def setup_train(self, model_file_path=None):
self.model = Model(model_file_path)
params = list(self.model.encoder.parameters()) + list(self.model.decoder.parameters()) + \
list(self.model.reduce_state.parameters())
initial_lr = config.lr_coverage if config.is_coverage else config.lr
self.optimizer = Adagrad(
params,
lr=initial_lr,
initial_accumulator_value=config.adagrad_init_acc)
start_iter, start_loss = 0, 0
if model_file_path is not None:
state = torch.load(model_file_path,
map_location=lambda storage, location: storage)
start_iter = state['iter']
start_loss = state['current_loss']
if not config.is_coverage:
self.optimizer.load_state_dict(state['optimizer'])
if use_cuda:
for state in self.optimizer.state.values():
for k, v in state.items():
if torch.is_tensor(v):
state[k] = v.cuda()
return start_iter, start_loss
def setup_train(self, model_file_path=None):
self.model = Model(model_file_path, vectors=self.vectors)
params = list(self.model.encoder.parameters()) + list(self.model.decoder.parameters()) + \
list(self.model.reduce_state.parameters())
pytorch_total_params = sum(p.numel() for p in params if p.requires_grad)
print(f"Parameters count: {pytorch_total_params}")
initial_lr = config.lr_coverage if config.is_coverage else config.lr
# self.optimizer = adagrad.Adagrad(params, lr=initial_lr, initial_accumulator_value=config.adagrad_init_acc)
self.optimizer = Adam(params, lr=initial_lr)
start_iter, start_training_loss, start_eval_loss = 0, 0, 0
if model_file_path is not None:
state = torch.load(model_file_path, map_location=lambda storage, location: storage)
start_iter = state['iter']
start_training_loss = state['current_train_loss']
start_eval_loss = state['current_eval_loss']
if not config.is_coverage:
self.optimizer.load_state_dict(state['optimizer'])
if use_cuda:
for state in self.optimizer.state.values():
for k, v in state.items():
print(k)
if isinstance(v, torch.Tensor):
state[k] = v.cuda()
self.chechpoint = Checkpoint(self.model,
self.optimizer,
self.model_dir,
start_eval_loss if start_eval_loss != 0 else float("inf"))
return start_iter, start_training_loss, start_eval_loss
def __init__(self, model_file_path, model_type="stem", load_batcher=True):
self.vocab = Vocab(config.vocab_path, config.vocab_size)
if load_batcher:
self.batcher = Batcher(config.decode_data_path,
self.vocab,
mode='decode',
batch_size=config.beam_size,
single_pass=True)
time.sleep(15)
self.model = Model(model_file_path, is_eval=True)
self.model_type = model_type
def __init__(self, model_file_path, is_word_level, is_combined, alpha):
self.vocab = Vocab(config.vocab_path, config.vocab_size)
# self.batcher = Batcher(config.eval_data_path, self.vocab, mode='eval',
# batch_size=config.batch_size, single_pass=True)
self.dataset = DailyMailDataset("val", self.vocab)
# time.sleep(15)
model_name = os.path.basename(model_file_path)
self.is_word_level = is_word_level
self.is_combined = is_combined
self.alpha = alpha
eval_dir = os.path.join(config.log_root, 'eval_%s' % (model_name))
if not os.path.exists(eval_dir):
os.mkdir(eval_dir)
self.model = Model(model_file_path, is_eval=True)
def predict(sentence, model_path):
if not os.path.exists(model_path):
raise Exception("Need to provide model path")
model = Model(model_path)
checkpoint = torch.load(model_path,
map_location=lambda storage, location: storage)
vocab = checkpoint['vocab']
target_field = Field(sequential=True,
init_token=START_DECODING,
eos_token=STOP_DECODING,
pad_token=PAD_TOKEN,
batch_first=True,
include_lengths=True,
unk_token=UNKNOWN_TOKEN,
lower=True)
source_field = Field(sequential=True,
init_token=SENTENCE_START,
eos_token=SENTENCE_END,
pad_token=PAD_TOKEN,
batch_first=True,
include_lengths=True,
unk_token=UNKNOWN_TOKEN,
lower=True)
source_field.vocab = vocab
target_field.vocab = vocab
data = [{'src': sentence, 'tgt': ''}]
predict_data = Mydataset(data=data,
fields=(('source', source_field), ('target',
target_field)))
setattr(args, 'vectors', source_field.vocab.vectors)
setattr(args, 'vocab_size', len(source_field.vocab.itos))
setattr(args, 'emb_dim', vectors.dim)
dest="is_word_level",
action="store_true")
parser.add_argument("--combined", dest="is_combined", action="store_true")
parser.set_defaults(is_word_level=False)
parser.set_defaults(is_combined=False)
args = parser.parse_args()
seq2seq_checkpoint_file = "./Seq2Seq_model_50000"
pg_losses = [
] #pickle.load(open("/home/lgpu0231/dumps_model_12_16_11_08/pg_losses_350.p", 'rb'))
run_avg_losses = [
] #pickle.load(open("/home/lgpu0231/dumps_model_12_16_11_08/run_avg_losses_350.p", 'rb'))
# Model
model = Model(seq2seq_checkpoint_file)
# model = Model()
# Load data
trainer = TrainSeq2Seq(is_word_level=args.is_word_level,
is_combined=args.is_combined)
# Prepare for training (e.g. optimizer)
iter, running_avg_loss = trainer.setup(model, model_file_path=None)
# GENERATOR MLE TRAINING - Pretrain
print('Starting Generator MLE Training...')
#trainer.train_nll(MLE_TRAIN_EPOCHS, iter, running_avg_loss)
# ADVERSARIAL TRAINING
print('\nStarting PG Training...')
trainer.train_pg(PG_TRAIN_EPOCHS, iter, running_avg_loss, pg_losses,
def train():
target_field = Field(sequential=True,
init_token=START_DECODING,
eos_token=STOP_DECODING,
pad_token=PAD_TOKEN,
batch_first=True,
include_lengths=True,
unk_token=UNKNOWN_TOKEN,
lower=True)
source_field = Field(sequential=True,
init_token=SENTENCE_START,
eos_token=SENTENCE_END,
pad_token=PAD_TOKEN,
batch_first=True,
include_lengths=True,
unk_token=UNKNOWN_TOKEN,
lower=True)
train_path = '../data/incar_alexa/train_public.pickle'
dev_path = '../data/incar_alexa/dev_public.pickle'
test_path = '../data/incar_alexa/test_public.pickle'
path = '../data/cnn_stories_tokenized'
summary_writer = SummaryWriter(config.summary_path)
train_src, train_tgt, train_id = load_data(train_path)
dev_src, dev_tgt, dev_id = load_data(dev_path)
test_src, test_tgt, test_id = load_data(test_path)
# train_data = prepare_data_cnn(path)
# # print(train_data[0])
# train_src = [dt['src'] for dt in train_data]
# train_tgt = [dt['tgt'] for dt in train_data]
# train_id = [dt['id'] for dt in train_data]
# train_src, test_src, train_tgt, test_tgt = train_test_split(
# train_src, train_tgt, test_size=0.15, random_state=123)
# train_id, test_id = train_test_split(
# train_id, test_size=0.15, random_state=123)
# # print(f"{len(train_src)}, {len(train_tgt)}")
# train_src, dev_src, train_tgt, dev_tgt = train_test_split(
# train_src, train_tgt, test_size=0.15, random_state=123)
# train_id, dev_id = train_test_split(
# train_id, test_size=0.15, random_state=123)
# print(source_field.preprocess(train_src[0]))
# exit()
train_src_preprocessed = [source_field.preprocess(x) for x in train_src]
dev_src_preprocessed = [source_field.preprocess(x) for x in dev_src]
test_src_preprocessed = [source_field.preprocess(x) for x in test_src]
train_tgt_preprocessed = [target_field.preprocess(x) for x in train_tgt]
dev_tgt_preprocessed = [target_field.preprocess(x) for x in dev_tgt]
test_tgt_preprocessed = [target_field.preprocess(x) for x in test_tgt]
# train_src_preprocessed = source_field.apply(lambda x: source_field.preprocess(x))
vectors = Vectors(
name='/home/binhna/Downloads/shared_resources/cc.en.300.vec',
cache='/home/binhna/Downloads/shared_resources/')
source_field.build_vocab([
train_src_preprocessed, dev_src_preprocessed, train_tgt_preprocessed,
dev_tgt_preprocessed
],
vectors=vectors)
target_field.build_vocab([
train_src_preprocessed, dev_src_preprocessed, train_tgt_preprocessed,
dev_tgt_preprocessed
],
vectors=vectors)
train_data = [{
'src': src,
'tgt': tgt,
'id': id
} for src, tgt, id in zip(train_src, train_tgt, train_id)]
train_data = Mydataset(data=train_data,
fields=(('source', source_field), ('target',
target_field)))
dev_data = [{
'src': src,
'tgt': tgt,
'id': id
} for src, tgt, id in zip(dev_src, dev_tgt, dev_id)]
# print(dev_data[0])
dev_data = Mydataset(data=dev_data,
fields=(('source', source_field), ('target',
target_field)))
test_data = [{
'src': src,
'tgt': tgt,
'id': id
} for src, tgt, id in zip(test_src, test_tgt, test_id)]
test_data = Mydataset(data=test_data,
fields=(('source', source_field), ('target',
target_field)))
# print(train_data[10].source)
# print(train_data[10].target)
# print(len(target_field.vocab))
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
train_iter, test_iter, dev_iter = BucketIterator.splits(
datasets=(train_data, test_data, dev_data),
batch_sizes=(config.batch_size, config.batch_size, config.batch_size),
device=device,
sort_key=lambda x: len(x.source),
sort_within_batch=True)
args = ARGS()
setattr(args, 'vectors', source_field.vocab.vectors)
setattr(args, 'vocab_size', len(source_field.vocab.itos))
setattr(args, 'emb_dim', vectors.dim)
model = Model(args)
params = list(model.encoder.parameters()) + list(
model.decoder.parameters()) + list(model.reduce_state.parameters())
initial_lr = config.lr_coverage if config.is_coverage else config.lr
optimizer = Adagrad(params,
lr=initial_lr,
initial_accumulator_value=config.adagrad_init_acc)
iter, running_avg_loss = 0, 0
start = time.time()
for epoch in range(500):
print(f"Epoch: {epoch+1}")
for i, batch in tqdm(enumerate(train_iter), total=len(train_iter)):
# print(batch.source[0].size())
# exit()
batch_size = batch.batch_size
# encoder part
enc_padding_mask = get_mask(batch.source, device)
enc_batch = batch.source[0]
enc_lens = batch.source[1]
encoder_outputs, encoder_feature, encoder_hidden = model.encoder(
enc_batch, enc_lens)
s_t_1 = model.reduce_state(encoder_hidden)
coverage = Variable(torch.zeros(batch.source[0].size())).to(device)
c_t_1 = Variable(torch.zeros(
(batch_size, 2 * config.hidden_dim))).to(device)
extra_zeros, enc_batch_extend_vocab, max_art_oovs = get_extra_features(
batch.source[0], source_field.vocab)
extra_zeros = extra_zeros.to(device)
enc_batch_extend_vocab = enc_batch_extend_vocab.to(device)
# decoder part
dec_batch = batch.target[0][:, :-1]
# print(dec_batch.size())
target_batch = batch.target[0][:, 0:]
dec_lens_var = batch.target[1]
dec_padding_mask = get_mask(batch.target, device)
max_dec_len = max(dec_lens_var)
step_losses = []
for di in range(min(max_dec_len, config.max_dec_steps) - 1):
y_t_1 = dec_batch[:, di] # Teacher forcing
final_dist, s_t_1, c_t_1, attn_dist, p_gen, next_coverage = model.decoder(
y_t_1, s_t_1, encoder_outputs, encoder_feature,
enc_padding_mask, c_t_1, extra_zeros,
enc_batch_extend_vocab, coverage, di)
target = target_batch[:, di]
gold_probs = torch.gather(final_dist, 1,
target.unsqueeze(1)).squeeze()
step_loss = -torch.log(gold_probs + config.eps)
if config.is_coverage:
step_coverage_loss = torch.sum(
torch.min(attn_dist, coverage), 1)
step_loss = step_loss + config.cov_loss_wt * step_coverage_loss
coverage = next_coverage
step_mask = dec_padding_mask[:, di]
step_loss = step_loss * step_mask
step_losses.append(step_loss)
sum_losses = torch.sum(torch.stack(step_losses, 1), 1)
batch_avg_loss = sum_losses / dec_lens_var
loss = torch.mean(batch_avg_loss)
loss.backward()
norm = clip_grad_norm_(model.encoder.parameters(),
config.max_grad_norm)
clip_grad_norm_(model.decoder.parameters(), config.max_grad_norm)
clip_grad_norm_(model.reduce_state.parameters(),
config.max_grad_norm)
optimizer.step()
running_avg_loss = calc_running_avg_loss(loss.item(),
running_avg_loss,
summary_writer, iter)
iter += 1
summary_writer.flush()
# print_interval = 10
# if iter % print_interval == 0:
# print(f'steps {iter}, batch number: {i} with {time.time() - start} seconds, loss: {loss}')
# start = time.time()
if iter % 300 == 0:
save_model(model, optimizer, running_avg_loss, iter,
config.model_dir)
dev_data = Mydataset(data=dev_data,
fields=(('source', source_field), ('target',
target_field)))
test_data = [{
'src': src,
'tgt': tgt,
'id': id
} for src, tgt, id in zip(test_src, test_tgt, test_id)]
test_data = Mydataset(data=test_data,
fields=(('source', source_field), ('target',
target_field)))
setattr(args, 'vectors', source_field.vocab.vectors)
setattr(args, 'vocab_size', len(source_field.vocab.itos))
setattr(args, 'emb_dim', vectors.dim)
model = Model(args)
trainer = Trainer(model=model,
args=args,
train_dataset=train_data,
eval_dataset=dev_data,
test_dataset=test_data,
vocab=source_field.vocab,
is_train=True)
trainer.train()
# for name in ['train', 'dev', 'test']:
# process_incar_data(f'../data/incar_alexa/{name}_public.json')
# vocabs = read_vocabs('../data/finished_files/vocab')
# print(len(vocabs))