def main(args):
features, targets = generate_synthetic_data(args.model_type,
args.num_samples)
# split train/test sets
x_train, x_val, y_train, y_val = train_test_split(features,
targets,
test_size=0.2)
db_train = tf.data.Dataset.from_tensor_slices(
(x_train, y_train)).batch(args.batch_size_train)
db_val = tf.data.Dataset.from_tensor_slices(
(x_val, y_val)).batch(args.batch_size_eval)
if args.model_type == 'MLP':
model = MLP(num_inputs=Constants._MLP_NUM_FEATURES,
num_layers=Constants._MLP_NUM_LAYERS,
num_dims=Constants._MLP_NUM_DIMS,
num_outputs=Constants._NUM_TARGETS,
dropout_rate=args.dropout)
elif args.model_type == 'TCN':
model = TCN(nb_filters=Constants._TCN_NUM_FILTERS,
kernel_size=Constants._TCN_KERNEL_SIZE,
nb_stacks=Constants._TCN_NUM_STACK,
dilations=Constants._TCN_DIALATIONS,
padding=Constants._TCN_PADDING,
dropout_rate=args.lr)
criteon = keras.losses.MeanSquaredError()
optimizer = keras.optimizers.Adam(learning_rate=args.lr)
for epoch in range(args.max_epoch):
for step, (x, y) in enumerate(db_train):
with tf.GradientTape() as tape:
logits = model(x)
loss = criteon(y, logits)
grads = tape.gradient(loss, model.trainable_variables)
optimizer.apply_gradients(zip(grads, model.trainable_variables))
if step % 100 == 0:
print('Epoch: {}, Step: {}/{}, Loss: {}'.format(
epoch, step, int(x_train.shape[0] / args.batch_size_train),
loss))
# Perform inference and measure the speed every epoch
start_time = time.time()
for _, (x, _) in enumerate(db_val):
_ = model.predict(x)
end_time = time.time()
print("Inference speed: {} samples/s\n".format(
x_val.shape[0] / (end_time - start_time)))
for x, y in tqdm(train_loader):
padded_text, lens, mask = seq_indexer.add_padding_tensor(
x, gpu=args.gpu)
label = label_indexer.instance2tensor(y, gpu=args.gpu)
y = model(padded_text, lens, mask)
loss = criterion(y, label)
loss.backward()
optimizer.step()
train_loss += loss.item()
optimizer.zero_grad()
print('epoch', epoch + 1, ' loss: ',
train_loss / (len(dataset.train_word) / args.batch_size))
pred = model.predict(dataset.train_word,
embedding_indexer=seq_indexer,
label_indexer=label_indexer,
batch_size=args.batch_size)
train_score = eval.get_socre(pred, dataset.train_label)
pred = model.predict(dataset.dev_word,
embedding_indexer=seq_indexer,
label_indexer=label_indexer,
batch_size=args.batch_size)
dev_score = eval.get_socre(pred, dataset.dev_label)
pred = model.predict(dataset.test_word,
embedding_indexer=seq_indexer,
label_indexer=label_indexer,
batch_size=args.batch_size)
test_score = eval.get_socre(pred, dataset.test_label)