BertMaskedLM Task Model and Preprocessor (#774)

* bert_masekd_lm init

* Merge branch 'master' into BertMaskedLM

* WIP : BERT MASKED LM

* Added Tests

* Black Formatting

* Fixed Format

* Fixed formatting

* black + lint.sh

* Reformat codew

* Updated Docstring for bert_tokenizer

* Updated masked_lm_generator.py

* fixed linting

* Changed Boolean Variables tp Numeric

* Formatted using shell/format.sh

* Updated bert_masked_lm.py

* typo fix

---------

Co-authored-by: Matt Watson <[email protected]>

keras_nlp/models/__init__.py

@@ -24,6 +24,10 @@
 from keras_nlp.models.bart.bart_tokenizer import BartTokenizer
 from keras_nlp.models.bert.bert_backbone import BertBackbone
 from keras_nlp.models.bert.bert_classifier import BertClassifier
+from keras_nlp.models.bert.bert_masked_lm import BertMaskedLM
+from keras_nlp.models.bert.bert_masked_lm_preprocessor import (
+    BertMaskedLMPreprocessor,
+)
 from keras_nlp.models.bert.bert_preprocessor import BertPreprocessor
 from keras_nlp.models.bert.bert_tokenizer import BertTokenizer
 from keras_nlp.models.deberta_v3.deberta_v3_backbone import DebertaV3Backbone

keras_nlp/models/bert/bert_masked_lm.py

@@ -0,0 +1,152 @@
+# Copyright 2022 The KerasNLP Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""BERT masked LM model."""
+
+import copy
+
+from tensorflow import keras
+
+from keras_nlp.layers.masked_lm_head import MaskedLMHead
+from keras_nlp.models.bert.bert_backbone import BertBackbone
+from keras_nlp.models.bert.bert_backbone import bert_kernel_initializer
+from keras_nlp.models.bert.bert_masked_lm_preprocessor import (
+    BertMaskedLMPreprocessor,
+)
+from keras_nlp.models.bert.bert_presets import backbone_presets
+from keras_nlp.models.task import Task
+from keras_nlp.utils.python_utils import classproperty
+
+
+@keras.utils.register_keras_serializable(package="keras_nlp")
+class BertMaskedLM(Task):
+    """An end-to-end BERT model for the masked language modeling task.
+
+    This model will train BERT on a masked language modeling task.
+    The model will predict labels for a number of masked tokens in the
+    input data. For usage of this model with pre-trained weights, see the
+    `from_preset()` method.
+
+    This model can optionally be configured with a `preprocessor` layer, in
+    which case inputs can be raw string features during `fit()`, `predict()`,
+    and `evaluate()`. Inputs will be tokenized and dynamically masked during
+    training and evaluation. This is done by default when creating the model
+    with `from_preset()`.
+
+    Disclaimer: Pre-trained models are provided on an "as is" basis, without
+    warranties or conditions of any kind.
+
+    Args:
+        backbone: A `keras_nlp.models.BertBackbone` instance.
+        preprocessor: A `keras_nlp.models.BertMaskedLMPreprocessor` or
+            `None`. If `None`, this model will not apply preprocessing, and
+            inputs should be preprocessed before calling the model.
+
+    Example usage:
+
+    Raw string inputs and pretrained backbone.
+    ```python
+    # Create a dataset with raw string features. Labels are inferred.
+    features = ["The quick brown fox jumped.", "I forgot my homework."]
+
+    # Create a BertMaskedLM with a pretrained backbone and further train
+    # on an MLM task.
+    masked_lm = keras_nlp.models.BertMaskedLM.from_preset(
+        "bert_base_en",
+    )
+    masked_lm.compile(
+        loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
+    )
+    masked_lm.fit(x=features, batch_size=2)
+    ```
+
+    Preprocessed inputs and custom backbone.
+    ```python
+    # Create a preprocessed dataset where 0 is the mask token.
+    preprocessed_features = {
+        "token_ids": tf.constant(
+            [[1, 2, 0, 4, 0, 6, 7, 8]] * 2, shape=(2, 8)
+        ),
+        "padding_mask": tf.constant(
+            [[1, 1, 1, 1, 1, 1, 1, 1]] * 2, shape=(2, 8)
+        ),
+        "mask_positions": tf.constant([[2, 4]] * 2, shape=(2, 2)),
+        "segment_ids": tf.constant([[0, 0, 0, 0, 0, 0, 0, 0]] * 2, shape=(2, 8))
+    }
+    # Labels are the original masked values.
+    labels = [[3, 5]] * 2
+
+    # Randomly initialize a BERT encoder
+    backbone = keras_nlp.models.BertBackbone(
+        vocabulary_size=50265,
+        num_layers=12,
+        num_heads=12,
+        hidden_dim=768,
+        intermediate_dim=3072,
+        max_sequence_length=12
+    )
+    # Create a BERT masked LM model and fit the data.
+    masked_lm = keras_nlp.models.BertMaskedLM(
+        backbone,
+        preprocessor=None,
+    )
+    masked_lm.compile(
+        loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
+    )
+    masked_lm.fit(x=preprocessed_features, y=labels, batch_size=2)
+    ```
+    """
+
+    def __init__(
+        self,
+        backbone,
+        preprocessor=None,
+        **kwargs,
+    ):
+        inputs = {
+            **backbone.input,
+            "mask_positions": keras.Input(
+                shape=(None,), dtype="int32", name="mask_positions"
+            ),
+        }
+        backbone_outputs = backbone(backbone.input)
+        outputs = MaskedLMHead(
+            vocabulary_size=backbone.vocabulary_size,
+            embedding_weights=backbone.token_embedding.embeddings,
+            intermediate_activation="gelu",
+            kernel_initializer=bert_kernel_initializer(),
+            name="mlm_head",
+        )(backbone_outputs["sequence_output"], inputs["mask_positions"])
+
+        # Instantiate using Functional API Model constructor
+        super().__init__(
+            inputs=inputs,
+            outputs=outputs,
+            include_preprocessing=preprocessor is not None,
+            **kwargs,
+        )
+        # All references to `self` below this line
+        self.backbone = backbone
+        self.preprocessor = preprocessor
+
+    @classproperty
+    def backbone_cls(cls):
+        return BertBackbone
+
+    @classproperty
+    def preprocessor_cls(cls):
+        return BertMaskedLMPreprocessor
+
+    @classproperty
+    def presets(cls):
+        return copy.deepcopy(backbone_presets)

keras_nlp/models/bert/bert_masked_lm_preprocessor.py

@@ -0,0 +1,138 @@
+# Copyright 2022 The KerasNLP Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""BERT masked language model preprocessor layer."""
+
+from absl import logging
+from tensorflow import keras
+
+from keras_nlp.layers.masked_lm_mask_generator import MaskedLMMaskGenerator
+from keras_nlp.models.bert.bert_preprocessor import BertPreprocessor
+from keras_nlp.utils.keras_utils import pack_x_y_sample_weight
+
+
+@keras.utils.register_keras_serializable(package="keras_nlp")
+class BertMaskedLMPreprocessor(BertPreprocessor):
+    """BERT preprocessing for the masked language modeling task.
+    This preprocessing layer will prepare inputs for a masked language modeling
+    task. It is primarily intended for use with the
+    `keras_nlp.models.BertMaskedLM` task model. Preprocessing will occur in
+    multiple steps.
+    - Tokenize any number of input segments using the `tokenizer`.
+    - Pack the inputs together using a `keras_nlp.layers.MultiSegmentPacker`.
+       with the appropriate `"[CLS]"`, `"[SEP]"`, `"[SEP]"`, `"[SEP]"` and `"[PAD]"` tokens.
+    - Randomly select non-special tokens to mask, controlled by
+      `mask_selection_rate`.
+    - Construct a `(x, y, sample_weight)` tuple suitable for training with a
+      `keras_nlp.models.BertMaskedLM` task model.
+    Examples:
+    ```python
+    # Load the preprocessor from a preset.
+    preprocessor = keras_nlp.models.BertMaskedLMPreprocessor.from_preset(
+        "bert_base_en"
+    )
+    # Tokenize and mask a single sentence.
+    sentence = tf.constant("The quick brown fox jumped.")
+    preprocessor(sentence)
+    # Tokenize and mask a batch of sentences.
+    sentences = tf.constant(
+        ["The quick brown fox jumped.", "Call me Ishmael."]
+    )
+    preprocessor(sentences)
+    # Tokenize and mask a dataset of sentences.
+    features = tf.constant(
+        ["The quick brown fox jumped.", "Call me Ishmael."]
+    )
+    ds = tf.data.Dataset.from_tensor_slices((features))
+    ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
+    # Alternatively, you can create a preprocessor from your own vocabulary.
+    # The usage is exactly the same as above.
+    vocab = ["[PAD]", "[UNK]", "[CLS]", "[SEP]", "[MASK]"]
+    vocab += ["THE", "QUICK", "BROWN", "FOX"]
+    vocab += ["Call", "me", "Ishmael"]
+    tokenizer = keras_nlp.models.BertTokenizer(vocabulary=vocab)
+    preprocessor = keras_nlp.models.BertMaskedLMPreprocessor(tokenizer)
+    ```
+    """
+
+    def __init__(
+        self,
+        tokenizer,
+        sequence_length=512,
+        truncate="round_robin",
+        mask_selection_rate=0.15,
+        mask_selection_length=96,
+        mask_token_rate=0.8,
+        random_token_rate=0.1,
+        **kwargs,
+    ):
+        super().__init__(
+            tokenizer,
+            sequence_length=sequence_length,
+            truncate=truncate,
+            **kwargs,
+        )
+
+        self.masker = MaskedLMMaskGenerator(
+            mask_selection_rate=mask_selection_rate,
+            mask_selection_length=mask_selection_length,
+            mask_token_rate=mask_token_rate,
+            random_token_rate=random_token_rate,
+            vocabulary_size=tokenizer.vocabulary_size(),
+            mask_token_id=tokenizer.mask_token_id,
+            unselectable_token_ids=[
+                tokenizer.cls_token_id,
+                tokenizer.sep_token_id,
+                tokenizer.pad_token_id,
+            ],
+        )
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "mask_selection_rate": self.masker.mask_selection_rate,
+                "mask_selection_length": self.masker.mask_selection_length,
+                "mask_token_rate": self.masker.mask_token_rate,
+                "random_token_rate": self.masker.random_token_rate,
+            }
+        )
+        return config
+
+    def call(self, x, y=None, sample_weight=None):
+        if y is not None or sample_weight is not None:
+            logging.warning(
+                f"{self.__class__.__name__} generates `y` and `sample_weight` "
+                "based on your input data, but your data already contains `y` "
+                "or `sample_weight`. Your `y` and `sample_weight` will be "
+                "ignored."
+            )
+
+        x = super().call(x)
+
+        token_ids, padding_mask, segment_ids = (
+            x["token_ids"],
+            x["padding_mask"],
+            x["segment_ids"],
+        )
+        masker_outputs = self.masker(token_ids)
+        x = {
+            "token_ids": masker_outputs["token_ids"],
+            "padding_mask": padding_mask,
+            "segment_ids": segment_ids,
+            "mask_positions": masker_outputs["mask_positions"],
+        }
+        y = masker_outputs["mask_ids"]
+        sample_weight = masker_outputs["mask_weights"]
+        return pack_x_y_sample_weight(x, y, sample_weight)