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

Jump to bottom

Update MaskedLMHead to support dtype=bfloat16/float16/float64 #1197

Draft
wants to merge 2 commits into
base: master
Choose a base branch
from
Draft

Update MaskedLMHead to support dtype=bfloat16/float16/float64 #1197

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
50697e5
Update MaskedLMHead to support dtype=bfloat16/float16/float64.
g-dspencer Aug 4, 2023
ff0fede
update
g-dspencer Aug 4, 2023
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
2 changes: 2 additions & 0 deletions keras_nlp/layers/modeling/masked_lm_head.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -153,9 +153,11 @@ def build(self, inputs_shape, masked_positions_shape=None):
activation=self.intermediate_activation,
Copy link
Member

@mattdangerw mattdangerw Aug 8, 2023 •
edited
Loading

Choose a reason for hiding this comment

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

It seems like we should really have this for all our our "composite" layers in KerasNLP, right?

  • token and position embedding
  • transformer decoder
  • transformer encoder
  • cached multi head attention
  • f net encoder

Are you interested in following up for other layers? (same PR or split PRs fine!)

kernel_initializer=self.kernel_initializer,
bias_initializer=self.bias_initializer,
dtype=self._dtype_policy,
Copy link
Member

Choose a reason for hiding this comment

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

It looks like self.dtype_policy should just work. Can we do that?

https://github.com/keras-team/keras/blob/b3ffea6602dbbb481e82312baa24fe657de83e11/keras/engine/base_layer.py#L2213C9-L2218

)
self._layer_norm = keras.layers.LayerNormalization(
epsilon=self.layer_norm_epsilon,
dtype=self._dtype_policy,
)
if masked_positions_shape:
gather_length = masked_positions_shape[1]
Expand Down
53 changes: 53 additions & 0 deletions keras_nlp/layers/modeling/masked_lm_head_test.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15,6 +15,9 @@

import os

import tensorflow as tf
from absl.testing import parameterized

from keras_nlp.backend import keras
from keras_nlp.backend import ops
from keras_nlp.layers.modeling import masked_lm_head
Expand All @@ -36,6 +39,30 @@ def test_valid_call(self):
position_data = ops.random.randint(minval=0, maxval=10, shape=(4, 5))
model((token_data, position_data))

@parameterized.named_parameters(
("bfloat16", tf.bfloat16),
Copy link
Member

Choose a reason for hiding this comment

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

because we now run our testing suite with jax/torch/tf with keras-core, we are generally just referring to these by string name, e.g. "float16" instead of tf.float16.

Does anything break if we switch to that?

("float16", tf.float16),
("float32", tf.float32),
("float64", tf.float64),
)
def test_valid_call_with_dtype(self, dtype):
head = masked_lm_head.MaskedLMHead(
vocabulary_size=100,
activation="softmax",
dtype=dtype,
)
encoded_tokens = keras.Input(shape=(10, 16))
positions = keras.Input(shape=(5,), dtype="int32")
outputs = head(encoded_tokens, masked_positions=positions)
Copy link
Member

Choose a reason for hiding this comment

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

This might need a rebase over master. This should be mask_positions now. This is causing a lot of test failures.

model = keras.Model((encoded_tokens, positions), outputs)

token_data = ops.random.uniform(shape=(4, 10, 16))
position_data = ops.random.randint(minval=0, maxval=10, shape=(4, 5))
model((token_data, position_data))

for w in head.weights:
self.assertEqual(w.dtype, dtype, "Wrong type: " + w.name)

def test_valid_call_with_embedding_weights(self):
embedding = keras.layers.Embedding(100, 16)
embedding.build((4, 10))
Expand Down Expand Up @@ -119,6 +146,32 @@ def test_one_train_step(self):
loss = model.train_on_batch(x=(token_data, position_data), y=label_data)
self.assertGreater(loss, 0)

@parameterized.named_parameters(
Copy link
Member

Choose a reason for hiding this comment

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

I would kill this test. Compiling a real loss function can make for slower tests, and with the parameterized testing this could slow down our suite.

("bfloat16", tf.bfloat16),
("float16", tf.float16),
("float32", tf.float32),
("float64", tf.float64),
)
def test_one_train_step_with_dtype(self, dtype):
head = masked_lm_head.MaskedLMHead(
vocabulary_size=100,
dtype=dtype,
)
encoded_tokens = keras.Input(shape=(10, 16))
positions = keras.Input(shape=(5,), dtype="int32")
outputs = head(encoded_tokens, masked_positions=positions)
model = keras.Model((encoded_tokens, positions), outputs)

token_data = ops.random.uniform(shape=(4, 10, 16))
position_data = ops.random.randint(minval=0, maxval=10, shape=(4, 5))
label_data = ops.random.randint(minval=0, maxval=2, shape=(4, 5, 1))

loss = keras.losses.SparseCategoricalCrossentropy(from_logits=False)
optimizer = keras.optimizers.Adam()
model.compile(loss=loss, optimizer=optimizer)
loss = model.train_on_batch(x=(token_data, position_data), y=label_data)
self.assertGreater(loss, 0)

def test_saved_model(self):
head = masked_lm_head.MaskedLMHead(
vocabulary_size=100,
Expand Down
Loading