def test(data, batch_size=64, filename='roc.png', **kwargs):
global args
assert args.checkpoint is not None
model = RNNModel(feature_dims=data[0].feature_dim,
model_dir=args.output_dir,
**kwargs)
model.restore(args.checkpoint)
data = list(filter(lambda d: d.seq is not None, data))
for i in tqdm(range(0, len(data), batch_size)):
x, y, length = get_feature_label(data[i:i + batch_size],
length_limit=10000)
predictions = model.predict(x, length)
for l, p in zip(data[i:i + batch_size], predictions):
l.prediction = p
# if SimpleLengthModel.data_filter(data[i]):
# x, y, length = get_feature_label(data[i:i+batch_size], length_limit=1000)
# predictions = model.predict(x, length)
# for l,p in zip(data[i:i+batch_size], predictions):
# l.prediction = p
# else:
# for l in data[i:i+batch_size]:
# l.prediction = 1 + l.length / 100000.0
predictions = list(map(attrgetter('prediction'), data))
labels = list(map(attrgetter('label'), data))
plot_roc(predictions, labels, filename=filename)
def train(train_data,
val_data,
steps=6000,
val_per_steps=300,
checkpoint_per_steps=100,
batch_size=64,
learning_rate=0.01,
**kwargs):
global args
# train_data = list(filter(SimpleLengthModel.data_filter, train_data))
# val_data = list(filter(SimpleLengthModel.data_filter, val_data))
model = RNNModel(feature_dims=train_data[0].feature_dim,
model_dir=args.output_dir,
**kwargs)
if args.checkpoint is not None:
model.restore(args.checkpoint)
data_provider = batch_data_provider(train_data, batch_size=batch_size)
for t in range(0, steps):
x, y, length = get_feature_label(next(data_provider),
length_limit=1000)
result = model.train(x, y, length, learning_rate)
logging.info("step = {}: {}".format(model.global_step, result))
if model.global_step % val_per_steps == 0:
result = val(model, val_data)
model.init_streaming()
logging.info("validation for step = {}: {}".format(
model.global_step, result))
if model.global_step % checkpoint_per_steps == 0:
model.save_checkpoint()
logging.info("save checkpoint at {}".format(model.global_step))
if model.global_step % 2000 == 0:
learning_rate *= 0.2
logging.info("current learning rate = {}".format(learning_rate))
