def _run_epoch(self, sess, xs, labels, is_training):
loss_batches, size_batches = [], []
if is_training:
xs, labels = shuffle_pairs(xs, labels)
batches_train_xs = batches(xs, self.conf["batch_size"])
batches_train_labels = batches(labels, self.conf["batch_size"])
for train_x, train_label in zip(batches_train_xs, batches_train_labels):
feed_dict = {
self.tf_x: train_x,
self.tf_label: train_label,
self.is_training: is_training
}
self.global_step += 1
if is_training:
_, loss_val, summary = sess.run([self.optimizer, self.loss, self.merged_summary_op], feed_dict=feed_dict)
self.summary_writer_train.add_summary(summary, self.global_step)
else:
loss_val, summary = sess.run([self.loss, self.merged_summary_op], feed_dict=feed_dict)
self.summary_writer_dev.add_summary(summary, self.global_step)
loss_batches.append(loss_val)
size_batches.append(train_x.shape[0])
loss = np.average(loss_batches, weights=size_batches)
return loss
def predict(self, dataset, save_dir):
self.scaler = pickle_load(save_dir + self.save_scaler_filename)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
self.saver.restore(sess, save_dir + self.save_filename)
xs = dataset2arrays(dataset, return_labels=False)
self.scaler.transform(np.reshape(xs, (-1, 1)))
batches_xs = batches(xs, self.conf["batch_size"])
batches_p = []
for x in batches_xs:
feed_dict = {
self.tf_x: x,
self.is_training: False
}
p_val = sess.run(self.p, feed_dict=feed_dict)
p_val = p_val[:, 0]
batches_p.append(p_val)
for example, p in zip(dataset, unbatch(batches_p)):
example["p"] = p
example["prediction"] = int(p >= 0.5)
def test_batch1(self):
x = [[1], [2], [3]]
y = ['a', 'b', 'c']
x_batches_correct = [([1], [2]), ([3], )]
y_batches_correct = [('a', 'b'), ('c', )]
x_batches, y_batches = [], []
for batch in batches(list(zip(x, y)), batch_size=2, shuffle=False):
x_batch, y_batch = zip(*batch)
x_batches.append(x_batch)
y_batches.append(y_batch)
self.assertEqual(x_batches, x_batches_correct)
self.assertEqual(y_batches, y_batches_correct)
def test_batch2(self):
x = np.array([[1, 2], [3, 4], [5, 6]])
batches_correct = [[[1, 2], [3, 4]], [[5, 6]]]
for i, batch in enumerate(batches(x, batch_size=2, shuffle=False)):
self.assertEqual(batches_correct[i], batch.tolist())