def main():
logging.info("Build dataset...")
prob_and_idx = get_prob_idx(opt.strategy)
train_dataset = build_dataset(opt,
opt.train,
opt.vocab,
device,
prob_and_idx,
train=True)
valid_dataset = build_dataset(opt,
opt.valid,
opt.vocab,
device,
prob_and_idx,
train=False)
fields = valid_dataset.fields = train_dataset.fields
logging.info("Build model...")
pad_ids = {"src": fields["src"].pad_id, "tgt": fields["tgt"].pad_id}
vocab_sizes = {
"src": len(fields["src"].vocab),
"tgt": len(fields["tgt"].vocab)
}
model = NMTModel.load_model(opt, pad_ids, vocab_sizes).to(device)
criterion = LabelSmoothingLoss(opt.label_smoothing, vocab_sizes["tgt"],
pad_ids["tgt"]).to(device)
n_step = int(opt.train_from.split("-")[-1]) if opt.train_from else 1
optimizer = WarmAdam(model.parameters(), opt.lr, opt.hidden_size,
opt.warm_up, n_step)
logging.info("start training...")
train(model, criterion, optimizer, train_dataset, valid_dataset)
def main():
logging.info("Build dataset...")
valid_dataset = build_dataset(opt, opt.valid, opt.vocab, device, train=False)
logging.info("Built dataset valid ...")
train_dataset = build_dataset(opt, opt.train, opt.vocab, device, train=True)
logging.info("Built dataset train ...")
fields = valid_dataset.fields = train_dataset.fields
logging.info("Build model...")
pad_ids = {"source": fields["source"].pad_id,
"summary_cn": fields["summary_cn"].pad_id,
"summary_en": fields["summary_en"].pad_id}
vocab_sizes = {"source": len(fields["source"].vocab),
"summary_cn": len(fields["summary_cn"].vocab),
"summary_en": len(fields["summary_en"].vocab)}
print(vocab_sizes)
model = NMTModel.load_model(opt, pad_ids, vocab_sizes).to(device)
criterion_cn = LabelSmoothingLoss(opt.label_smoothing, vocab_sizes["summary_cn"], pad_ids["summary_cn"]).to(device)
criterion_en = LabelSmoothingLoss(opt.label_smoothing, vocab_sizes["summary_en"], pad_ids["summary_en"]).to(device)
n_step = int(opt.train_from.split("-")[-1]) if opt.train_from else 1
optimizer = WarmAdam(model.parameters(), opt.lr, opt.hidden_size, opt.warm_up, n_step)
logging.info("start training...")
train(model, pad_ids, vocab_sizes, criterion_cn, criterion_en, optimizer, train_dataset, valid_dataset)
def main():
sim_model = load_sim_model(opt.sim_model_file)
logging.info("Build dataset...")
train_dataset = build_dataset(opt, opt.train, opt.vocab, device, train=True)
valid_dataset = build_dataset(opt, opt.valid, opt.vocab, device, train=False)
fields = valid_dataset.fields = train_dataset.fields
logging.info("Build model...")
pad_ids = {"src": fields["src"].pad_id,
"task1_tgt": fields["task1_tgt"].pad_id,
"task2_tgt": fields["task2_tgt"].pad_id}
vocab_sizes = {"src": len(fields["src"].vocab),
"task1_tgt": len(fields["task1_tgt"].vocab),
"task2_tgt": len(fields["task2_tgt"].vocab)}
model = NMTModel.load_model(opt, pad_ids, vocab_sizes).to(device)
criterion_task1 = MyLabelSmoothingLoss(opt.label_smoothing, vocab_sizes["task1_tgt"], pad_ids["task1_tgt"]).to(device)
criterion_task2 = MyLabelSmoothingLoss(opt.label_smoothing, vocab_sizes["task2_tgt"], pad_ids["task2_tgt"]).to(device)
n_step = 1
optimizer = WarmAdam(model.parameters(), opt.lr, opt.hidden_size, opt.warm_up, n_step)
logging.info("start training...")
train(sim_model, model, criterion_task1, criterion_task2, optimizer, train_dataset, valid_dataset, fields, opt.alpha, opt.beta)
def main():
logging.info("Build dataset...")
train_dataset = build_dataset(opt,
opt.train,
opt.vocab,
device,
train=True)
valid_dataset = build_dataset(opt,
opt.valid,
opt.vocab,
device,
train=False)
fields = valid_dataset.fields = train_dataset.fields
logging.info("Build model...")
pad_ids = {
"src": fields["src"].pad_id,
"task1_tgt": fields["task1_tgt"].pad_id,
"task2_tgt": fields["task2_tgt"].pad_id,
"task3_tgt": fields["task3_tgt"].pad_id
}
vocab_sizes = {
"src": len(fields["src"].vocab),
"task1_tgt": len(fields["task1_tgt"].vocab),
"task2_tgt": len(fields["task2_tgt"].vocab),
"task3_tgt": len(fields["task3_tgt"].vocab)
}
print(vocab_sizes)
model = NMTModel.load_model(opt, pad_ids, vocab_sizes).to(device)
# if torch.cuda.device_count() > 1:
# print("Let's use", torch.cuda.device_count(), "GPUs!")
# model = nn.DataParallel(model, device_ids=[0,1])
# for MT
criterion_task1 = LabelSmoothingLoss(opt.label_smoothing,
vocab_sizes["task1_tgt"],
pad_ids["task1_tgt"]).to(device)
# for MS
criterion_task2 = LabelSmoothingLoss(opt.label_smoothing,
vocab_sizes["task2_tgt"],
pad_ids["task2_tgt"]).to(device)
# for CLS
criterion_task3 = LabelSmoothingLoss(opt.label_smoothing,
vocab_sizes["task3_tgt"],
pad_ids["task3_tgt"]).to(device)
n_step = int(opt.train_from.split("-")[-1]) if opt.train_from else 1
optimizer = WarmAdam(model.parameters(), opt.lr, opt.hidden_size,
opt.warm_up, n_step)
logging.info("start training...")
train(model, pad_ids, vocab_sizes, criterion_task1, criterion_task2,
criterion_task3, optimizer, train_dataset, valid_dataset)
def main():
logger.info("Build dataset...")
train_dataset = build_dataset(opt, opt.train, opt.vocab, device, train=True)
valid_dataset = build_dataset(opt, opt.valid, opt.vocab, device, train=False)
fields = valid_dataset.fields = train_dataset.fields
logger.info("Build model...")
model = NMTModel.load_model(loader, fields)
criterion = LabelSmoothingLoss(opt.label_smoothing, len(fields["tgt"].vocab), fields["tgt"].pad_id)
model = nn.DataParallel(FullModel(model, criterion)).to(device)
optimizer = WarmAdam(model.module.model.parameters(), opt.lr, opt.hidden_size, opt.warm_up, loader.step)
logger.info("start training...")
train(model, optimizer, train_dataset, valid_dataset)
def main():
logging.info("Build dataset...")
prob_and_idx = get_prob_idx()
dataset = build_dataset(opt, [opt.input, opt.truth or opt.input],
opt.vocab,
device,
prob_and_idx,
train=False)
fields = dataset.fields
pad_ids = {"src": fields["src"].pad_id, "tgt": fields["tgt"].pad_id}
vocab_sizes = {
"src": len(fields["src"].vocab),
"tgt": len(fields["tgt"].vocab)
}
# load checkpoint from model_path
logging.info("Load checkpoint from %s." % opt.model_path)
checkpoint = torch.load(opt.model_path,
map_location=lambda storage, loc: storage)
logging.info("Build model...")
model = NMTModel.load_model(checkpoint["opt"], pad_ids, vocab_sizes,
checkpoint["model"]).to(device).eval()
logging.info("Start translation...")
with torch.set_grad_enabled(False):
translate(dataset, fields, model)
def main():
logger.info("Build dataset...")
dataset = build_dataset(opt, [opt.input, opt.input], opt.vocab, device, train=False)
logger.info("Build model...")
model = NMTModel.load_model(loader, dataset.fields).to(device).eval()
logger.info("Start translation...")
with torch.set_grad_enabled(False):
translate(dataset, dataset.fields, model)
def main():
logging.info("Build dataset...")
valid_dataset = build_dataset(opt,
opt.valid,
opt.vocab,
device,
train=False)
fields = valid_dataset.fields
pad_ids = {
"source": fields["source"].pad_id,
"summary_cn": fields["summary_cn"].pad_id,
"summary_en": fields["summary_en"].pad_id
}
vocab_sizes = {
"source": len(fields["source"].vocab),
"summary_cn": len(fields["summary_cn"].vocab),
"summary_en": len(fields["summary_en"].vocab)
}
# load checkpoint from model_path
for ckpt in range(1100000, 1345001, 5000):
model_path = opt.model_path + '/checkpoint-step-' + str(ckpt)
logging.info("Load checkpoint from %s." % model_path)
checkpoint = torch.load(model_path,
map_location=lambda storage, loc: storage)
logging.info("Build model...")
model = NMTModel.load_model(checkpoint["opt"], pad_ids, vocab_sizes,
checkpoint["model"]).to(device).eval()
criterion_cn = LabelSmoothingLoss(opt.label_smoothing,
vocab_sizes["summary_cn"],
pad_ids["summary_cn"]).to(device)
criterion_en = LabelSmoothingLoss(opt.label_smoothing,
vocab_sizes["summary_en"],
pad_ids["summary_en"]).to(device)
n_step = ckpt
logging.info("Start translation...")
with torch.set_grad_enabled(False):
valid(model, criterion_cn, criterion_en, valid_dataset, n_step)
encoder += torch.numel(tensor)
if 'cn_decoder' in layer_tensor_name:
task1 += torch.numel(tensor)
if 'en_decoder' in layer_tensor_name:
task2 += torch.numel(tensor)
print('encoder: {}, task1: {}, task2: {}, Total parameters: {}'.format(
encoder, task1, task2, total))
pytorch_total_params = sum(p.numel() for p in model.parameters()
if p.requires_grad)
print(pytorch_total_params)
if __name__ == '__main__':
train_dataset = build_dataset(opt,
opt.train,
opt.vocab,
device,
train=True)
fields = train_dataset.fields
logging.info("Build model...")
pad_ids = {
"source": fields["source"].pad_id,
"summary_cn": fields["summary_cn"].pad_id,
"summary_en": fields["summary_en"].pad_id
}
vocab_sizes = {
"source": len(fields["source"].vocab),
"summary_cn": len(fields["summary_cn"].vocab),
"summary_en": len(fields["summary_en"].vocab)
}
