel('answer length')
plt.show()
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 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)
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('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 __init__(self, config_file="/mnt/sdb/cjm/Match-LSTM/config/id19.yaml"):
logger.info('------------MODEL TEST INPUT--------------')
logger.info('loading config file...')
#self.global_config = read_config("config/id19.yaml")
self.global_config = read_config(config_file)
# set random seed
seed = self.global_config['global']['random_seed']
torch.manual_seed(seed)
torch.set_grad_enabled(
False) # make sure all tensors below have require_grad=False
# set default gpu
os.environ["CUDA_VISIBLE_DEVICES"] = str(
self.global_config['test']["gpu_id"])
enable_cuda = self.global_config['test']['enable_cuda']
self.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...')
self.dataset = SquadDataset(self.global_config)
logger.info('constructing model...')
model_choose = self.global_config['global']['model']
dataset_h5_path = self.global_config['data']['dataset_h5']
if model_choose == 'base':
model = BaseModel(
dataset_h5_path,
model_config=read_config('config/base_model.yaml'))
elif model_choose == 'match-lstm':
model = MatchLSTM(dataset_h5_path)
elif model_choose == 'match-lstm+':
model = MatchLSTMPlus(
dataset_h5_path,
self.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(self.device)
model.eval() # let training = False, make sure right dropout
logging.info('model parameters count: %d' % count_parameters(model))
# load model weight
logger.info('loading model weight...')
model_weight_path = self.global_config['data']['model_path']
is_exist_model_weight = os.path.exists(model_weight_path)
assert is_exist_model_weight, "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)
self.model = model
self.nlp = spacy.load('en')
self.metadata = {
k: v.tolist()
for k, v in self.dataset.meta_data.items()
}
class mrc:
def __init__(self, config_file="/mnt/sdb/cjm/Match-LSTM/config/id19.yaml"):
logger.info('------------MODEL TEST INPUT--------------')
logger.info('loading config file...')
#self.global_config = read_config("config/id19.yaml")
self.global_config = read_config(config_file)
# set random seed
seed = self.global_config['global']['random_seed']
torch.manual_seed(seed)
torch.set_grad_enabled(
False) # make sure all tensors below have require_grad=False
# set default gpu
os.environ["CUDA_VISIBLE_DEVICES"] = str(
self.global_config['test']["gpu_id"])
enable_cuda = self.global_config['test']['enable_cuda']
self.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...')
self.dataset = SquadDataset(self.global_config)
logger.info('constructing model...')
model_choose = self.global_config['global']['model']
dataset_h5_path = self.global_config['data']['dataset_h5']
if model_choose == 'base':
model = BaseModel(
dataset_h5_path,
model_config=read_config('config/base_model.yaml'))
elif model_choose == 'match-lstm':
model = MatchLSTM(dataset_h5_path)
elif model_choose == 'match-lstm+':
model = MatchLSTMPlus(
dataset_h5_path,
self.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(self.device)
model.eval() # let training = False, make sure right dropout
logging.info('model parameters count: %d' % count_parameters(model))
# load model weight
logger.info('loading model weight...')
model_weight_path = self.global_config['data']['model_path']
is_exist_model_weight = os.path.exists(model_weight_path)
assert is_exist_model_weight, "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)
self.model = model
self.nlp = spacy.load('en')
self.metadata = {
k: v.tolist()
for k, v in self.dataset.meta_data.items()
}
def mrcqa_batch(self, contexts, question, single_question=True):
if single_question:
questions = []
for i in range(len(contexts)):
questions.append(question)
data_nopad = self.build_data(contexts, questions)
else:
data_nopad = self.build_data(contexts, question)
data_pad = {
'context':
self.dict2array(data_nopad['context']),
'question':
self.dict2array(data_nopad['question']),
'answer_range':
pad_sequences(data_nopad['answer_range'], padding='post',
value=-1),
'samples_id':
np.array(data_nopad['samples_id'])
}
batch_data = self.dataset.get_input_dataloader(
self.global_config['test']['batch_size'],
self.global_config['global']['num_data_workers'],
shuffle=False,
input_data=data_pad)
# batch_data = dataset.get_dataloader_test(32, 5)
batch_cnt = len(batch_data)
answer = []
cdict = data_pad['context']
right_space = cdict['right_space']
cnt = 0
for bnum, batch in enumerate(batch_data):
batch = [x.to(self.device) if x is not None else x for x in batch]
bat_context = batch[0]
bat_answer_range = batch[-1]
# forward
batch_input = batch[:len(batch) - 1]
_, tmp_ans_range, _ = self.model.forward(*batch_input)
tmp_context_ans = zip(bat_context.cpu().data.numpy(),
tmp_ans_range.cpu().data.numpy())
# generate initial answer text
i = 0
for c, a in tmp_context_ans:
cur_no = cnt + i
tmp_ans = self.dataset.sentence_id2word(c[a[0]:(a[1] + 1)])
cur_space = right_space[cur_no][a[0]:(a[1] + 1)]
cur_ans = ''
for j, word in enumerate(tmp_ans):
cur_ans += word
if cur_space[j]:
cur_ans += ' '
answer.append(cur_ans.strip())
i += 1
cnt += i
logging.info('batch=%d/%d' % (bnum, batch_cnt))
# manual release memory, todo: really effect?
del bat_context, bat_answer_range, batch, batch_input
del tmp_ans_range
# torch.cuda.empty_cache()
return answer
def build_data(self, contexts, question):
contexts_ids = []
questions_ids = []
answers_range_wid = [
] # each answer use the [start,end] representation, all the answer horizontal concat
samples_id = []
cnt = 0
for i, context in enumerate(contexts):
context_doc = DocText(self.nlp, context,
self.global_config['preprocess'])
question_doc = DocText(self.nlp, question[i],
self.global_config['preprocess'])
context_doc.update_em(question_doc)
question_doc.update_em(context_doc)
context_f = context_doc.to_id_dict(self.metadata)
question_f = question_doc.to_id_dict(self.metadata)
contexts_ids.append(context_f)
questions_ids.append(question_f)
samples_id.append(str(cnt))
answers_range_wid.append([0, 1])
cnt += 1
return {
'context': contexts_ids,
'question': questions_ids,
'answer_range': answers_range_wid,
'samples_id': samples_id
}
def dict2array(self, data_doc, use_domain_tag=False):
"""
transform dict to numpy array
:param data_doc: [{'token': [], 'pos': [], 'ent': [], 'em': [], 'em_lemma': [], 'right_space': [], 'domain_tag':[]]
:return:
"""
data = {
'token': [],
'pos': [],
'ent': [],
'em': [],
'em_lemma': [],
'right_space': []
}
if use_domain_tag:
data['domain_tag'] = []
max_len = 0
for ele in data_doc:
assert ele.keys() == data.keys()
if len(ele['token']) > max_len:
max_len = len(ele['token'])
for k in ele.keys():
if len(ele[k]) > 0:
data[k].append(ele[k])
for k in data.keys():
if len(data[k]) > 0:
data[k] = pad_sequences(data[k],
maxlen=max_len,
padding='post',
value=0)
return data
def main():
logger.info('------------MODEL TEST INPUT--------------')
logger.info('loading config file...')
global_config = read_config()
# set random seed
seed = global_config['global']['random_seed']
torch.manual_seed(seed)
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 = BaseModel(dataset_h5_path,
model_config=read_config('config/base_model.yaml'))
elif model_choose == 'match-lstm':
model = MatchLSTM(dataset_h5_path)
elif model_choose == 'match-lstm+':
model = MatchLSTMPlus(dataset_h5_path)
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.eval() # let training = False, make sure right dropout
logging.info('model parameters count: %d' % count_parameters(model))
# load model weight
logger.info('loading model weight...')
model_weight_path = global_config['data']['model_path']
is_exist_model_weight = os.path.exists(model_weight_path)
assert is_exist_model_weight, "not found model weight file on '%s'" % model_weight_path
weight = torch.load(model_weight_path, map_location=lambda storage, loc: storage)
model.load_state_dict(weight, strict=False)
# manual input qa
context = "In 1870, Tesla moved to Karlovac, to attend school at the Higher Real Gymnasium, where he was " \
"profoundly influenced by a math teacher Martin Sekuli\u0107.:32 The classes were held in German, " \
"as it was a school within the Austro-Hungarian Military Frontier. Tesla was able to perform integral " \
"calculus in his head, which prompted his teachers to believe that he was cheating. He finished a " \
"four-year term in three years, graduating in 1873.:33 "
question1 = "What language were classes held in at Tesla's school?"
answer1 = ["German"]
question2 = "Who was Tesla influenced by while in school?"
answer2 = ["Martin Sekuli\u0107"]
question3 = "Why did Tesla go to Karlovac?"
answer3 = ["attend school at the Higher Real Gymnasium", 'to attend school']
# change here to select questions
question = question1
answer = answer1[0]
# preprocess
nlp = spacy.load('en')
context_doc = DocText(nlp, context, global_config['preprocess'])
question_doc = DocText(nlp, question, global_config['preprocess'])
context_doc.update_em(question_doc)
question_doc.update_em(context_doc)
context_token = context_doc.token
question_token = question_doc.token
context_id_char = to_long_tensor(dataset.sentence_char2id(context_token))
question_id_char = to_long_tensor(dataset.sentence_char2id(question_token))
context_id, context_f = context_doc.to_id(dataset.meta_data)
question_id, question_f = question_doc.to_id(dataset.meta_data)
bat_input = [context_id, question_id, context_id_char, question_id_char, context_f, question_f]
bat_input = [x.unsqueeze(0) if x is not None else x for x in bat_input]
out_ans_prop, out_ans_range, vis_param = model.forward(*bat_input)
out_ans_range = out_ans_range.numpy()
start = out_ans_range[0][0]
end = out_ans_range[0][1] + 1
out_answer_id = context_id[start:end]
out_answer = dataset.sentence_id2word(out_answer_id)
logging.info('Predict Answer: ' + ' '.join(out_answer))
# to show on visdom
s = 0
e = 48
x_left = vis_param['match']['left']['alpha'][0, :, s:e].numpy()
x_right = vis_param['match']['right']['alpha'][0, :, s:e].numpy()
x_left_gated = vis_param['match']['left']['gated'][0, :, s:e].numpy()
x_right_gated = vis_param['match']['right']['gated'][0, :, s:e].numpy()
draw_heatmap_sea(x_left,
xlabels=context_token[s:e],
ylabels=question_token,
answer=answer,
save_path='data/test-left.png',
bottom=0.45)
draw_heatmap_sea(x_right,
xlabels=context_token[s:e],
ylabels=question_token,
answer=answer,
save_path='data/test-right.png',
bottom=0.45)
enable_self_match = False
if enable_self_match:
x_self_left = vis_param['self']['left']['alpha'][0, s:e, s:e].numpy()
x_self_right = vis_param['self']['right']['alpha'][0, s:e, s:e].numpy()
draw_heatmap_sea(x_self_left,
xlabels=context_token[s:e],
ylabels=context_token[s:e],
answer=answer,
save_path='data/test-self-left.png',
inches=(11, 11),
bottom=0.2)
draw_heatmap_sea(x_self_right,
xlabels=context_token[s:e],
ylabels=context_token[s:e],
answer=answer,
save_path='data/test-self-right.png',
inches=(11, 11),
bottom=0.2)
def main():
logger.info('------------MODEL TEST INPUT--------------')
logger.info('loading config file...')
global_config = read_config()
# set random seed
seed = global_config['global']['random_seed']
torch.manual_seed(seed)
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 = BaseModel(dataset_h5_path,
model_config=read_config('config/base_model.yaml'))
elif model_choose == 'match-lstm':
model = MatchLSTM(dataset_h5_path)
elif model_choose == 'match-lstm+':
model = MatchLSTMPlus(dataset_h5_path)
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.eval() # let training = False, make sure right dropout
logging.info('model parameters count: %d' % count_parameters(model))
# load model weight
logger.info('loading model weight...')
model_weight_path = global_config['data']['model_path']
is_exist_model_weight = os.path.exists(model_weight_path)
assert is_exist_model_weight, "not found model weight file on '%s'" % model_weight_path
weight = torch.load(model_weight_path,
map_location=lambda storage, loc: storage)
model.load_state_dict(weight, strict=False)
# manual input qa
context = "In 1870, Tesla moved to Karlovac, to attend school at the Higher Real Gymnasium, where he was " \
"profoundly influenced by a math teacher Martin Sekuli\u0107.:32 The classes were held in German, " \
"as it was a school within the Austro-Hungarian Military Frontier. Tesla was able to perform integral " \
"calculus in his head, which prompted his teachers to believe that he was cheating. He finished a " \
"four-year term in three years, graduating in 1873.:33 "
question1 = "What language were classes held in at Tesla's school?"
answer1 = ["German"]
question2 = "Who was Tesla influenced by while in school?"
answer2 = ["Martin Sekuli\u0107"]
question3 = "Why did Tesla go to Karlovac?"
answer3 = [
"attend school at the Higher Real Gymnasium", 'to attend school'
]
# change here to select questions
question = question1
answer = answer1[0]
# preprocess
nlp = spacy.load('en')
context_doc = DocText(nlp, context, global_config['preprocess'])
question_doc = DocText(nlp, question, global_config['preprocess'])
context_doc.update_em(question_doc)
question_doc.update_em(context_doc)
context_token = context_doc.token
question_token = question_doc.token
context_id_char = to_long_tensor(dataset.sentence_char2id(context_token))
question_id_char = to_long_tensor(dataset.sentence_char2id(question_token))
context_id, context_f = context_doc.to_id(dataset.meta_data)
question_id, question_f = question_doc.to_id(dataset.meta_data)
bat_input = [
context_id, question_id, context_id_char, question_id_char, context_f,
question_f
]
bat_input = [x.unsqueeze(0) if x is not None else x for x in bat_input]
out_ans_prop, out_ans_range, vis_param = model.forward(*bat_input)
out_ans_range = out_ans_range.numpy()
start = out_ans_range[0][0]
end = out_ans_range[0][1] + 1
out_answer_id = context_id[start:end]
out_answer = dataset.sentence_id2word(out_answer_id)
logging.info('Predict Answer: ' + ' '.join(out_answer))
# to show on visdom
s = 0
e = 48
x_left = vis_param['match']['left']['alpha'][0, :, s:e].numpy()
x_right = vis_param['match']['right']['alpha'][0, :, s:e].numpy()
x_left_gated = vis_param['match']['left']['gated'][0, :, s:e].numpy()
x_right_gated = vis_param['match']['right']['gated'][0, :, s:e].numpy()
draw_heatmap_sea(x_left,
xlabels=context_token[s:e],
ylabels=question_token,
answer=answer,
save_path='data/test-left.png',
bottom=0.45)
draw_heatmap_sea(x_right,
xlabels=context_token[s:e],
ylabels=question_token,
answer=answer,
save_path='data/test-right.png',
bottom=0.45)
enable_self_match = False
if enable_self_match:
x_self_left = vis_param['self']['left']['alpha'][0, s:e, s:e].numpy()
x_self_right = vis_param['self']['right']['alpha'][0, s:e, s:e].numpy()
draw_heatmap_sea(x_self_left,
xlabels=context_token[s:e],
ylabels=context_token[s:e],
answer=answer,
save_path='data/test-self-left.png',
inches=(11, 11),
bottom=0.2)
draw_heatmap_sea(x_self_right,
xlabels=context_token[s:e],
ylabels=context_token[s:e],
answer=answer,
save_path='data/test-self-right.png',
inches=(11, 11),
bottom=0.2)
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():
logger.info('------------Match-LSTM TEST INPUT--------------')
logger.info('loading config file...')
global_config = read_config()
# set random seed
seed = global_config['model']['global']['random_seed']
torch.manual_seed(seed)
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)
model.eval() # let training = False, make sure right dropout
logging.info('model parameters count: %d' % count_parameters(model))
# load model weight
logger.info('loading model weight...')
model_weight_path = global_config['data']['model_path']
is_exist_model_weight = os.path.exists(model_weight_path)
assert is_exist_model_weight, "not found model weight file on '%s'" % model_weight_path
weight = torch.load(model_weight_path,
map_location=lambda storage, loc: storage)
model.load_state_dict(weight, strict=False)
# manual input qa
context = "In 1870, Tesla moved to Karlovac, to attend school at the Higher Real Gymnasium, where he was " \
"profoundly influenced by a math teacher Martin Sekuli\u0107.:32 The classes were held in German, " \
"as it was a school within the Austro-Hungarian Military Frontier. Tesla was able to perform integral " \
"calculus in his head, which prompted his teachers to believe that he was cheating. He finished a " \
"four-year term in three years, graduating in 1873.:33 "
question1 = "What language were classes held in at Tesla's school?"
answer1 = ["German"]
question2 = "Who was Tesla influenced by while in school?"
answer2 = ["Martin Sekuli\u0107"]
question3 = "Why did Tesla go to Karlovac?"
answer3 = [
"attend school at the Higher Real Gymnasium", 'to attend school'
]
# change here to select questions
question = question1
answer = answer1[0]
# preprocess
context_token = nltk.word_tokenize(context)
question_token = nltk.word_tokenize(question)
a = np.array(context_token)
context_id = dataset.sentence_word2id(context_token)
question_id = dataset.sentence_word2id(question_token)
context_id_char = dataset.sentence_char2id(context_token)
question_id_char = dataset.sentence_char2id(question_token)
context_var, question_var, context_var_char, question_var_char = [
to_long_tensor(x).unsqueeze(0)
for x in [context_id, question_id, context_id_char, question_id_char]
]
out_ans_prop, out_ans_range, vis_param = model.forward(
context_var, question_var, context_var_char, question_var_char)
out_ans_range = out_ans_range.cpu().data.numpy()
start = out_ans_range[0][0]
end = out_ans_range[0][1] + 1
out_answer_id = context_id[start:end]
out_answer = dataset.sentence_id2word(out_answer_id)
logging.info('Predict Answer: ' + ' '.join(out_answer))
# to show on visdom
s = 0
e = 48
x_left = vis_param['match']['left'][0, :, s:e].cpu().data.numpy()
x_right = vis_param['match']['right'][0, :, s:e].cpu().data.numpy()
draw_heatmap_sea(x_left,
xlabels=context_token[s:e],
ylabels=question_token,
answer=answer,
save_path='data/test-left.png',
bottom=0.45)
draw_heatmap_sea(x_right,
xlabels=context_token[s:e],
ylabels=question_token,
answer=answer,
save_path='data/test-right.png',
bottom=0.45)
if global_config['model']['interaction']['enable_self_match']:
x_self_left = vis_param['self']['left'][0, s:e, s:e].cpu().data.numpy()
x_self_right = vis_param['self']['right'][0, s:e,
s:e].cpu().data.numpy()
draw_heatmap_sea(x_self_left,
xlabels=context_token[s:e],
ylabels=context_token[s:e],
answer=answer,
save_path='data/test-self-left.png',
inches=(11, 11),
bottom=0.2)
draw_heatmap_sea(x_self_right,
xlabels=context_token[s:e],
ylabels=context_token[s:e],
answer=answer,
save_path='data/test-self-right.png',
inches=(11, 11),
bottom=0.2)
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.')
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 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)
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']
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:
criterion = MyNLLLoss()
score_em, score_f1, sum_loss = eval_on_model(model=model,
criterion=criterion,
batch_data=batch_dev_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()
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)
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.')