def eval_batch(self, session, mtest, batch, batch_size, get_probs=False):
batch_size_original = len(batch)
ids = tf_utils.make_array(
[item['ids'] for item in batch] + \
[batch[0]['ids'] for _ in xrange(batch_size - len(batch))],
length=self.memory_size)
ids_lengths = np.array(
[len(item['ids']) for item in batch] + \
[len(batch[0]['ids']) for _ in xrange(batch_size - len(batch))])
labels = tf_utils.make_array(
[item['labels'] for item in batch] + \
[batch[0]['labels'] for _ in xrange(batch_size - len(batch))])
labels = labels[:, list(self.label_types)]
if get_probs:
all_probs = session.run(
mtest.all_probs,
tf_utils.filter_none_keys({
mtest.ids: ids,
mtest.ids_lengths: ids_lengths,
mtest.labels: labels,
}))
return all_probs[:batch_size_original]
(preds, ) = session.run([mtest.preds],
tf_utils.filter_none_keys({
mtest.ids: ids,
mtest.ids_lengths: ids_lengths,
mtest.labels: labels,
}))
preds = preds[:batch_size_original]
labels = labels[:batch_size_original]
correct = (preds == labels)
incorrects = []
for i, e in enumerate(correct):
if (False in e):
incorrects.append({
'ids': ids[i],
'preds': preds[i],
'labels': labels[i]
})
if self.label_types == [0, 1, 2, 3]:
correct = np.concatenate(
(correct, np.all(correct[:, [0, 2]], axis=1)[:, np.newaxis],
np.all(correct[:, [1, 3]], axis=1)[:, np.newaxis]),
axis=1)
return correct, len(preds), preds, incorrects
def get_batch_full(self, batch, batch_size):
ids = tf_utils.make_array([item['ids'] for item in batch], length=self.memory_size, batch_size=batch_size)
ids_lengths = np.zeros(batch_size)
for i, item in enumerate(batch):
ids_lengths[i] = len(item['ids'])
labels = []
for i, item in enumerate(batch):
lab = [x.lower() for x in item['label_names']]
labels.append(lab)
return ids, ids_lengths, labels
def read_batch(self, current_bucket, current_bucket_start):
bucket = self.bucketed_train[current_bucket]
batch = bucket[current_bucket_start:current_bucket_start +
self.optim_config['batch_size']]
batch = batch + [
random.choice(bucket)
for _ in range(self.optim_config['batch_size'] - len(batch))
]
# batch x time
ids = tf_utils.make_array([item['ids'] for item in batch],
length=self.memory_size)
# batch
ids_lengths = np.array([len(item['ids']) for item in batch])
# batch x label types
labels = tf_utils.make_array([item['labels']
for item in batch])[:, self.label_types]
# batch x time x label types
expanded_labels = np.zeros(
(labels.shape[0], ids.shape[1], labels.shape[1]))
expanded_labels[:, ids_lengths - 1] = labels
return ids, ids_lengths, labels, expanded_labels
def get_batch(self, batch):
ids = tf_utils.make_array([item['ids'] for item in batch], length=self.memory_size)
ids_lengths = np.array([len(item['ids']) for item in batch]) # [batch_size(32)]
labels = np.zeros((self.config['batch_size'], 4), dtype=np.int32)
batch_mask = [np.zeros((self.config['batch_size'], self.label_size[i]), dtype=np.int32) for i in range(4)]
for i, item in enumerate(batch):
lab = [x.lower() for x in item['label_names']]
lab = [self.label2index[j][lab[j]] for j in range(4)] # for ext exp
labels[i, :] = lab
mask = self.get_mask(item['label_names'])
for j in range(4):
batch_mask[j][i, :] = mask[j]
return ids, ids_lengths, labels, batch_mask
def pred_one(self, session, mtest, batch, batch_size, get_probs=False):
batch_size_original = len(batch)
ids = tf_utils.make_array(
[item['ids'] for item in batch] + \
[batch[0]['ids'] for _ in xrange(batch_size - len(batch))],
length=self.memory_size)
ids_lengths = np.array(
[len(item['ids']) for item in batch] + \
[len(batch[0]['ids']) for _ in xrange(batch_size - len(batch))])
labels = tf_utils.make_array(
[item['labels'] for item in batch] + \
[batch[0]['labels'] for _ in xrange(batch_size - len(batch))])
labels = labels[:, list(self.label_types)]
(preds, ) = session.run([mtest.preds],
tf_utils.filter_none_keys({
mtest.ids: ids,
mtest.ids_lengths: ids_lengths,
mtest.labels: labels,
}))
preds = preds[:batch_size_original]
return preds