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}"
)
tokenizer = get_tokenizer(dataset_config, model_config)
checkpoint_manager = CheckpointManager(exp_dir)
checkpoint = checkpoint_manager.load_checkpoint("best.tar")
model = CharCNN(num_classes=model_config.num_classes,
embedding_dim=model_config.embedding_dim,
vocab=tokenizer.vocab)
model.load_state_dict(checkpoint["model_state_dict"])
summary_manager = SummaryManager(exp_dir)
filepath = getattr(dataset_config, args.data)
ds = Corpus(filepath, tokenizer.split_and_transform)
dl = DataLoader(ds, batch_size=args.batch_size, num_workers=4)
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)
summary_manager.load("summary.json")
summary_manager.update({f"{args.data}": summary})
summary_manager.save("summary.json")
print(f"loss: {summary['loss']:.3f}, acc: {summary['acc']:.2%}")
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=split_to_jamo, pad_fn=padder)
# model (restore)
save_path = cwd / params['filepath'].get('ckpt')
ckpt = torch.load(save_path)
num_classes = params['model'].get('num_classes')
embedding_dim = params['model'].get('embedding_dim')
model = CharCNN(num_classes=num_classes,
embedding_dim=embedding_dim,
vocab=tokenizer.vocab)
model.load_state_dict(ckpt['model_state_dict'])
# evaluation
batch_size = params['training'].get('batch_size')
tr_path = cwd / params['filepath'].get('tr')
val_path = cwd / params['filepath'].get('val')
tst_path = cwd / params['filepath'].get('tst')
tr_ds = Corpus(tr_path, tokenizer.split_and_transform)
tr_dl = DataLoader(tr_ds, batch_size=batch_size, num_workers=4)
val_ds = Corpus(val_path, tokenizer.split_and_transform)
val_dl = DataLoader(val_ds, batch_size=batch_size, num_workers=4)
tst_ds = Corpus(tst_path, tokenizer.split_and_transform)
tst_dl = DataLoader(tst_ds, batch_size=batch_size, num_workers=4)
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
model.to(device)
tr_acc = get_accuracy(model, tr_dl, device)
val_acc = get_accuracy(model, val_dl, device)
tst_acc = get_accuracy(model, tst_dl, device)
print('tr_acc: {:.2%}, val_acc: {:.2%}, tst_acc: {:.2%}'.format(
tr_acc, val_acc, tst_acc))
def evaluate(cfgpath):
# parsing json
with open(os.path.join(os.getcwd(), cfgpath)) as io:
params = json.loads(io.read())
# restoring model
savepath = os.path.join(os.getcwd(), params['filepath'].get('ckpt'))
ckpt = torch.load(savepath)
tokenizer = JamoTokenizer()
padder = PadSequence(300)
model = CharCNN(num_classes=params['model'].get('num_classes'),
embedding_dim=params['model'].get('embedding_dim'),
dic=tokenizer.token2idx)
model.load_state_dict(ckpt['model_state_dict'])
model.eval()
device = torch.device(
'cuda') if torch.cuda.is_available() else torch.device('cpu')
model.to(device)
# creating dataset, dataloader
tst_filepath = os.path.join(os.getcwd(), params['filepath'].get('tst'))
tst_ds = Corpus(tst_filepath, tokenizer, padder)
tst_dl = DataLoader(tst_ds, batch_size=128, num_workers=4)
# evaluation
correct_count = 0
for x_mb, y_mb in tqdm(tst_dl):
x_mb = x_mb.to(device)
y_mb = y_mb.to(device)
with torch.no_grad():
y_mb_hat = model(x_mb)
y_mb_hat = torch.max(y_mb_hat, 1)[1]
correct_count += (y_mb_hat == y_mb).sum().item()
print('Acc : {:.2%}'.format(correct_count / len(tst_ds)))
model_config = Config(json_path=model_dir / 'config.json')
# 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 (restore)
checkpoint_manager = CheckpointManager(model_dir)
checkpoint = checkpoint_manager.load_checkpoint(args.restore_file + '.tar')
model = CharCNN(num_classes=model_config.num_classes,
embedding_dim=model_config.embedding_dim,
vocab=tokenizer.vocab)
model.load_state_dict(checkpoint['model_state_dict'])
# evaluation
summary_manager = SummaryManager(model_dir)
filepath = getattr(data_config, args.data_name)
ds = Corpus(filepath, tokenizer.split_and_transform)
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 = evaluate(model, dl, {
'loss': nn.CrossEntropyLoss(),
model_dir = Path(args.model_dir)
data_config = Config(json_path=data_dir / 'config.json')
model_config = Config(json_path=model_dir / 'config.json')
# 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)
loss_fn = nn.CrossEntropyLoss()
opt = optim.Adam(params=model.parameters(), lr=model_config.learning_rate)
scheduler = ReduceLROnPlateau(opt, patience=5)
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 = CharCNN(num_classes=model_config.num_classes,
embedding_dim=model_config.embedding_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 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 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))
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, pad_val=vocab.to_indices(vocab.padding_token))
tokenizer = Tokenizer(vocab=vocab, split_fn=split_to_jamo, pad_fn=padder)
# model
num_classes = params['model'].get('num_classes')
embedding_dim = params['model'].get('embedding_dim')
model = CharCNN(num_classes=num_classes,
embedding_dim=embedding_dim,
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, 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)
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()
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)