subchar=args.subchar)
# model (restore)
checkpoint_manager = CheckpointManager(model_dir)
checkpoint = checkpoint_manager.load_checkpoint('best_snu_{}.tar'.format(
args.pretrained_config))
config = BertConfig(ptr_config.config)
model = SentenceClassifier(config,
num_classes=model_config.num_classes,
vocab=preprocessor.vocab)
model.load_state_dict(checkpoint['model_state_dict'])
# evaluation
filepath = getattr(data_config, args.dataset)
ds = Corpus(filepath, preprocessor.preprocess)
dl = DataLoader(ds, batch_size=model_config.batch_size, num_workers=4)
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
model.to(device)
summary_manager = SummaryManager(model_dir)
summary = evaluate(model, dl, {
'loss': nn.CrossEntropyLoss(),
'acc': acc
}, device)
summary_manager.load('summary_snu_{}.json'.format(args.pretrained_config))
summary_manager.update({'{}'.format(args.dataset): summary})
summary_manager.save('summary_snu_{}.json'.format(args.pretrained_config))
def train(cfgpath):
# parsing json
with open(os.path.join(os.getcwd(), cfgpath)) as io:
params = json.loads(io.read())
# creating preprocessor
tokenizer = JamoTokenizer()
padder = PadSequence(300)
# creating model
model = CharCNN(num_classes=params['model'].get('num_classes'),
embedding_dim=params['model'].get('embedding_dim'),
dic=tokenizer.token2idx)
# creating dataset, dataloader
tr_filepath = os.path.join(os.getcwd(), params['filepath'].get('tr'))
val_filepath = os.path.join(os.getcwd(), params['filepath'].get('val'))
batch_size = params['training'].get('batch_size')
tr_ds = Corpus(tr_filepath, tokenizer, padder)
tr_dl = DataLoader(tr_ds,
batch_size=batch_size,
shuffle=True,
num_workers=4,
drop_last=True)
val_ds = Corpus(val_filepath, tokenizer, padder)
val_dl = DataLoader(val_ds, batch_size=batch_size, num_workers=4)
# training
loss_fn = nn.CrossEntropyLoss()
opt = optim.Adam(params=model.parameters(),
lr=params['training'].get('learning_rate'))
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
model.to(device)
epochs = params['training'].get('epochs')
for epoch in tqdm(range(epochs), desc='epochs'):
avg_tr_loss = 0
avg_val_loss = 0
tr_step = 0
val_step = 0
model.train()
for x_mb, y_mb in tqdm(tr_dl, desc='iters'):
x_mb = x_mb.to(device)
y_mb = y_mb.to(device)
score = model(x_mb)
opt.zero_grad()
tr_loss = loss_fn(score, y_mb)
tr_loss.backward()
opt.step()
avg_tr_loss += tr_loss.item()
tr_step += 1
else:
avg_tr_loss /= tr_step
model.eval()
for x_mb, y_mb in tqdm(val_dl):
x_mb = x_mb.to(device)
y_mb = y_mb.to(device)
with torch.no_grad():
score = model(x_mb)
val_loss = loss_fn(score, y_mb)
avg_val_loss += val_loss.item()
val_step += 1
else:
avg_val_loss /= val_step
tqdm.write('epoch : {}, tr_loss : {:.3f}, val_loss : {:.3f}'.format(
epoch + 1, avg_tr_loss, avg_val_loss))
ckpt = {
'epoch': epoch,
'model_state_dict': model.state_dict(),
'opt_state_dict': opt.state_dict()
}
savepath = os.path.join(os.getcwd(), params['filepath'].get('ckpt'))
torch.save(ckpt, savepath)
def get_data_loaders(dataset_config, tokenizer, batch_size):
tr_ds = Corpus(dataset_config.train, tokenizer.split_and_transform)
tr_dl = DataLoader(tr_ds, batch_size=batch_size, shuffle=True, num_workers=4, drop_last=True)
val_ds = Corpus(dataset_config.validation, tokenizer.split_and_transform)
val_dl = DataLoader(val_ds, batch_size=batch_size, num_workers=4)
return tr_dl, val_dl
def main():
train_path = Path.cwd() / 'data_in' / 'train.txt'
val_path = Path.cwd() / 'data_in' / 'val.txt'
vocab_path = Path.cwd() / 'data_in' / 'vocab.pkl'
with open(vocab_path, mode='rb') as io:
vocab = pickle.load(io)
tokenizer = MeCab()
padder = PadSequence(length=70, pad_val=vocab.token_to_idx['<pad>'])
tr_ds = Corpus(train_path, vocab, tokenizer, padder)
tr_dl = DataLoader(tr_ds, batch_size=1024, shuffle=True, num_workers=1, drop_last=True)
val_ds = Corpus(val_path, vocab, tokenizer, padder)
val_dl = DataLoader(val_ds, batch_size=1024)
model = Net(vocab_len=len(vocab))
loss_fn = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=0.01)
for epoch in range(1):
model.train()
index = 0
acc = 0
for label, sen1, sen2 in tqdm(tr_dl, disable=True):
optimizer.zero_grad()
pre_label = model(sen1, sen2)
loss = loss_fn(pre_label, label)
loss.backward()
optimizer.step()
pred_cls = pre_label.data.max(1)[1]
acc += pred_cls.eq(label.data).cpu().sum()
print("epoch: {}, index: {}, loss: {}".format((epoch + 1), index, loss.item()))
index += len(label)
print('Accuracy : %d %%' % (
100 * acc / index))
for epoch in range(1):
model.train()
index = 0
acc = 0
for label, sen1, sen2 in tqdm(val_dl, disable=True):
optimizer.zero_grad()
pre_label = model(sen1, sen2)
loss = loss_fn(pre_label, label)
loss.backward()
optimizer.step()
pred_cls = pre_label.data.max(1)[1]
acc += pred_cls.eq(label.data).cpu().sum()
print("epoch: {}, index: {}, loss: {}".format((epoch + 1), index, loss.item()))
index += len(label)
print('Accuracy : %d %%' % (
100 * acc / index))
# model(restore)
checkpoint_manager = CheckpointManager(model_dir)
checkpoint = checkpoint_manager.load_checkpoint(args.restore_file + '.tar')
model = ConvRec(num_classes=model_config.num_classes,
embedding_dim=model_config.embedding_dim,
hidden_dim=model_config.hidden_dim,
vocab=tokenizer.vocab)
model.load_state_dict(checkpoint['model_state_dict'])
# evaluation
model.eval()
summary_manager = SummaryManager(model_dir)
filepath = getattr(data_config, args.data_name)
ds = Corpus(filepath,
tokenizer.split_and_transform,
min_length=model_config.min_length,
pad_val=tokenizer.vocab.to_indices(' '))
dl = DataLoader(ds,
batch_size=model_config.batch_size,
num_workers=4,
collate_fn=batchify)
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
model.to(device)
summary = evaluate(model, dl, {
'loss': nn.CrossEntropyLoss(),
'acc': acc
}, device)
# model (restore)
checkpoint_manager = CheckpointManager(model_dir)
checkpoint = checkpoint_manager.load_checkpoint("best.tar")
model = SAN(
num_classes=model_config.num_classes,
lstm_hidden_dim=model_config.lstm_hidden_dim,
hidden_dim=model_config.hidden_dim,
da=model_config.da,
r=model_config.r,
vocab=tokenizer.vocab,
)
model.load_state_dict(checkpoint["model_state_dict"])
# evaluation
filepath = getattr(data_config, args.dataset)
ds = Corpus(filepath, tokenizer.split_and_transform)
dl = DataLoader(
ds, batch_size=model_config.batch_size, num_workers=4, collate_fn=batchify
)
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
model.to(device)
summary_manager = SummaryManager(model_dir)
summary = evaluate(model, dl, {"loss": nn.CrossEntropyLoss(), "acc": acc}, device)
summary_manager.load("summary.json")
summary_manager.update({"{}".format(args.dataset): summary})
summary_manager.save("summary.json")
print("loss: {:.3f}, acc: {:.2%}".format(summary["loss"], summary["acc"]))
def validation_dataloader(self):
val_ds = Corpus(self.hparams.validation,
self.tokenizer.split_and_transform)
return DataLoader(val_ds,
batch_size=self.hparams.batch_size,
num_workers=4)
def main():
train_path = Path.cwd() / '..' / 'data_in' / 'train.txt'
val_path = Path.cwd() / '..' / 'data_in' / 'val.txt'
vocab_path = Path.cwd() / '..' / 'data_in' / 'vocab.pkl'
length = 70
dim = 300
batch_size = 1024
learning_rate = 0.01
epochs = 10
hidden = 50
with open(vocab_path, mode='rb') as io:
vocab = pickle.load(io)
train = tf.data.TextLineDataset(str(train_path)).shuffle(
buffer_size=batch_size).batch(batch_size=batch_size,
drop_remainder=True)
eval = tf.data.TextLineDataset(str(val_path)).batch(batch_size=batch_size,
drop_remainder=True)
tokenizer = MeCab()
corpus = Corpus(vocab, tokenizer)
malstm = MaLSTM(length, dim, len(vocab))
opt = tf.optimizers.Adam(learning_rate=learning_rate)
loss_fn = tf.losses.SparseCategoricalCrossentropy(from_logits=True)
'''
loss, accuracy
'''
train_loss_metric = tf.keras.metrics.Mean(name='train_loss')
train_acc_metric = tf.keras.metrics.SparseCategoricalAccuracy(
name='train_accuracy')
val_loss_metric = tf.keras.metrics.Mean(name='val_loss')
val_acc_metric = tf.keras.metrics.SparseCategoricalAccuracy(
name='val_accuracy')
for epoch in range(epochs):
train_loss_metric.reset_states()
train_acc_metric.reset_states()
val_loss_metric.reset_states()
val_acc_metric.reset_states()
tf.keras.backend.set_learning_phase(1)
for step, val in tqdm(enumerate(train)):
sen1, sen2, label = corpus.token2idx(val)
with tf.GradientTape() as tape:
logits = malstm(sen1, sen2)
train_loss = loss_fn(label, logits)
grads = tape.gradient(target=train_loss,
sources=malstm.trainable_variables)
opt.apply_gradients(
grads_and_vars=zip(grads, malstm.trainable_variables))
train_loss_metric.update_state(train_loss)
train_acc_metric.update_state(label, logits)
tr_loss = train_loss_metric.result()
tqdm.write('epoch : {}, tr_acc : {:.3f}%, tr_loss : {:.3f}'.format(
epoch + 1,
train_acc_metric.result() * 100, tr_loss))
tokenizer = Tokenizer(vocab=vocab,
split_fn=ptr_tokenizer.tokenize,
pad_fn=pad_sequence)
# model (restore)
checkpoint_manager = CheckpointManager(model_dir)
checkpoint = checkpoint_manager.load_checkpoint(args.restore_file + '.tar')
config = BertConfig('pretrained/bert_config.json')
model = BertClassifier(config,
num_labels=model_config.num_classes,
vocab=tokenizer.vocab)
model.load_state_dict(checkpoint['model_state_dict'])
# evaluation
filepath = getattr(data_config, args.data_name)
ds = Corpus(filepath, tokenizer.preprocess)
dl = DataLoader(ds, batch_size=model_config.batch_size, num_workers=4)
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
model.to(device)
summary_manager = SummaryManager(model_dir)
summary = evaluate(model, dl, {
'loss': nn.CrossEntropyLoss(),
'acc': acc
}, device)
summary_manager.load('summary.json')
summary_manager.update({'{}'.format(args.data_name): summary})
summary_manager.save('summary.json')
def main(argv):
train_data = Path.cwd() / '..' / 'data_in' / 'train.txt'
val_data = Path.cwd() / '..' / 'data_in' / 'val.txt'
test_data = Path.cwd() / '..' / 'data_in' / 'test.txt'
dev_data = Path.cwd() / '..' / 'data_in' / 'dev.txt'
# init params
classes = FLAGS.classes
max_length = FLAGS.length
epochs = FLAGS.epochs
learning_rate = FLAGS.learning_rate
dim = FLAGS.embedding_dim
global_step = 1000
batch_size = FLAGS.batch_size
with open(Path.cwd() / '..' / 'data_in' / 'vocab.pkl', mode='rb') as io:
vocab = pickle.load(io)
train = tf.data.TextLineDataset(str(train_data)).shuffle(
buffer_size=batch_size).batch(batch_size=batch_size)
eval = tf.data.TextLineDataset(str(val_data)).batch(batch_size=batch_size)
test = tf.data.TextLineDataset(str(test_data)).batch(batch_size=batch_size)
dev = tf.data.TextLineDataset(str(dev_data)).batch(batch_size=batch_size)
padder = PadSequence(max_length,
pad_val=vocab.to_indices(vocab.padding_token))
processing = Corpus(vocab=vocab, split_fn=Split(), pad_fn=padder)
# create model
char_cnn = CharCNN(vocab=vocab, classes=classes, dim=dim)
# create optimizer & loss_fn
opt = tf.optimizers.Adam(learning_rate=learning_rate)
loss_fn = tf.losses.SparseCategoricalCrossentropy()
train_loss_metric = tf.keras.metrics.Mean(name='train_loss')
train_acc_metric = tf.keras.metrics.SparseCategoricalAccuracy(
name='train_accuracy')
val_loss_metric = tf.keras.metrics.Mean(name='val_loss')
val_acc_metric = tf.keras.metrics.SparseCategoricalAccuracy(
name='val_accuracy')
# train_summary_writer = tf.summary.create_file_writer('./data_out/summaries/train')
# eval_summary_writer = tf.summary.create_file_writer('./data_out/summaries/eval')
# ckpt = tf.train.Checkpoint(step=tf.Variable(1), optimizer=opt, net=char_cnn)
# manager = tf.train.CheckpointManager(ckpt, './data_out/tf_ckpts', max_to_keep=3)
# ckpt.restore(manager.latest_checkpoint)
#
# if manager.latest_checkpoint:
# print("Restored from {}".format(manager.latest_checkpoint))
# else:
# print("Initializing from scratch.")
#training
for epoch in tqdm(range(epochs), desc='epochs'):
train_loss_metric.reset_states()
train_acc_metric.reset_states()
val_loss_metric.reset_states()
val_acc_metric.reset_states()
tf.keras.backend.set_learning_phase(1)
#with train_summary_writer.as_default():
for step, val in tqdm(enumerate(train), desc='steps'):
data, label = processing.token2idex(val)
with tf.GradientTape() as tape:
logits = char_cnn(data)
train_loss = loss_fn(label, logits)
#ckpt.step.assign_add(1)
grads = tape.gradient(target=train_loss,
sources=char_cnn.trainable_variables)
opt.apply_gradients(
grads_and_vars=zip(grads, char_cnn.trainable_variables))
train_loss_metric.update_state(train_loss)
train_acc_metric.update_state(label, logits)
# if tf.equal(opt.iterations % global_step, 0):
# tf.summary.scalar('loss', train_loss_metric.result(), step=opt.iterations)
tr_loss = train_loss_metric.result()
#save_path = manager.save()
#print(save_path)
tqdm.write('epoch : {}, tr_acc : {:.3f}%, tr_loss : {:.3f}'.format(
epoch + 1,
train_acc_metric.result() * 100, tr_loss))
label_vocab = pickle.load(io)
token_tokenizer = Tokenizer(token_vocab, split_to_self)
label_tokenizer = Tokenizer(label_vocab, split_to_self)
# model (restore)
checkpoint_manager = CheckpointManager(model_dir)
checkpoint = checkpoint_manager.load_checkpoint(args.restore_file + ".tar")
model = BilstmCRF(label_vocab, token_vocab, model_config.lstm_hidden_dim)
model.load_state_dict(checkpoint["model_state_dict"])
# evaluation
summary_manager = SummaryManager(model_dir)
filepath = getattr(data_config, args.data_name)
ds = Corpus(
filepath,
token_tokenizer.split_and_transform,
label_tokenizer.split_and_transform,
)
dl = DataLoader(ds,
batch_size=model_config.batch_size,
num_workers=4,
collate_fn=batchify)
device = torch.device(
"cuda") if torch.cuda.is_available() else torch.device("cpu")
model.to(device)
f1_score = get_f1_score(model, dl, device)
summary_manager.load("summary.json")
summary_manager._summary[args.data_name].update({"f1": f1_score})
summary_manager.save("summary.json")
# loading trained model
save_path = params['filepath'].get('ckpt')
# save_path = 'tokenize.pth'
ckpt = torch.load(save_path)
config = BertConfig('bert/bert_config.json')
model = BertTagger(config=config, num_labels=len(label_vocab.token_to_idx), vocab=token_vocab)
model.load_state_dict(ckpt['model_state_dict'])
# loading datasets
batch_size = params['training'].get('batch_size')
train_path = params['filepath'].get('train')
val_path = params['filepath'].get('val')
test_path = params['filepath'].get('test')
train_data = Corpus(train_path, token_vocab.to_indices, label_vocab.to_indices)
train_loader = DataLoader(train_data, batch_size=batch_size, shuffle=False, num_workers=16,
drop_last=True, collate_fn=batchify)
val_data = Corpus(val_path, token_vocab.to_indices, label_vocab.to_indices)
val_loader = DataLoader(val_data, batch_size=batch_size, shuffle=False, num_workers=16,
drop_last=True, collate_fn=batchify)
test_data = Corpus(test_path, token_vocab.to_indices, label_vocab.to_indices)
test_loader = DataLoader(test_data, batch_size=batch_size, shuffle=False, num_workers=16,
drop_last=True, collate_fn=batchify)
# using gpu
device = torch.device('cuda:1') if torch.cuda.is_available() else torch.device('cpu')
model.to(device)
pad_sequence = PadSequence(length=model_config.length,
pad_val=vocab.to_indices(vocab.padding_token))
preprocessor = PreProcessor(vocab=vocab,
split_fn=ptr_tokenizer.tokenize,
pad_fn=pad_sequence)
# model
config = BertConfig('pretrained/bert_config.json')
model = PairwiseClassifier(config,
num_classes=model_config.num_classes,
vocab=preprocessor.vocab)
bert_pretrained = torch.load('pretrained/pytorch_model.bin')
model.load_state_dict(bert_pretrained, strict=False)
# training
tr_ds = Corpus(data_config.tr, preprocessor.preprocess)
tr_dl = DataLoader(tr_ds,
batch_size=model_config.batch_size,
shuffle=True,
num_workers=4,
drop_last=True)
val_ds = Corpus(data_config.val, preprocessor.preprocess)
val_dl = DataLoader(val_ds, batch_size=model_config.batch_size)
loss_fn = nn.CrossEntropyLoss()
opt = optim.Adam([
{
"params": model.bert.parameters(),
"lr": model_config.learning_rate / 100
},
{
model_config = Config(json_path=model_dir / 'config.json')
# tokenizer
with open(data_config.vocab, mode='rb') as io:
vocab = pickle.load(io)
tokenizer = Tokenizer(vocab=vocab, split_fn=split_to_jamo)
# model
model = ConvRec(num_classes=model_config.num_classes,
embedding_dim=model_config.embedding_dim,
hidden_dim=model_config.hidden_dim,
vocab=tokenizer.vocab)
# training
tr_ds = Corpus(data_config.train,
tokenizer.split_and_transform,
min_length=model_config.min_length,
pad_val=tokenizer.vocab.to_indices(' '))
tr_dl = DataLoader(tr_ds,
batch_size=model_config.batch_size,
shuffle=True,
num_workers=4,
collate_fn=batchify,
drop_last=True)
val_ds = Corpus(data_config.validation,
tokenizer.split_and_transform,
min_length=model_config.min_length,
pad_val=tokenizer.vocab.to_indices(' '))
val_dl = DataLoader(val_ds,
batch_size=model_config.batch_size,
collate_fn=batchify)
def main(json_path):
cwd = Path.cwd()
with open(cwd / json_path) as io:
params = json.loads(io.read())
# tokenizer
vocab_path = params['filepath'].get('vocab')
with open(cwd / vocab_path, mode='rb') as io:
vocab = pickle.load(io)
tokenizer = Tokenizer(vocab=vocab, split_fn=MeCab().morphs)
# model
num_classes = params['model'].get('num_classes')
lstm_hidden_dim = params['model'].get('lstm_hidden_dim')
hidden_dim = params['model'].get('hidden_dim')
da = params['model'].get('da')
r = params['model'].get('r')
model = SAN(num_classes=num_classes,
lstm_hidden_dim=lstm_hidden_dim,
hidden_dim=hidden_dim,
da=da,
r=r,
vocab=tokenizer.vocab)
# training
epochs = params['training'].get('epochs')
batch_size = params['training'].get('batch_size')
learning_rate = params['training'].get('learning_rate')
global_step = params['training'].get('global_step')
tr_path = cwd / params['filepath'].get('tr')
val_path = cwd / params['filepath'].get('val')
tr_ds = Corpus(tr_path, tokenizer.split_and_transform)
tr_dl = DataLoader(tr_ds,
batch_size=batch_size,
shuffle=True,
num_workers=4,
drop_last=True,
collate_fn=batchify)
val_ds = Corpus(val_path, tokenizer.split_and_transform)
val_dl = DataLoader(val_ds,
batch_size=batch_size,
num_workers=4,
collate_fn=batchify)
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)
writer = SummaryWriter('./runs/{}'.format(params['version']))
for epoch in tqdm(range(epochs), desc='epochs'):
tr_loss = 0
model.train()
for step, mb in tqdm(enumerate(tr_dl), desc='steps', total=len(tr_dl)):
queries_a_mb, queries_b_mb, y_mb = map(lambda elm: elm.to(device),
mb)
queries_mb = (queries_a_mb, queries_b_mb)
opt.zero_grad()
score, queries_a_attn_mat, queries_b_attn_mat = model(queries_mb)
a_reg = regularize(queries_a_attn_mat, r, device)
b_reg = regularize(queries_b_attn_mat, r, device)
mb_loss = loss_fn(score, y_mb)
mb_loss.add_(a_reg)
mb_loss.add_(b_reg)
mb_loss.backward()
opt.step()
tr_loss += mb_loss.item()
if (epoch * len(tr_dl) + step) % global_step == 0:
val_loss = evaluate(model, val_dl, loss_fn, device)
writer.add_scalars('loss', {
'train': tr_loss / (step + 1),
'validation': val_loss
},
epoch * len(tr_dl) + step)
model.train()
else:
tr_loss /= (step + 1)
val_loss = evaluate(model, val_dl, loss_fn, device)
scheduler.step(val_loss)
tqdm.write('epoch : {}, tr_loss : {:.3f}, val_loss : {:.3f}'.format(
epoch + 1, tr_loss, val_loss))
ckpt = {
'model_state_dict': model.state_dict(),
'opt_state_dict': opt.state_dict()
}
save_path = cwd / params['filepath'].get('ckpt')
torch.save(ckpt, save_path)
preprocessor = PreProcessor(ptr_tokenizer, model_config.max_len)
# Load Model
config_filepath = ptr_dir / "{}-config.json".format(args.type)
config = BertConfig.from_pretrained(config_filepath,
output_hidden_states=False)
model = BIIN(config,
vocab,
model_config.hidden_size,
enc_num_layers=len(model_config.hidden_size))
# Data Loader
tr_ds = Corpus(data_config.tr_path,
preprocessor.preprocess,
sep='\t',
doc_col='question1',
label_col='is_duplicate',
is_pair=True,
doc_col_second='question2')
val_ds = Corpus(data_config.dev_path,
preprocessor.preprocess,
sep='\t',
doc_col='question1',
label_col='is_duplicate',
is_pair=True,
doc_col_second='question2')
tr_dl = DataLoader(tr_ds,
batch_size=model_config.batch_size,
shuffle=True,
num_workers=4,
drop_last=True)
pad_val=vocab.to_indices(
vocab.padding_token))
preprocessor = PreProcessor(vocab=vocab,
split_fn=ptr_tokenizer,
pad_fn=pad_sequence)
# model
config = BertConfig(ptr_config.config)
model = PairwiseClassifier(config,
num_classes=model_config.num_classes,
vocab=preprocessor.vocab)
# bert_pretrained = torch.load(ptr_config.bert)
# model.load_state_dict(bert_pretrained, strict=False)
# training
tr_ds = Corpus(data_config.train, preprocessor.preprocess)
tr_dl = DataLoader(tr_ds,
batch_size=model_config.batch_size,
shuffle=True,
num_workers=4,
drop_last=True)
val_ds = Corpus(data_config.validation, preprocessor.preprocess)
val_dl = DataLoader(val_ds,
batch_size=model_config.batch_size,
num_workers=4)
loss_fn = nn.CrossEntropyLoss()
opt = optim.Adam([
{
"params": model.bert.parameters(),
"lr": model_config.learning_rate / 100
def main(json_path):
cwd = Path.cwd()
with open(cwd / json_path) as io:
params = json.loads(io.read())
# tokenizer
vocab_path = params['filepath'].get('vocab')
with open(cwd / vocab_path, mode='rb') as io:
vocab = pickle.load(io)
length = params['padder'].get('length')
padder = PadSequence(length=length, pad_val=vocab.to_indices(vocab.padding_token))
tokenizer = Tokenizer(vocab=vocab, split_fn=MeCab().morphs, pad_fn=padder)
# model
num_classes = params['model'].get('num_classes')
model = SenCNN(num_classes=num_classes, vocab=tokenizer.vocab)
# training
epochs = params['training'].get('epochs')
batch_size = params['training'].get('batch_size')
learning_rate = params['training'].get('learning_rate')
global_step = params['training'].get('global_step')
tr_path = cwd / params['filepath'].get('tr')
val_path = cwd / params['filepath'].get('val')
tr_ds = Corpus(tr_path, tokenizer.split_and_transform)
tr_dl = DataLoader(tr_ds, batch_size=batch_size, shuffle=True, num_workers=4, drop_last=True)
val_ds = Corpus(val_path, tokenizer.split_and_transform)
val_dl = DataLoader(val_ds, batch_size=batch_size)
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)
writer = SummaryWriter('./runs/{}'.format(params['version']))
for epoch in tqdm(range(epochs), desc='epochs'):
tr_loss = 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()
mb_loss = loss_fn(model(x_mb), y_mb)
mb_loss.backward()
clip_grad_norm_(model._fc.weight, 5)
opt.step()
tr_loss += mb_loss.item()
if (epoch * len(tr_dl) + step) % global_step == 0:
val_loss = evaluate(model, val_dl, loss_fn, device)
writer.add_scalars('loss', {'train': tr_loss / (step + 1),
'val': val_loss}, epoch * len(tr_dl) + step)
model.train()
else:
tr_loss /= (step + 1)
val_loss = evaluate(model, val_dl, loss_fn, device)
scheduler.step(val_loss)
tqdm.write('epoch : {}, tr_loss : {:.3f}, val_loss : {:.3f}'.format(epoch + 1, tr_loss, val_loss))
ckpt = {'model_state_dict': model.state_dict(),
'opt_state_dict': opt.state_dict()}
save_path = cwd / params['filepath'].get('ckpt')
torch.save(ckpt, save_path)
# tokenizer
with open(data_config.vocab, mode='rb') as io:
vocab = pickle.load(io)
pad_sequence = PadSequence(length=model_config.length,
pad_val=vocab.to_indices(vocab.padding_token))
tokenizer = Tokenizer(vocab=vocab,
split_fn=split_to_jamo,
pad_fn=pad_sequence)
# model
model = CharCNN(num_classes=model_config.num_classes,
embedding_dim=model_config.embedding_dim,
vocab=tokenizer.vocab)
# training
tr_ds = Corpus(data_config.train, tokenizer.split_and_transform)
tr_dl = DataLoader(tr_ds,
batch_size=model_config.batch_size,
shuffle=True,
num_workers=4,
drop_last=True)
val_ds = Corpus(data_config.validation, tokenizer.split_and_transform)
val_dl = DataLoader(val_ds,
batch_size=model_config.batch_size,
num_workers=4)
loss_fn = nn.CrossEntropyLoss()
opt = optim.Adam(params=model.parameters(), lr=model_config.learning_rate)
scheduler = ReduceLROnPlateau(opt, patience=5)
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
def main(argv):
train_data = Path.cwd() / 'data_in' / 'train.txt'
val_data = Path.cwd() / 'data_in' / 'val.txt'
with open(Path.cwd() / 'data_in' / 'vocab.pkl', mode='rb') as io:
vocab = pickle.load(io)
train = tf.data.TextLineDataset(str(train_data)).shuffle(buffer_size=1000).batch(batch_size=FLAGS.batch_size,
drop_remainder=True)
eval = tf.data.TextLineDataset(str(val_data)).batch(batch_size=FLAGS.batch_size, drop_remainder=True)
tokenized = MeCab()
processing = Corpus(vocab=vocab, tokenizer=tokenized)
# init params
classes = FLAGS.classes
max_length = FLAGS.length
epochs = FLAGS.epochs
learning_rate = FLAGS.learning_rate
global_step = 1000
# create model
sen_cnn = SenCNN(vocab=vocab, classes=classes)
# create optimizer & loss_fn
opt = tf.optimizers.Adam(learning_rate=learning_rate)
loss_fn = tf.losses.SparseCategoricalCrossentropy(from_logits=True)
train_loss_metric = tf.keras.metrics.Mean(name='train_loss')
train_acc_metric = tf.keras.metrics.SparseCategoricalAccuracy(name='train_accuracy')
val_loss_metric = tf.keras.metrics.Mean(name='val_loss')
val_acc_metric = tf.keras.metrics.SparseCategoricalAccuracy(name='val_accuracy')
train_summary_writer = tf.summary.create_file_writer('./data_out/summaries/train')
eval_summary_writer = tf.summary.create_file_writer('./data_out/summaries/eval')
ckpt = tf.train.Checkpoint(step=tf.Variable(1), optimizer=opt, net=sen_cnn)
manager = tf.train.CheckpointManager(ckpt, './data_out/tf_ckpts', max_to_keep=3)
ckpt.restore(manager.latest_checkpoint)
if manager.latest_checkpoint:
print("Restored from {}".format(manager.latest_checkpoint))
else:
print("Initializing from scratch.")
# training
for epoch in tqdm(range(epochs), desc='epochs'):
train_loss_metric.reset_states()
train_acc_metric.reset_states()
val_loss_metric.reset_states()
val_acc_metric.reset_states()
tf.keras.backend.set_learning_phase(1)
tr_loss = 0
with train_summary_writer.as_default():
for step, val in tqdm(enumerate(train), desc='steps'):
data, label = processing.token2idex(val)
with tf.GradientTape() as tape:
logits = sen_cnn(data)
train_loss = loss_fn(label, logits)
ckpt.step.assign_add(1)
grads = tape.gradient(target=train_loss, sources=sen_cnn.trainable_variables)
opt.apply_gradients(grads_and_vars=zip(grads, sen_cnn.trainable_variables))
# tr_loss += pred_loss.numpy()
train_loss_metric.update_state(train_loss)
train_acc_metric.update_state(label, logits)
if tf.equal(opt.iterations % global_step, 0):
tf.summary.scalar('loss', train_loss_metric.result(), step=opt.iterations)
# else:
# tr_loss /= (step + 1)
# print("t_loss {}".format(tr_loss))
tr_loss = train_loss_metric.result()
save_path = manager.save()
print(save_path)
tf.keras.backend.set_learning_phase(0)
val_loss = 0
with eval_summary_writer.as_default():
for step, val in tqdm(enumerate(eval), desc='steps'):
data, label = processing.token2idex(val)
logits = sen_cnn(data)
val_loss = loss_fn(label, logits)
# val_loss += mb_loss.numpy()
val_loss_metric.update_state(val_loss)
val_acc_metric.update_state(label, logits)
tf.summary.scalar('loss', val_loss_metric.result(), step=step)
val_loss = val_loss_metric.result()
tqdm.write(
'epoch : {}, tr_acc : {:.3f}%, tr_loss : {:.3f}, val_acc : {:.3f}%, val_loss : {:.3f}'.format(epoch + 1,
train_acc_metric.result() * 100,
tr_loss,
val_acc_metric.result() * 100,
val_loss))