Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Mask vs sequence #5565

Merged
merged 8 commits into from
Apr 15, 2020
Merged

Mask vs sequence #5565

Show file tree
Hide file tree
Changes from 4 commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
8d8f421
merge conflict
evgeniiaraz Apr 2, 2020
d4564ec
small changes
evgeniiaraz Apr 2, 2020
18e4b79
small changed by comments
evgeniiaraz Apr 2, 2020
e0201f1
None return type
evgeniiaraz Apr 2, 2020
d4b7b51
Update docstring
evgeniiaraz Apr 3, 2020
ae6dd7b
sequence length as a separate method
evgeniiaraz Apr 3, 2020
2bf5f23
Merge branch 'mask-vs-sequence' of https://github.com/RasaHQ/rasa int…
evgeniiaraz Apr 3, 2020
75ac652
black
evgeniiaraz Apr 3, 2020
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
43 changes: 24 additions & 19 deletions rasa/nlu/classifiers/diet_classifier.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -87,10 +87,10 @@

TEXT_FEATURES = f"{TEXT}_features"
LABEL_FEATURES = f"{LABEL}_features"
TEXT_MASK = f"{TEXT}_mask"
LABEL_MASK = f"{LABEL}_mask"
LABEL_IDS = f"{LABEL}_ids"
TAG_IDS = "tag_ids"
TEXT_SEQ_LENGTH = f"{TEXT}_lengths"
LABEL_SEQ_LENGTH = f"{LABEL}_lengths"


class DIETClassifier(IntentClassifier, EntityExtractor):
Expand Down Expand Up @@ -484,7 +484,7 @@ def _create_label_data(
# to track correctly dynamic sequences
label_data.add_features(LABEL_IDS, [np.expand_dims(label_ids, -1)])

label_data.add_mask(LABEL_MASK, LABEL_FEATURES)
label_data.add_lengths(LABEL_SEQ_LENGTH, LABEL_FEATURES)

return label_data

Expand Down Expand Up @@ -558,8 +558,8 @@ def _create_model_data(
model_data.add_features(LABEL_IDS, [np.expand_dims(label_ids, -1)])
model_data.add_features(TAG_IDS, [tag_ids])

model_data.add_mask(TEXT_MASK, TEXT_FEATURES)
model_data.add_mask(LABEL_MASK, LABEL_FEATURES)
model_data.add_lengths(TEXT_SEQ_LENGTH, TEXT_FEATURES)
model_data.add_lengths(LABEL_SEQ_LENGTH, LABEL_FEATURES)

return model_data

Expand Down Expand Up @@ -1165,10 +1165,6 @@ def _prepare_entity_recognition_layers(self) -> None:
average="micro",
)

@staticmethod
def _get_sequence_lengths(mask: tf.Tensor) -> tf.Tensor:
return tf.cast(tf.reduce_sum(mask[:, :, 0], axis=1), tf.int32)

def _combine_sparse_dense_features(
self,
features: List[Union[np.ndarray, tf.Tensor, tf.SparseTensor]],
Expand Down Expand Up @@ -1255,10 +1251,12 @@ def _create_sequence(

def _create_all_labels(self) -> Tuple[tf.Tensor, tf.Tensor]:
all_label_ids = self.tf_label_data[LABEL_IDS][0]

label_lengths = tf.cast(self.tf_label_data[LABEL_SEQ_LENGTH][0], dtype=tf.int32)
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Should we add a helper method for this? Something like sequence_length_for? This logic is repeated all over the place.

mask_label = self._compute_mask(label_lengths)

x = self._create_bow(
self.tf_label_data[LABEL_FEATURES],
self.tf_label_data[LABEL_MASK][0],
self.label_name,
self.tf_label_data[LABEL_FEATURES], mask_label, self.label_name,
)
all_labels_embed = self._tf_layers[f"embed.{LABEL}"](x)

Expand Down Expand Up @@ -1352,13 +1350,19 @@ def _calculate_entity_loss(

return loss, f1

def _compute_mask(self, sequence_lengths: tf.Tensor) -> tf.Tensor:
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can be static.

mask = tf.sequence_mask(sequence_lengths, dtype=tf.float32)
# explicitly add last dimension to mask
# to track correctly dynamic sequences
return tf.expand_dims(mask, -1)

def batch_loss(
self, batch_in: Union[Tuple[tf.Tensor], Tuple[np.ndarray]]
) -> tf.Tensor:
tf_batch_data = self.batch_to_model_data_format(batch_in, self.data_signature)

mask_text = tf_batch_data[TEXT_MASK][0]
sequence_lengths = self._get_sequence_lengths(mask_text)
sequence_lengths = tf.cast(tf_batch_data[TEXT_SEQ_LENGTH][0], dtype=tf.int32)
mask_text = self._compute_mask(sequence_lengths)

(
text_transformed,
Expand Down Expand Up @@ -1387,11 +1391,12 @@ def batch_loss(
# get _cls_ vector for intent classification
cls = self._last_token(text_transformed, sequence_lengths)

label_lengths = tf.cast(tf_batch_data[LABEL_SEQ_LENGTH][0], dtype=tf.int32)
mask_label = self._compute_mask(label_lengths)

label_ids = tf_batch_data[LABEL_IDS][0]
label = self._create_bow(
tf_batch_data[LABEL_FEATURES],
tf_batch_data[LABEL_MASK][0],
self.label_name,
tf_batch_data[LABEL_FEATURES], mask_label, self.label_name,
)
loss, acc = self._calculate_label_loss(cls, label, label_ids)
self.intent_loss.update_state(loss)
Expand All @@ -1417,8 +1422,8 @@ def batch_predict(
batch_in, self.predict_data_signature
)

mask_text = tf_batch_data[TEXT_MASK][0]
sequence_lengths = self._get_sequence_lengths(mask_text)
sequence_lengths = tf.cast(tf_batch_data[TEXT_SEQ_LENGTH][0], dtype=tf.int32)
mask_text = self._compute_mask(sequence_lengths)

text_transformed, _, _, _ = self._create_sequence(
tf_batch_data[TEXT_FEATURES], mask_text, self.text_name
Expand Down
28 changes: 18 additions & 10 deletions rasa/nlu/selectors/response_selector.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,9 +17,9 @@
DIET,
TEXT_FEATURES,
LABEL_FEATURES,
TEXT_MASK,
LABEL_MASK,
LABEL_IDS,
TEXT_SEQ_LENGTH,
LABEL_SEQ_LENGTH,
)
from rasa.utils.tensorflow.constants import (
LABEL,
Expand Down Expand Up @@ -432,8 +432,10 @@ def _prepare_layers(self) -> None:
def _create_all_labels(self) -> Tuple[tf.Tensor, tf.Tensor]:
all_label_ids = self.tf_label_data[LABEL_IDS][0]

mask_label = self.tf_label_data[LABEL_MASK][0]
sequence_lengths_label = self._get_sequence_lengths(mask_label)
sequence_lengths_label = tf.cast(
self.tf_label_data[LABEL_SEQ_LENGTH][0], dtype=tf.int32
)
mask_label = self._compute_mask(sequence_lengths_label)

label_transformed, _, _, _ = self._create_sequence(
self.tf_label_data[LABEL_FEATURES], mask_label, self.label_name
Expand All @@ -449,8 +451,10 @@ def batch_loss(
) -> tf.Tensor:
tf_batch_data = self.batch_to_model_data_format(batch_in, self.data_signature)

mask_text = tf_batch_data[TEXT_MASK][0]
sequence_lengths_text = self._get_sequence_lengths(mask_text)
sequence_lengths_text = tf.cast(
tf_batch_data[TEXT_SEQ_LENGTH][0], dtype=tf.int32
)
mask_text = self._compute_mask(sequence_lengths_text)

(
text_transformed,
Expand All @@ -465,8 +469,10 @@ def batch_loss(
sequence_ids=True,
)

mask_label = tf_batch_data[LABEL_MASK][0]
sequence_lengths_label = self._get_sequence_lengths(mask_label)
sequence_lengths_label = tf.cast(
tf_batch_data[LABEL_SEQ_LENGTH][0], dtype=tf.int32
)
mask_label = self._compute_mask(sequence_lengths_label)

label_transformed, _, _, _ = self._create_sequence(
tf_batch_data[LABEL_FEATURES], mask_label, self.label_name
Expand Down Expand Up @@ -506,8 +512,10 @@ def batch_predict(
batch_in, self.predict_data_signature
)

mask_text = tf_batch_data[TEXT_MASK][0]
sequence_lengths_text = self._get_sequence_lengths(mask_text)
sequence_lengths_text = tf.cast(
tf_batch_data[TEXT_SEQ_LENGTH][0], dtype=tf.int32
)
mask_text = self._compute_mask(sequence_lengths_text)

text_transformed, _, _, _ = self._create_sequence(
tf_batch_data[TEXT_FEATURES], mask_text, self.text_name
Expand Down
13 changes: 6 additions & 7 deletions rasa/utils/tensorflow/model_data.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -144,20 +144,19 @@ def add_features(self, key: Text, features: List[np.ndarray]):
# update number of examples
self.num_examples = self.number_of_examples()

def add_mask(self, key: Text, from_key: Text):
"""Calculate mask for given key and put it under specified key."""

def add_lengths(self, key: Text, from_key: Text) -> None:
"""
Adds np.array of lengths of sequences to data under given key
"""
if not self.data.get(from_key):
return

self.data[key] = []

for data in self.data[from_key]:
if data.size > 0:
# explicitly add last dimension to mask
# to track correctly dynamic sequences
mask = np.array([np.ones((x.shape[0], 1)) for x in data])
self.data[key].append(mask)
lengths = np.array([x.shape[0] for x in data])
self.data[key].append(lengths)
break

def split(
Expand Down