def batches(self):
self._shuffle()
for bidx in range(self.n_batches):
s_idxs = compute_batch_idxs(bidx * self.half_batch_size,
self.half_batch_size, self.n_s_samples)
t_idxs = compute_batch_idxs(bidx * self.half_batch_size,
self.half_batch_size, self.n_t_samples)
yield dict(xs=self.xs[s_idxs],
ys=self.ys[s_idxs],
ws=self.ws[s_idxs],
xt=self.xt[t_idxs],
yt=self.yt[t_idxs],
wt=self.wt[t_idxs])
def predict(self, inputs, lengths):
sess = tf.get_default_session()
n_samples = len(inputs)
n_batches = utils.compute_n_batches(n_samples, self.batch_size)
probs = np.zeros((len(inputs), self.max_len, self.output_dim))
for bidx in range(n_batches):
idxs = utils.compute_batch_idxs(bidx * self.batch_size, self.batch_size, n_samples)
probs[idxs] = sess.run(self.probs, feed_dict={
self.inputs:inputs[idxs],
self.lengths:lengths[idxs],
self.dropout_keep_prop_ph: 1.
})
preds = np.argmax(probs, axis=-1)
return probs, preds
def generate(self, z=None, c=None, batch_size=100):
# setup
sess = tf.get_default_session()
# sample noise if not provided
if z is None and c is None:
z, c, _ = self._sample_noise(batch_size)
elif z is None:
batch_size = len(c)
z, _, _ = self._sample_noise(batch_size)
else:
batch_size = len(c)
# at this point, z dictates number of samples
n_samples = len(z)
n_batches = compute_n_batches(n_samples, batch_size)
# allocate return containers
gx = np.zeros((n_samples, self.input_dim))
scores = np.zeros((n_samples, 1))
# formulate outputs
outputs = [self.gx, self.gen_scores]
# run the batches
for bidx in range(n_batches):
idxs = compute_batch_idxs(bidx * batch_size, batch_size, n_samples)
feed = {self.z: z[idxs], self.c: c[idxs]}
fetched = sess.run(outputs, feed_dict=feed)
# unpack
gx[idxs] = fetched[0]
scores[idxs] = fetched[1]
# return the relevant info
return dict(gx=gx, scores=scores)
def predict(self, x, tgt=True, batch_size=100):
# setup
sess = tf.get_default_session()
n_samples = len(x)
n_batches = compute_n_batches(n_samples, batch_size)
probs = np.zeros((n_samples, self.output_dim))
# decide between src or tgt probs
outputs_list = [self.tgt_task_probs] if tgt else [self.src_task_probs]
x_ph = self.xt if tgt else self.xs
# compute probs
for bidx in range(n_batches):
idxs = compute_batch_idxs(bidx * batch_size,
batch_size,
n_samples,
fill='none')
probs[idxs] = sess.run(outputs_list,
feed_dict={
x_ph: x[idxs],
self.dropout_keep_prob_ph: 1.
})
return probs
def batches(self):
self._shuffle()
for bidx in range(self.n_batches):
idxs = compute_batch_idxs(bidx * self.batch_size, self.batch_size,
self.n_samples)
yield dict(x=self.x[idxs])
def train(self,
data,
n_epochs=100,
batch_size=100,
writer=None,
val_writer=None,
stop_early=False):
sess = tf.get_default_session()
if self.batch_size is not None:
batch_size = self.batch_size
n_samples = len(data['x_train'])
n_batches = utils.compute_n_batches(n_samples, batch_size)
n_val_samples = len(data['x_val'])
n_val_batches = utils.compute_n_batches(n_val_samples, batch_size)
last_val_losses = collections.deque([np.inf] * 2)
for epoch in range(n_epochs):
# shuffle train set
idxs = np.random.permutation(len(data['x_train']))
data['x_train'] = data['x_train'][idxs]
data['y_train'] = data['y_train'][idxs]
# train
total_loss = 0
for bidx in range(n_batches):
idxs = utils.compute_batch_idxs(bidx * batch_size, batch_size,
n_samples)
feed_dict = {
self.inputs: data['x_train'][idxs],
self.targets: data['y_train'][idxs],
}
outputs_list = [
self.loss, self.summary_op, self.global_step, self.train_op
]
loss, summary, step, _ = sess.run(outputs_list,
feed_dict=feed_dict)
total_loss += loss
if writer is not None:
writer.add_summary(summary, step)
sys.stdout.write(
'\repoch: {} / {} batch: {} / {} loss: {}'.format(
epoch + 1, n_epochs, bidx + 1, n_batches,
total_loss / (bidx + 1)))
print('\n')
# val
total_loss = 0
for bidx in range(n_val_batches):
s = bidx * batch_size
e = s + batch_size
idxs = utils.compute_batch_idxs(bidx * batch_size, batch_size,
n_val_samples)
feed_dict = {
self.inputs: data['x_val'][idxs],
self.targets: data['y_val'][idxs],
self.dropout_keep_prop_ph: 1.
}
outputs_list = [self.loss, self.summary_op, self.global_step]
loss, summary, step = sess.run(outputs_list,
feed_dict=feed_dict)
total_loss += loss
if val_writer is not None:
val_writer.add_summary(summary, step)
sys.stdout.write(
'\rval epoch: {} / {} batch: {} / {} loss: {}'.format(
epoch + 1, n_epochs, bidx + 1, n_val_batches,
total_loss / (bidx + 1)))
print('\n')
if stop_early:
if all(total_loss > v for v in last_val_losses):
break
last_val_losses.popleft()
last_val_losses.append(total_loss)
def train(
self,
data,
n_epochs=100,
writer=None,
val_writer=None):
sess = tf.get_default_session()
n_samples = len(data['train_x'])
n_batches = utils.compute_n_batches(n_samples, self.batch_size)
n_val_samples = len(data['val_x'])
n_val_batches = utils.compute_n_batches(n_val_samples, self.batch_size)
for epoch in range(n_epochs):
# shuffle train set
idxs = np.random.permutation(len(data['train_x']))
data['train_x'] = data['train_x'][idxs]
data['train_y'] = data['train_y'][idxs]
data['train_lengths'] = data['train_lengths'][idxs]
# train
total_loss = 0
for bidx in range(n_batches):
idxs = utils.compute_batch_idxs(bidx * self.batch_size, self.batch_size, n_samples)
feed_dict = {
self.inputs:data['train_x'][idxs],
self.targets:data['train_y'][idxs],
self.lengths:data['train_lengths'][idxs]
}
outputs_list = [self.loss, self.summary_op, self.global_step, self.train_op]
loss, summary, step, _ = sess.run(outputs_list, feed_dict=feed_dict)
total_loss += loss
writer.add_summary(summary, step)
sys.stdout.write('\repoch: {} / {} batch: {} / {} loss: {}'.format(
epoch+1, n_epochs, bidx+1, n_batches,
total_loss / (self.batch_size * (bidx+1))))
self.validate(data['train_x'], data['train_y'], data['train_lengths'], writer, epoch)
print('\n')
# val
total_loss = 0
for bidx in range(n_val_batches):
s = bidx * self.batch_size
e = s + self.batch_size
idxs = utils.compute_batch_idxs(bidx * self.batch_size, self.batch_size, n_val_samples)
feed_dict = {
self.inputs:data['val_x'][idxs],
self.targets:data['val_y'][idxs],
self.lengths:data['val_lengths'][idxs],
self.dropout_keep_prop_ph: 1.
}
outputs_list = [self.loss, self.summary_op, self.global_step]
loss, summary, step = sess.run(outputs_list, feed_dict=feed_dict)
total_loss += loss
val_writer.add_summary(summary, step)
sys.stdout.write('\rval epoch: {} / {} batch: {} / {} loss: {}'.format(
epoch+1, n_epochs, bidx+1, n_val_batches,
total_loss / (self.batch_size * (bidx+1))))
self.validate(data['val_x'], data['val_y'], data['val_lengths'], val_writer, epoch)
print('\n')