def load_model_decode(data, name):
print("Load Model from file: ", data.model_dir)
model = SeqLabel(data)
model.load_state_dict(torch.load(data.load_model_dir))
pytorch_total_params = sum(p.numel() for p in model.parameters())
print("Number_parameters:", pytorch_total_params)
print("Decode %s data, nbest: %s ..." % (name, data.nbest))
start_time = time.time()
summary = evaluate(data, model, name, True, data.nbest)
pred_results_tasks = []
pred_scores_tasks = []
range_tasks = len(data.index_of_main_tasks)
for idtask in range(range_tasks):
speed, acc, p, r, f, pred_results, pred_scores = summary[idtask]
pred_results_tasks.append(pred_results)
pred_scores_tasks.append(pred_scores)
end_time = time.time()
time_cost = end_time - start_time
if data:
print(
"%s: time:%.2fs, speed:%.2fst/s; acc: %.4f, p: %.4f, r: %.4f, f: %.4f"
% (name, time_cost, speed, acc, p, r, f))
else:
print("%s: time:%.2fs, speed:%.2fst/s; acc: %.4f" %
(name, time_cost, speed, acc))
return pred_results_tasks, pred_scores_tasks
def load_model_decode(data, name):
print("Load Model from file: ", data.model_dir)
model = SeqLabel(data)
## load model need consider if the model trained in GPU and load in CPU, or vice versa
# if not gpu:
# model.load_state_dict(torch.load(model_dir))
# # model.load_state_dict(torch.load(model_dir), map_location=lambda storage, loc: storage)
# # model = torch.load(model_dir, map_location=lambda storage, loc: storage)
# else:
# model.load_state_dict(torch.load(model_dir))
# # model = torch.load(model_dir)
model.load_state_dict(torch.load(data.load_model_dir))
print("Decode %s data, nbest: %s ..." % (name, data.nbest))
start_time = time.time()
speed, acc, p, r, f, pred_results, pred_scores = evaluate(
data, model, name, data.nbest)
end_time = time.time()
time_cost = end_time - start_time
if data.seg:
print(
"%s: time:%.2fs, speed:%.2fst/s; acc: %.4f, p: %.4f, r: %.4f, f: %.4f"
% (name, time_cost, speed, acc, p, r, f))
else:
print("%s: time:%.2fs, speed:%.2fst/s; acc: %.4f" %
(name, time_cost, speed, acc))
return pred_results, pred_scores
def load_model_test(data, name):
print("Load Model from file: ", data.dset_dir)
model = SeqLabel(data)
model.load_state_dict(torch.load(data.load_model_dir))
start_time = time.time()
speed, p, r, f, pred_results, pred_scores = evaluate(data, model, name)
end_time = time.time()
time_cost = end_time - start_time
if data.seg:
print("Test: time: %.2fs, speed: %.2fst/s; [p: %.4f, r: %.4f, f: %.4f]"%(time_cost, speed, p, r, f))
else:
print("%s: time:%.2fs, speed:%.2fst/s; acc: %.4f"%(name, time_cost, speed, acc))
return pred_results, pred_scores
def train(data):
save_data_name = data.model_dir + ".dset"
data.save(save_data_name)
model = SeqLabel(data)
# 加载预训练
print('loading model %s' % model_path)
model.load_state_dict(torch.load(model_path, map_location=map_location))
print('data.seg:', data.seg)
optimizer = ''
if data.optimizer.lower() == "sgd":
optimizer = optim.SGD(model.parameters(),
lr=data.HP_lr,
momentum=data.HP_momentum,
weight_decay=data.HP_l2)
best_dev = -10
print('data.HP_gpu:', data.HP_gpu)
for idx in range(data.HP_iteration):
epoch_start = time.time()
temp_start = epoch_start
print("Epoch: %s/%s" % (idx, data.HP_iteration))
if data.optimizer == "SGD":
optimizer = lr_decay(optimizer, idx, data.HP_lr_decay, data.HP_lr)
instance_count = 0
sample_id = 0
sample_loss = 0 # 每500个batch清零
total_loss = 0 # 一个epoch里的完整loss
right_token = 0
whole_token = 0
# print("Before Shuffle: first input word list:", data.train_Ids[0][0])
random.shuffle(data.train_Ids)
print("Shuffle: first input word list:", data.train_Ids[0][0])
model.train()
model.zero_grad()
batch_size = data.HP_batch_size
# batch_id = 0
train_num = len(data.train_Ids)
print('train_num:', train_num) # 训练样本的数量
total_batch = train_num // batch_size + 1
print('total_batch:', total_batch)
for batch_id in range(total_batch):
start = batch_id * batch_size
end = (batch_id + 1) * batch_size
if end > train_num:
end = train_num
instance = data.train_Ids[start:end]
if not instance:
continue
batch_word, batch_features, batch_wordlen, batch_wordrecover, batch_char, batch_charlen, batch_charrecover, \
batch_label, mask = batchify_with_label(instance, data.HP_gpu, True, data.sentence_classification)
instance_count += 1
loss, tag_seq = model.calculate_loss(batch_word, batch_features,
batch_wordlen, batch_char,
batch_charlen,
batch_charrecover,
batch_label, mask)
right, whole = predict_check(
tag_seq, batch_label, mask,
data.sentence_classification) # pred与gold的校验
right_token += right
whole_token += whole
sample_loss += loss.item()
total_loss += loss.item()
if end % 500 == 0:
temp_time = time.time()
temp_cost = temp_time - temp_start
temp_start = temp_time
print(
" Instance: %s; Time: %.2fs; loss: %.4f; acc: %s/%s=%.4f"
% (end, temp_cost, sample_loss, right_token, whole_token,
(right_token + 0.) / whole_token))
if sample_loss > 1e8 or str(sample_loss) == "nan":
print(
"ERROR: LOSS EXPLOSION (>1e8) ! PLEASE SET PROPER PARAMETERS AND STRUCTURE! EXIT...."
)
exit(1)
sys.stdout.flush()
sample_loss = 0
loss.backward()
optimizer.step()
model.zero_grad()
temp_time = time.time()
temp_cost = temp_time - temp_start
print(" Instance: %s; Time: %.2fs; loss: %.4f; acc: %s/%s=%.4f" %
(end, temp_cost, sample_loss, right_token, whole_token,
(right_token + 0.) / whole_token))
epoch_finish = time.time()
epoch_cost = epoch_finish - epoch_start
print(
"Epoch: %s training finished. Time: %.2fs, speed: %.2fst/s, total loss: %s"
% (idx, epoch_cost, train_num / epoch_cost, total_loss))
print("total_loss:", total_loss)
if total_loss > 1e8 or str(total_loss) == "nan":
print(
"ERROR: LOSS EXPLOSION (>1e8) ! PLEASE SET PROPER PARAMETERS AND STRUCTURE! EXIT...."
)
exit(1)
speed, acc, p, r, f, _, _ = evaluate(data, model, "dev")
dev_finish = time.time()
dev_cost = dev_finish - epoch_finish
current_score = f
print(
"Dev: time: %.2fs, speed: %.2fst/s; acc: %.4f, p: %.4f, r: %.4f, f: %.4f"
% (dev_cost, speed, acc, p, r, f))
if current_score > best_dev:
print("Exceed previous best f score:", best_dev)
model_name = data.model_dir + '.' + str(idx) + ".model"
print("Save current best torch_model in file:",
model_name) # 保存当前epoch结束的模型
torch.save(model.state_dict(), model_name)
best_dev = current_score
# 每50轮保存一下
if idx % 50 == 0:
model_name = data.model_dir + '.' + str(idx) + ".model"
print('Save every 50 epoch in file: %s' % model_name)
torch.save(model.state_dict(), model_name)
speed, acc, p, r, f, _, _ = evaluate(data, model, "test")
test_finish = time.time()
test_cost = test_finish - dev_finish
print(
"Test: time: %.2fs, speed: %.2fst/s; acc: %.4f, p: %.4f, r: %.4f, f: %.4f"
% (test_cost, speed, acc, p, r, f))
# 对自己add对样本做一下evaluate:
speed, acc, p, r, f, _, _ = evaluate(data, model, "raw")
raw_finish = time.time()
raw_cost = raw_finish - test_finish
print(
"Raw: time: %.2fs, speed: %.2fst/s; acc: %.4f, p: %.4f, r: %.4f, f: %.4f"
% (raw_cost, speed, acc, p, r, f))
gc.collect()
model_name = data.model_dir + '.' + str(idx) + ".model"
print('Save every 50 epoch in file: %s' % model_name)
torch.save(model.state_dict(), model_name)
speed, acc, p, r, f, _, _ = evaluate(data, model, "test")
test_finish = time.time()
test_cost = test_finish - dev_finish
print(
"Test: time: %.2fs, speed: %.2fst/s; acc: %.4f, p: %.4f, r: %.4f, f: %.4f"
% (test_cost, speed, acc, p, r, f))
# 对自己add对样本做一下evaluate:
speed, acc, p, r, f, _, _ = evaluate(data, model, "raw")
raw_finish = time.time()
raw_cost = raw_finish - test_finish
print(
"Raw: time: %.2fs, speed: %.2fst/s; acc: %.4f, p: %.4f, r: %.4f, f: %.4f"
% (raw_cost, speed, acc, p, r, f))
gc.collect()
if __name__ == '__main__':
# train(data)
model = SeqLabel(data)
model.load_state_dict(
torch.load('transfer_model/transfer.1.model',
map_location=map_location))
speed, acc, p, r, f, _, _ = evaluate(data, model, "raw")
raw_finish = time.time()
class NCRF:
def __init__(self):
# print("Python Version: %s.%s"%(sys.version_info[0],sys.version_info[1]))
# print("PyTorch Version:%s"%(torch.__version__))
# print("Process ID: ", os.getpid())
self.data = Data()
self.data.HP_gpu = torch.cuda.is_available()
if self.data.HP_gpu:
self.data.device = 'cuda'
# print("GPU:", self.data.HP_gpu, "; device:", self.data.device)
self.optimizer = None
self.model = None
def read_data_config_file(self, config_dir):
self.data.read_config(config_dir)
def manual_data_setting(self, setting_dict):
## set data through manual dict, all value should be in string format.
self.data.manual_config(setting_dict)
def initialize_model_and_optimizer(self):
if self.data.sentence_classification:
self.model = SentClassifier(self.data)
else:
self.model = SeqLabel(self.data)
if self.data.optimizer.lower() == "sgd":
self.optimizer = optim.SGD(self.model.parameters(),
lr=self.data.HP_lr,
momentum=self.data.HP_momentum,
weight_decay=self.data.HP_l2)
elif self.data.optimizer.lower() == "adagrad":
self.optimizer = optim.Adagrad(self.model.parameters(),
lr=self.data.HP_lr,
weight_decay=self.data.HP_l2)
elif self.data.optimizer.lower() == "adadelta":
self.optimizer = optim.Adadelta(self.model.parameters(),
lr=self.data.HP_lr,
weight_decay=self.data.HP_l2)
elif self.data.optimizer.lower() == "rmsprop":
self.optimizer = optim.RMSprop(self.model.parameters(),
lr=self.data.HP_lr,
weight_decay=self.data.HP_l2)
elif self.data.optimizer.lower() == "adam":
self.optimizer = optim.Adam(self.model.parameters(),
lr=self.data.HP_lr,
weight_decay=self.data.HP_l2)
else:
print("Optimizer illegal: %s" % (self.data.optimizer))
exit(1)
def initialize_data(self, input_list=None):
self.data.initial_alphabets(input_list)
if self.data.use_word_emb and self.data.use_word_seq:
self.data.build_pretrain_emb()
def initialization(self, input_list=None):
## must initialize data before initialize model and optimizer, as alphabet size and pretrain emb matters
self.num_ = '''
input_list: [train_list, dev_list, test_list]
train_list/dev_list/test_list: [sent_list, label_list, feature_list]
sent_list: list of list [[word1, word2,...],...,[wordx, wordy]...]
label_list: if sentence_classification:
list of labels [label1, label2,...labelx, labely,...]
else:
list of list [[label1, label2,...],...,[labelx, labely,...]]
feature_list: if sentence_classification:
list of labels [[feat1, feat2,..],...,[feat1, feat2,..]], len(feature_list)= sentence_num
else:
list of list [[[feat1, feat2,..],...,[feat1, feat2,..]],...,[[feat1, feat2,..],...,[feat1, feat2,..]]], , len(feature_list)= sentence_num
'''
self.initialize_data(input_list)
self.and_optimizer = self.initialize_model_and_optimizer()
def self_generate_instances(self):
self.data.generate_instance('train')
self.data.generate_instance('dev')
self.data.generate_instance('test')
def generate_instances_from_list(self, input_list, name):
return self.data.generate_instance_from_list(input_list, name)
def save(self, model_dir="ncrf.model"):
# print("Save model to file: ", model_dir)
the_dict = {
'data': self.data,
'state_dict': self.model.state_dict(),
'optimizer': self.optimizer.state_dict()
}
torch.save(the_dict, model_dir)
def load(self, model_dir="ncrf.model"):
the_dict = torch.load(model_dir)
self.data = the_dict['data']
self.data.silence = True
## initialize the model and optimizer befor load state dict
self.initialize_model_and_optimizer()
self.model.load_state_dict(the_dict['state_dict'])
self.optimizer.load_state_dict(the_dict['optimizer'])
# print("Model loaded from file: ", model_dir)
def train(self, train_Ids=None, save_model_dir=None):
'''
train_Ids: list of words, chars and labels, various length. [[words, features, chars, labels],[words, features, chars,labels],...]
words: word ids for one sentence. (batch_size, sent_len)
features: features ids for one sentence. (batch_size, sent_len, feature_num)
chars: char ids for on sentences, various length. (batch_size, sent_len, each_word_length)
labels: label ids for one sentence. (batch_size, sent_len)
save_model_dir: model name to be saved
'''
if train_Ids:
self.data.train_Ids = train_Ids
# print(self.data.train_Ids[0])
print('-----begin train------')
# exit(0)
best_dev = -10
best_model = None
for idx in range(self.data.HP_iteration):
epoch_start = time.time()
temp_start = epoch_start
# print("Epoch: %s/%s" %(idx,self.data.HP_iteration))
if self.data.optimizer == "SGD":
self.optimizer = lr_decay(self.optimizer, idx,
self.data.HP_lr_decay,
self.data.HP_lr)
instance_count = 0
sample_id = 0
sample_loss = 0
total_loss = 0
right_token = 0
whole_token = 0
random.shuffle(self.data.train_Ids)
first_list = ", ".join([
self.data.word_alphabet.get_instance(a)
for a in self.data.train_Ids[0][0]
])
# print("Shuffle: first input: [%s]" %(first_list))
## set model in train model
self.model.train()
batch_size = self.data.HP_batch_size
batch_id = 0
train_num = len(self.data.train_Ids)
total_batch = train_num // batch_size + 1
for batch_id in range(total_batch):
self.optimizer.zero_grad()
start = batch_id * batch_size
end = (batch_id + 1) * batch_size
if end > train_num:
end = train_num
instance = self.data.train_Ids[start:end]
if not instance:
continue
batch_word, batch_features, batch_wordlen, batch_wordrecover, batch_char, batch_charlen, batch_charrecover, batch_word_text, batch_label, mask = batchify_with_label(
instance, self.data.HP_gpu, True,
self.data.sentence_classification)
instance_count += 1
loss, tag_seq = self.model.calculate_loss(
batch_word, batch_features, batch_wordlen, batch_char,
batch_charlen, batch_charrecover, batch_word_text,
batch_label, mask)
right, whole = predict_check(tag_seq, batch_label, mask,
self.data.sentence_classification)
right_token += right
whole_token += whole
# print("loss:",loss.item())
sample_loss += loss.item()
total_loss += loss.item()
if end % 300000 == 0:
temp_time = time.time()
temp_cost = temp_time - temp_start
temp_start = temp_time
print(
" Instance: %s; Time: %.2fs; loss: %.4f; acc: %s/%s=%.4f"
%
(end, temp_cost, sample_loss, right_token, whole_token,
(right_token + 0.) / whole_token))
if sample_loss > 1e8 or str(sample_loss) == "nan":
print(
"ERROR: LOSS EXPLOSION (>1e8) ! PLEASE SET PROPER PARAMETERS AND STRUCTURE! EXIT...."
)
exit(1)
sys.stdout.flush()
sample_loss = 0
loss.backward()
self.optimizer.step()
temp_time = time.time()
temp_cost = temp_time - temp_start
# print(" Instance: %s; Time: %.2fs; loss: %.4f; acc: %s/%s=%.4f"%(end, temp_cost, sample_loss, right_token, whole_token,(right_token+0.)/whole_token))
epoch_finish = time.time()
epoch_cost = epoch_finish - epoch_start
print(
"Epoch: %s training finished. Time: %.2fs, speed: %.2fst/s, total loss: %s"
% (idx, epoch_cost, train_num / epoch_cost, total_loss))
# print("totalloss:", total_loss)
if total_loss > 1e8 or str(total_loss) == "nan":
print(
"ERROR: LOSS EXPLOSION (>1e8) ! PLEASE SET PROPER PARAMETERS AND STRUCTURE! EXIT...."
)
exit(1)
# continue
speed, f = evaluate(self.data, self.model, "dev")
dev_finish = time.time()
dev_cost = dev_finish - epoch_finish
current_score = f
print("Dev: time: %.2fs, speed: %.2fst/s; f: %.4f" %
(dev_cost, speed, f))
if current_score > best_dev:
# if self.data.seg:
print("Exceed previous best f score:", best_dev)
_, f = evaluate(self.data, self.model, "test")
print("Test: f: %.4f" % (f))
# _ , f = evaluate(self.data, self.model, "test")
# if self.data.seg:
# print("Test: f: %.4f"%(f))
# else:
# print("Exceed previous best f score:", best_dev)
# if save_model_dir == None:
# model_name = self.data.model_dir + ".model"
# else:
# model_name = save_model_dir + ".model"
# self.save(model_name)
# torch.save(model.state_dict(), model_name)
best_dev = current_score
# best_model = model_name
## decode test
# else:
# print("Test: time: %.2fs, speed: %.2fst/s; acc: %.4f"%(test_cost, speed, acc))
gc.collect()
# if best_model != None:
# self.load(best_model)
# _ , f = evaluate(self.data, self.model, "test")
# print("Test: f: %.4f"%(f))
# def evaluate(self):
def decode(self, raw_Ids):
'''
raw_Ids: list of words, chars and labels, various length. [[words, features, chars, labels],[words, features, chars,labels],...]
words: word ids for one sentence. (batch_size, sent_len)
features: features ids for one sentence. (batch_size, sent_len, feature_num)
chars: char ids for on sentences, various length. (batch_size, sent_len, each_word_length)
labels: label ids for one sentence. (batch_size, sent_len)
## label should be padded in raw input
'''
instances = raw_Ids
## set model in eval model
self.model.eval()
batch_size = self.data.HP_batch_size
instance_num = len(instances)
total_batch = instance_num // batch_size + 1
decode_label = []
for batch_id in tqdm(range(total_batch)):
start = batch_id * batch_size
end = (batch_id + 1) * batch_size
if end > instance_num:
end = instance_num
instance = instances[start:end]
if not instance:
continue
batch_word, batch_features, batch_wordlen, batch_wordrecover, batch_char, batch_charlen, batch_charrecover, batch_word_text, batch_label, mask = batchify_with_label(
instance, self.data.HP_gpu, False,
self.data.sentence_classification)
tag_seq = self.model(batch_word, batch_features, batch_wordlen,
batch_char, batch_charlen, batch_charrecover,
batch_word_text, None, mask)
tag_seq = tag_seq[batch_wordrecover.cpu()]
decode_label += tag_seq.cpu().data.numpy().tolist()
return decode_label
def decode_prob(self, raw_Ids):
'''
raw_Ids: list of words, chars and labels, various length. [[words, features, chars, labels],[words, features, chars,labels],...]
words: word ids for one sentence. (batch_size, sent_len)
features: features ids for one sentence. (batch_size, sent_len, feature_num)
chars: char ids for on sentences, various length. (batch_size, sent_len, each_word_length)
labels: label ids for one sentence. (batch_size, sent_len)
## label should be padded in raw input
'''
if not self.data.sentence_classification:
print(
"decode probability is only valid in sentence classification task. Exit."
)
exit(0)
instances = raw_Ids
target_probability_list = []
target_result_list = []
## set model in eval model
self.model.eval()
batch_size = self.data.HP_batch_size
instance_num = len(instances)
total_batch = instance_num // batch_size + 1
for batch_id in tqdm(range(total_batch)):
start = batch_id * batch_size
end = (batch_id + 1) * batch_size
if end > instance_num:
end = instance_num
instance = instances[start:end]
if start % 10000 == 0:
print("Decode: ", start)
if not instance:
continue
batch_word, batch_features, batch_wordlen, batch_wordrecover, batch_char, batch_charlen, batch_charrecover, batch_word_text, batch_label, mask = batchify_with_label(
instance, self.data.HP_gpu, False,
self.data.sentence_classification)
target_probability, _ = self.model.get_target_probability(
batch_word, batch_features, batch_wordlen, batch_char,
batch_charlen, batch_charrecover, batch_word_text, None, mask)
target_probability = target_probability[batch_wordrecover.cpu()]
target_probability_list.append(target_probability)
target_probabilities = np.concatenate(target_probability_list, axis=0)
return target_probabilities
def decode_prob_and_attention_weights(self, raw_Ids):
'''
raw_Ids: list of words, chars and labels, various length. [[words, features, chars, labels],[words, features, chars,labels],...]
words: word ids for one sentence. (batch_size, sent_len)
features: features ids for one sentence. (batch_size, sent_len, feature_num)
chars: char ids for on sentences, various length. (batch_size, sent_len, each_word_length)
labels: label ids for one sentence. (batch_size, sent_len)
## label should be padded in raw input
'''
if not self.data.sentence_classification:
print(
"decode probability is only valid in sentence classification task. Exit."
)
exit(0)
if self.data.words2sent_representation.upper(
) != "ATTENTION" and self.data.words2sent_representation.upper(
) != "ATT":
print(
"attention weights are only valid in attention model. Current: %s, Exit."
% (self.data.words2sent_representation))
exit(0)
instances = raw_Ids
target_probability_list = []
sequence_attention_weight_list = []
## set model in eval model
self.model.eval()
batch_size = self.data.HP_batch_size
instance_num = len(instances)
total_batch = instance_num // batch_size + 1
for batch_id in tqdm(range(total_batch)):
start = batch_id * batch_size
end = (batch_id + 1) * batch_size
if end > instance_num:
end = instance_num
instance = instances[start:end]
if start % 10000 == 0:
print("Decode: ", start)
if not instance:
continue
batch_word, batch_features, batch_wordlen, batch_wordrecover, batch_char, batch_charlen, batch_charrecover, batch_word_text, batch_label, mask = batchify_with_label(
instance, self.data.HP_gpu, False,
self.data.sentence_classification)
target_probability, weights = self.model.get_target_probability(
batch_word, batch_features, batch_wordlen, batch_char,
batch_charlen, batch_charrecover, batch_word_text, None, mask)
## target_probability, weights are both numpy
target_probability = target_probability[batch_wordrecover.cpu()]
weights = weights[batch_wordrecover.cpu()]
target_probability_list.append(target_probability)
sequence_attention_weight_list += weights.tolist()
target_probabilities = np.concatenate(target_probability_list, axis=0)
print(len(sequence_attention_weight_list))
## sequence_attention_weight_list: list with different batch size and many padded 0
return target_probabilities, sequence_attention_weight_list
def load_ncrf_model(data):
model = SeqLabel(data)
print('loading model:', data.load_model_dir)
model.load_state_dict(
torch.load(data.load_model_dir, map_location=torch.device('cpu')))
return model
