for epoch in tqdm(range(model_config.epochs), desc='epochs'):
tr_loss = 0
tr_acc = 0
model.train()
for step, mb in tqdm(enumerate(tr_dl), desc='steps', total=len(tr_dl)):
x_mb, y_mb = map(lambda elm: elm.to(device), mb)
opt.zero_grad()
y_hat_mb = model(x_mb)
mb_loss = loss_fn(y_hat_mb, y_mb)
mb_loss.backward()
opt.step()
with torch.no_grad():
mb_acc = acc(y_hat_mb, y_mb)
tr_loss += mb_loss.item()
tr_acc += mb_acc.item()
if (epoch * len(tr_dl) + step) % model_config.summary_step == 0:
val_loss = evaluate(model, val_dl, {'loss': loss_fn}, device)['loss']
writer.add_scalars('loss', {'train': tr_loss / (step + 1),
'val': val_loss}, epoch * len(tr_dl) + step)
tqdm.write('global_step: {:3}, tr_loss: {:.3f}, val_loss: {:.3f}'.format(epoch * len(tr_dl) + step,
tr_loss / (step + 1),
val_loss))
model.train()
else:
tr_loss /= (step + 1)
tr_acc /= (step + 1)
def main(args):
dataset_config = Config(args.dataset_config)
model_config = Config(args.model_config)
exp_dir = Path("experiments") / model_config.type
exp_dir = exp_dir.joinpath(
f"epochs_{args.epochs}_batch_size_{args.batch_size}_learning_rate_{args.learning_rate}"
)
if not exp_dir.exists():
exp_dir.mkdir(parents=True)
if args.fix_seed:
torch.manual_seed(777)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
tokenizer = get_tokenizer(dataset_config)
tr_dl, val_dl = get_data_loaders(dataset_config, model_config, tokenizer,
args.batch_size)
# model
model = ConvRec(num_classes=model_config.num_classes,
embedding_dim=model_config.embedding_dim,
hidden_dim=model_config.hidden_dim,
vocab=tokenizer.vocab)
loss_fn = nn.CrossEntropyLoss()
opt = optim.Adam(params=model.parameters(), lr=args.learning_rate)
scheduler = ReduceLROnPlateau(opt, patience=5)
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
model.to(device)
writer = SummaryWriter('{}/runs'.format(exp_dir))
checkpoint_manager = CheckpointManager(exp_dir)
summary_manager = SummaryManager(exp_dir)
best_val_loss = 1e+10
for epoch in tqdm(range(args.epochs), desc='epochs'):
tr_loss = 0
tr_acc = 0
model.train()
for step, mb in tqdm(enumerate(tr_dl), desc='steps', total=len(tr_dl)):
x_mb, y_mb = map(lambda elm: elm.to(device), mb)
opt.zero_grad()
y_hat_mb = model(x_mb)
mb_loss = loss_fn(y_hat_mb, y_mb)
mb_loss.backward()
opt.step()
with torch.no_grad():
mb_acc = acc(y_hat_mb, y_mb)
tr_loss += mb_loss.item()
tr_acc += mb_acc.item()
if (epoch * len(tr_dl) + step) % args.summary_step == 0:
val_loss = evaluate(model, val_dl, {'loss': loss_fn},
device)['loss']
writer.add_scalars('loss', {
'train': tr_loss / (step + 1),
'val': val_loss
},
epoch * len(tr_dl) + step)
model.train()
else:
tr_loss /= (step + 1)
tr_acc /= (step + 1)
tr_summary = {'loss': tr_loss, 'acc': tr_acc}
val_summary = evaluate(model, val_dl, {
'loss': loss_fn,
'acc': acc
}, device)
scheduler.step(val_summary['loss'])
tqdm.write('epoch : {}, tr_loss: {:.3f}, val_loss: '
'{:.3f}, tr_acc: {:.2%}, val_acc: {:.2%}'.format(
epoch + 1, tr_summary['loss'], val_summary['loss'],
tr_summary['acc'], val_summary['acc']))
val_loss = val_summary['loss']
is_best = val_loss < best_val_loss
if is_best:
state = {
'epoch': epoch + 1,
'model_state_dict': model.state_dict(),
'opt_state_dict': opt.state_dict()
}
summary = {'train': tr_summary, 'validation': val_summary}
summary_manager.update(summary)
summary_manager.save('summary.json')
checkpoint_manager.save_checkpoint(state, 'best.tar')
best_val_loss = val_loss
def main(args):
dataset_config = Config(args.dataset_config)
model_config = Config(args.model_config)
exp_dir = Path("experiments") / model_config.type
exp_dir = exp_dir.joinpath(
f"epochs_{args.epochs}_batch_size_{args.batch_size}_learning_rate_{args.learning_rate}"
)
if not exp_dir.exists():
exp_dir.mkdir(parents=True)
if args.fix_seed:
torch.manual_seed(777)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
tokenizer = get_tokenizer(dataset_config, model_config)
tr_dl, val_dl = get_data_loaders(dataset_config, tokenizer, args.batch_size)
model = SenCNN(num_classes=model_config.num_classes, vocab=tokenizer.vocab)
loss_fn = nn.CrossEntropyLoss()
opt = optim.Adam(params=model.parameters(), lr=args.learning_rate)
scheduler = ReduceLROnPlateau(opt, patience=5)
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
model.to(device)
writer = SummaryWriter(f"{exp_dir}/runs")
checkpoint_manager = CheckpointManager(exp_dir)
summary_manager = SummaryManager(exp_dir)
best_val_loss = 1e10
for epoch in tqdm(range(args.epochs), desc="epochs"):
tr_loss = 0
tr_acc = 0
model.train()
for step, mb in tqdm(enumerate(tr_dl), desc="steps", total=len(tr_dl)):
x_mb, y_mb = map(lambda elm: elm.to(device), mb)
opt.zero_grad()
y_hat_mb = model(x_mb)
mb_loss = loss_fn(y_hat_mb, y_mb)
mb_loss.backward()
clip_grad_norm_(model._fc.weight, 5)
opt.step()
with torch.no_grad():
mb_acc = acc(y_hat_mb, y_mb)
tr_loss += mb_loss.item()
tr_acc += mb_acc.item()
if (epoch * len(tr_dl) + step) % args.summary_step == 0:
val_loss = evaluate(model, val_dl, {"loss": loss_fn}, device)["loss"]
writer.add_scalars("loss", {"train": tr_loss / (step + 1), "validation": val_loss},
epoch * len(tr_dl) + step)
model.train()
else:
tr_loss /= step + 1
tr_acc /= step + 1
tr_summary = {"loss": tr_loss, "acc": tr_acc}
val_summary = evaluate(model, val_dl, {"loss": loss_fn, "acc": acc}, device)
scheduler.step(val_summary["loss"])
tqdm.write(f"epoch: {epoch+1}\n"
f"tr_loss: {tr_summary['loss']:.3f}, val_loss: {val_summary['loss']:.3f}\n"
f"tr_acc: {tr_summary['acc']:.2%}, val_acc: {val_summary['acc']:.2%}")
val_loss = val_summary["loss"]
is_best = val_loss < best_val_loss
if is_best:
state = {
"epoch": epoch + 1,
"model_state_dict": model.state_dict(),
"opt_state_dict": opt.state_dict(),
}
summary = {
"epoch": epoch + 1,
"train": tr_summary,
"validation": val_summary,
}
summary_manager.update(summary)
summary_manager.save("summary.json")
checkpoint_manager.save_checkpoint(state, "best.tar")
best_val_loss = val_loss
tr_acc = 0
model.train()
for step, mb in tqdm(enumerate(tr_dl), desc='steps', total=len(tr_dl)):
x_mb, y_mb = map(lambda elm: elm.to(device), mb)
opt.zero_grad()
score, attn_mat = model(x_mb)
reg = regularize(attn_mat, model_config.r, device)
mb_loss = loss_fn(score, y_mb)
mb_loss.add_(reg)
mb_loss.backward()
opt.step()
with torch.no_grad():
mb_acc = acc(score, y_mb)
tr_loss += mb_loss.item()
tr_acc += mb_acc.item()
if (epoch * len(tr_dl) + step) % model_config.summary_step == 0:
val_loss = evaluate(model, val_dl, {'loss': loss_fn},
device)['loss']
writer.add_scalars('loss', {
'train': tr_loss / (step + 1),
'val': val_loss
},
epoch * len(tr_dl) + step)
model.train()
else:
tr_loss /= (step + 1)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--epoch', default=10, type=int)
parser.add_argument('--batch_size', default=128, type=int)
# parser.add_argument('--data_type', default='senCNN')
parser.add_argument('--classes', default=2, type=int)
parser.add_argument('--gpu', default=0, type=int)
parser.add_argument('--learning_rate', default=1e-3, type=float)
# parser.add_argument('--print_freq', default=3000, type=int)
# parser.add_argument('--weight_decay', default=5e-5, type=float)
parser.add_argument('--word_dim', default=16, type=int)
parser.add_argument('--word_max_len', default=300, type=int)
parser.add_argument('--global_step', default=1000, type=int)
parser.add_argument('--data_path', default='../data_in')
parser.add_argument('--file_path', default='../nsmc-master')
# parser.add_argument('--build_preprocessing', default=False)
# parser.add_argument('--build_vocab', default=False)
args = parser.parse_args()
# p = Preprocessing(args)
# p.makeProcessing()
# v = Build_Vocab(args)
# v.make_vocab()
with open(args.data_path + '/' + 'vocab_char.pkl', mode='rb') as io:
vocab = pickle.load(io)
padder = PadSequence(length=args.word_max_len,
pad_val=vocab.to_indices(vocab.padding_token))
tokenizer = Tokenizer(vocab=vocab, split_fn=split_to_jamo, pad_fn=padder)
model = EfficientCharCRNN(args, vocab)
epochs = args.epoch
batch_size = args.batch_size
learning_rate = args.learning_rate
global_step = args.global_step
tr_ds = Corpus(args.data_path + '/train.txt',
tokenizer.split_and_transform)
tr_dl = DataLoader(tr_ds,
batch_size=batch_size,
shuffle=True,
num_workers=4,
drop_last=True)
val_ds = Corpus(args.data_path + '/val.txt', tokenizer.split_and_transform)
val_dl = DataLoader(val_ds, batch_size=batch_size, num_workers=4)
loss_fn = nn.CrossEntropyLoss()
opt = optim.Adam(params=model.parameters(), lr=learning_rate)
scheduler = ReduceLROnPlateau(opt, patience=5)
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
model.to(device)
best_val_loss = 1e+10
for epoch in tqdm(range(args.epoch), desc='epochs'):
tr_loss = 0
tr_acc = 0
model.train()
for step, mb in tqdm(enumerate(tr_dl), desc='steps', total=len(tr_dl)):
x, y = map(lambda elm: elm.to(device), mb)
opt.zero_grad()
y_h = model(x)
m_loss = loss_fn(y_h, y)
m_loss.backward()
clip_grad_norm_(model._fc.weight, 5)
opt.step()
with torch.no_grad():
m_acc = acc(y_h, y)
tr_loss += m_loss.item()
tr_acc += m_acc.item()
else:
tr_loss /= (step + 1)
tr_acc /= (step + 1)
tr_summ = {'loss': tr_loss, 'acc': tr_acc}
val_summ = evaluate(model, val_dl, {
'loss': loss_fn,
'acc': acc
}, device)
scheduler.step(val_summ['loss'])
tqdm.write('epoch : {}, tr_loss: {:.3f}, val_loss: '
'{:.3f}, tr_acc: {:.2%}, val_acc: {:.2%}'.format(
epoch + 1, tr_summ['loss'], val_summ['loss'],
tr_summ['acc'], val_summ['acc']))
val_loss = val_summ['loss']
is_best = val_loss < best_val_loss
if is_best:
state = {
'epoch': epoch + 1,
'model_state_dict': model.state_dict(),
'opt_state_dict': opt.state_dict()
}
summary = {'tr': tr_summ, 'val': val_summ}
# manager.update_summary(summary)
# manager.save_summary('summary.json')
# manager.save_checkpoint(state, 'best.tar')
best_val_loss = val_loss
def main(args):
dataset_config = Config(args.dataset_config)
model_config = Config(args.model_config)
exp_dir = Path("experiments") / model_config.type
exp_dir = exp_dir.joinpath(
f"epochs_{args.epochs}_batch_size_{args.batch_size}_learning_rate_{args.learning_rate}"
)
if not exp_dir.exists():
exp_dir.mkdir(parents=True)
if args.fix_seed:
torch.manual_seed(777)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
preprocessor = get_preprocessor(dataset_config,
coarse_split_fn=split_morphs,
fine_split_fn=split_jamos)
tr_dl, val_dl = get_data_loaders(dataset_config,
preprocessor,
args.batch_size,
collate_fn=batchify)
# model
model = SAN(model_config.num_classes, preprocessor.coarse_vocab,
preprocessor.fine_vocab, model_config.fine_embedding_dim,
model_config.hidden_dim, model_config.multi_step,
model_config.prediction_drop_ratio)
opt = optim.Adam(model.parameters(), lr=args.learning_rate)
device = torch.device(
"cuda") if torch.cuda.is_available() else torch.device("cpu")
model.to(device)
writer = SummaryWriter(f"{exp_dir}/runs")
checkpoint_manager = CheckpointManager(exp_dir)
summary_manager = SummaryManager(exp_dir)
best_val_loss = 1e10
for epoch in tqdm(range(args.epochs), desc="epochs"):
tr_loss = 0
tr_acc = 0
model.train()
for step, mb in tqdm(enumerate(tr_dl), desc="steps", total=len(tr_dl)):
qa_mb, qb_mb, y_mb = map(
lambda elm: (el.to(device) for el in elm)
if isinstance(elm, tuple) else elm.to(device), mb)
opt.zero_grad()
y_hat_mb = model((qa_mb, qb_mb))
mb_loss = log_loss(y_hat_mb, y_mb)
mb_loss.backward()
opt.step()
with torch.no_grad():
mb_acc = acc(y_hat_mb, y_mb)
tr_loss += mb_loss.item()
tr_acc += mb_acc.item()
if (epoch * len(tr_dl) + step) % args.summary_step == 0:
val_loss = evaluate(model, val_dl, {"loss": log_loss},
device)["loss"]
writer.add_scalars("loss", {
"train": tr_loss / (step + 1),
"val": val_loss
},
epoch * len(tr_dl) + step)
model.train()
else:
tr_loss /= step + 1
tr_acc /= step + 1
tr_summary = {"loss": tr_loss, "acc": tr_acc}
val_summary = evaluate(model, val_dl, {
"loss": log_loss,
"acc": acc
}, device)
tqdm.write(
f"epoch: {epoch+1}\n"
f"tr_loss: {tr_summary['loss']:.3f}, val_loss: {val_summary['loss']:.3f}\n"
f"tr_acc: {tr_summary['acc']:.2%}, val_acc: {val_summary['acc']:.2%}"
)
val_loss = val_summary["loss"]
is_best = val_loss < best_val_loss
if is_best:
state = {
"epoch": epoch + 1,
"model_state_dict": model.state_dict(),
"opt_state_dict": opt.state_dict(),
}
summary = {"train": tr_summary, "validation": val_summary}
summary_manager.update(summary)
summary_manager.save("summary.json")
checkpoint_manager.save_checkpoint(state, "best.tar")
best_val_loss = val_loss
def main(args):
dataset_config = Config(args.dataset_config)
model_config = Config(args.model_config)
ptr_config_info = Config(f"conf/pretrained/{model_config.type}.json")
exp_dir = Path("experiments") / model_config.type
exp_dir = exp_dir.joinpath(
f"epochs_{args.epochs}_batch_size_{args.batch_size}_learning_rate_{args.learning_rate}"
f"_weight_decay_{args.weight_decay}")
if not exp_dir.exists():
exp_dir.mkdir(parents=True)
if args.fix_seed:
torch.manual_seed(777)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
preprocessor = get_preprocessor(ptr_config_info, model_config)
with open(ptr_config_info.config, mode="r") as io:
ptr_config = json.load(io)
# model
config = BertConfig()
config.update(ptr_config)
model = PairwiseClassifier(config,
num_classes=model_config.num_classes,
vocab=preprocessor.vocab)
bert_pretrained = torch.load(ptr_config_info.bert)
model.load_state_dict(bert_pretrained, strict=False)
tr_dl, val_dl = get_data_loaders(dataset_config, preprocessor,
args.batch_size)
loss_fn = nn.CrossEntropyLoss()
opt = optim.Adam([
{
"params": model.bert.parameters(),
"lr": args.learning_rate / 100
},
{
"params": model.classifier.parameters(),
"lr": args.learning_rate
},
],
weight_decay=args.weight_decay)
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
model.to(device)
writer = SummaryWriter(f'{exp_dir}/runs')
checkpoint_manager = CheckpointManager(exp_dir)
summary_manager = SummaryManager(exp_dir)
best_val_loss = 1e+10
for epoch in tqdm(range(args.epochs), desc='epochs'):
tr_loss = 0
tr_acc = 0
model.train()
for step, mb in tqdm(enumerate(tr_dl), desc='steps', total=len(tr_dl)):
x_mb, x_types_mb, y_mb = map(lambda elm: elm.to(device), mb)
opt.zero_grad()
y_hat_mb = model(x_mb, x_types_mb)
mb_loss = loss_fn(y_hat_mb, y_mb)
mb_loss.backward()
opt.step()
with torch.no_grad():
mb_acc = acc(y_hat_mb, y_mb)
tr_loss += mb_loss.item()
tr_acc += mb_acc.item()
if (epoch * len(tr_dl) + step) % args.summary_step == 0:
val_loss = evaluate(model, val_dl, {'loss': loss_fn},
device)['loss']
writer.add_scalars('loss', {
'train': tr_loss / (step + 1),
'val': val_loss
},
epoch * len(tr_dl) + step)
model.train()
else:
tr_loss /= (step + 1)
tr_acc /= (step + 1)
tr_summary = {'loss': tr_loss, 'acc': tr_acc}
val_summary = evaluate(model, val_dl, {
'loss': loss_fn,
'acc': acc
}, device)
tqdm.write(
f"epoch: {epoch+1}\n"
f"tr_loss: {tr_summary['loss']:.3f}, val_loss: {val_summary['loss']:.3f}\n"
f"tr_acc: {tr_summary['acc']:.2%}, val_acc: {val_summary['acc']:.2%}"
)
val_loss = val_summary['loss']
is_best = val_loss < best_val_loss
if is_best:
state = {
'epoch': epoch + 1,
'model_state_dict': model.state_dict(),
'opt_state_dict': opt.state_dict()
}
summary = {'train': tr_summary, 'validation': val_summary}
summary_manager.update(summary)
summary_manager.save('summary.json')
checkpoint_manager.save_checkpoint(state, 'best.tar')
best_val_loss = val_loss
