keras-hub-nightly 0.15.0.dev20240823171555__py3-none-any.whl

keras_hub/src/models/distil_bert/distil_bert_preprocessor.py

@@ -0,0 +1,175 @@
+# Copyright 2024 The KerasHub 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.
+
+
+import keras
+
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.layers.preprocessing.multi_segment_packer import (
+    MultiSegmentPacker,
+)
+from keras_hub.src.models.distil_bert.distil_bert_tokenizer import (
+    DistilBertTokenizer,
+)
+from keras_hub.src.models.preprocessor import Preprocessor
+from keras_hub.src.utils.keras_utils import (
+    convert_inputs_to_list_of_tensor_segments,
+)
+
+
+@keras_hub_export("keras_hub.models.DistilBertPreprocessor")
+class DistilBertPreprocessor(Preprocessor):
+    """A DistilBERT preprocessing layer which tokenizes and packs inputs.
+
+    This preprocessing layer will do three things:
+
+     1. Tokenize any number of input segments using the `tokenizer`.
+     2. Pack the inputs together using a `keras_hub.layers.MultiSegmentPacker`.
+       with the appropriate `"[CLS]"`, `"[SEP]"` and `"[PAD]"` tokens.
+     3. Construct a dictionary of with keys `"token_ids"` and `"padding_mask"`,
+       that can be passed directly to a DistilBERT model.
+
+    This layer can be used directly with `tf.data.Dataset.map` to preprocess
+    string data in the `(x, y, sample_weight)` format used by
+    `keras.Model.fit`.
+
+    Args:
+        tokenizer: A `keras_hub.models.DistilBertTokenizer` instance.
+        sequence_length: The length of the packed inputs.
+        truncate: string. The algorithm to truncate a list of batched segments
+            to fit within `sequence_length`. The value can be either
+            `round_robin` or `waterfall`:
+                - `"round_robin"`: Available space is assigned one token at a
+                    time in a round-robin fashion to the inputs that still need
+                    some, until the limit is reached.
+                - `"waterfall"`: The allocation of the budget is done using a
+                    "waterfall" algorithm that allocates quota in a
+                    left-to-right manner and fills up the buckets until we run
+                    out of budget. It supports an arbitrary number of segments.
+
+    Call arguments:
+        x: A tensor of single string sequences, or a tuple of multiple
+            tensor sequences to be packed together. Inputs may be batched or
+            unbatched. For single sequences, raw python inputs will be converted
+            to tensors. For multiple sequences, pass tensors directly.
+        y: Any label data. Will be passed through unaltered.
+        sample_weight: Any label weight data. Will be passed through unaltered.
+
+    Examples:
+
+    Directly calling the layer on data.
+    ```python
+    preprocessor = keras_hub.models.DistilBertPreprocessor.from_preset(
+        "distil_bert_base_en_uncased"
+    )
+    preprocessor(["The quick brown fox jumped.", "Call me Ishmael."])
+
+    # Custom vocabulary.
+    vocab = ["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"]
+    vocab += ["The", "quick", "brown", "fox", "jumped", "."]
+    tokenizer = keras_hub.models.DistilBertTokenizer(vocabulary=vocab)
+    preprocessor = keras_hub.models.DistilBertPreprocessor(tokenizer)
+    preprocessor("The quick brown fox jumped.")
+    ```
+
+    Mapping with `tf.data.Dataset`.
+    ```python
+    preprocessor = keras_hub.models.DistilBertPreprocessor.from_preset(
+        "distil_bert_base_en_uncased"
+    )
+
+    first = tf.constant(["The quick brown fox jumped.", "Call me Ishmael."])
+    second = tf.constant(["The fox tripped.", "Oh look, a whale."])
+    label = tf.constant([1, 1])
+    # Map labeled single sentences.
+    ds = tf.data.Dataset.from_tensor_slices((first, label))
+    ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
+
+
+    # Map unlabeled single sentences.
+    ds = tf.data.Dataset.from_tensor_slices(first)
+    ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
+
+    # Map labeled sentence pairs.
+    ds = tf.data.Dataset.from_tensor_slices(((first, second), label))
+    ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
+    # Map unlabeled sentence pairs.
+    ds = tf.data.Dataset.from_tensor_slices((first, second))
+
+    # Watch out for tf.data's default unpacking of tuples here!
+    # Best to invoke the `preprocessor` directly in this case.
+    ds = ds.map(
+        lambda first, second: preprocessor(x=(first, second)),
+        num_parallel_calls=tf.data.AUTOTUNE,
+    )
+    ```
+    """
+
+    tokenizer_cls = DistilBertTokenizer
+
+    def __init__(
+        self,
+        tokenizer,
+        sequence_length=512,
+        truncate="round_robin",
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.tokenizer = tokenizer
+        self.packer = None
+        self.sequence_length = sequence_length
+        self.truncate = truncate
+
+    def build(self, input_shape):
+        super().build(input_shape)
+        # Defer masker creation to `build()` so that we can be sure tokenizer
+        # assets have loaded when restoring a saved model.
+        self.packer = MultiSegmentPacker(
+            start_value=self.tokenizer.cls_token_id,
+            end_value=self.tokenizer.sep_token_id,
+            pad_value=self.tokenizer.pad_token_id,
+            truncate=self.truncate,
+            sequence_length=self.sequence_length,
+        )
+
+    def call(self, x, y=None, sample_weight=None):
+        x = convert_inputs_to_list_of_tensor_segments(x)
+        x = [self.tokenizer(segment) for segment in x]
+        token_ids, _ = self.packer(x)
+        x = {
+            "token_ids": token_ids,
+            "padding_mask": token_ids != self.tokenizer.pad_token_id,
+        }
+        return keras.utils.pack_x_y_sample_weight(x, y, sample_weight)
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "sequence_length": self.sequence_length,
+                "truncate": self.truncate,
+            }
+        )
+        return config
+
+    @property
+    def sequence_length(self):
+        """The padded length of model input sequences."""
+        return self._sequence_length
+
+    @sequence_length.setter
+    def sequence_length(self, value):
+        self._sequence_length = value
+        if self.packer is not None:
+            self.packer.sequence_length = value

keras_hub/src/models/distil_bert/distil_bert_presets.py

@@ -0,0 +1,57 @@
+# Copyright 2024 The KerasHub 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.
+"""DistilBERT model preset configurations."""
+
+backbone_presets = {
+    "distil_bert_base_en_uncased": {
+        "metadata": {
+            "description": (
+                "6-layer DistilBERT model where all input is lowercased. "
+                "Trained on English Wikipedia + BooksCorpus using BERT as the "
+                "teacher model."
+            ),
+            "params": 66362880,
+            "official_name": "DistilBERT",
+            "path": "distil_bert",
+            "model_card": "https://huggingface.co/distilbert-base-uncased",
+        },
+        "kaggle_handle": "kaggle://keras/distil_bert/keras/distil_bert_base_en_uncased/2",
+    },
+    "distil_bert_base_en": {
+        "metadata": {
+            "description": (
+                "6-layer DistilBERT model where case is maintained. "
+                "Trained on English Wikipedia + BooksCorpus using BERT as the "
+                "teacher model."
+            ),
+            "params": 65190912,
+            "official_name": "DistilBERT",
+            "path": "distil_bert",
+            "model_card": "https://huggingface.co/distilbert-base-cased",
+        },
+        "kaggle_handle": "kaggle://keras/distil_bert/keras/distil_bert_base_en/2",
+    },
+    "distil_bert_base_multi": {
+        "metadata": {
+            "description": (
+                "6-layer DistilBERT model where case is maintained. Trained on Wikipedias of 104 languages"
+            ),
+            "params": 134734080,
+            "official_name": "DistilBERT",
+            "path": "distil_bert",
+            "model_card": "https://huggingface.co/distilbert-base-multilingual-cased",
+        },
+        "kaggle_handle": "kaggle://keras/distil_bert/keras/distil_bert_base_multi/2",
+    },
+}

keras_hub/src/models/distil_bert/distil_bert_tokenizer.py

@@ -0,0 +1,114 @@
+# Copyright 2024 The KerasHub 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.
+
+
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.tokenizers.word_piece_tokenizer import WordPieceTokenizer
+
+
+@keras_hub_export("keras_hub.models.DistilBertTokenizer")
+class DistilBertTokenizer(WordPieceTokenizer):
+    """A DistilBERT tokenizer using WordPiece subword segmentation.
+
+    This tokenizer class will tokenize raw strings into integer sequences and
+    is based on `keras_hub.tokenizers.WordPieceTokenizer`. Unlike the
+    underlying tokenizer, it will check for all special tokens needed by DistilBERT
+    models and provides a `from_preset()` method to automatically download
+    a matching vocabulary for a DistilBERT preset.
+
+    This tokenizer does not provide truncation or padding of inputs. It can be
+    combined with a `keras_hub.models.DistilBertPreprocessor` layer for input packing.
+
+    If input is a batch of strings (rank > 0), the layer will output a
+    `tf.RaggedTensor` where the last dimension of the output is ragged.
+
+    If input is a scalar string (rank == 0), the layer will output a dense
+    `tf.Tensor` with static shape `[None]`.
+
+    Args:
+        vocabulary: A list of strings or a string filename path. If
+            passing a list, each element of the list should be a single word
+            piece token string. If passing a filename, the file should be a
+            plain text file containing a single word piece token per line.
+        lowercase: If `True`, the input text will be first lowered before
+            tokenization.
+        special_tokens_in_strings: bool. A bool to indicate if the tokenizer
+            should expect special tokens in input strings that should be
+            tokenized and mapped correctly to their ids. Defaults to False.
+
+    Examples:
+
+    ```python
+    # Unbatched input.
+    tokenizer = keras_hub.models.DistilBertTokenizer.from_preset(
+        "distil_bert_base_en_uncased",
+    )
+    tokenizer("The quick brown fox jumped.")
+
+    # Batched input.
+    tokenizer(["The quick brown fox jumped.", "The fox slept."])
+
+    # Detokenization.
+    tokenizer.detokenize(tokenizer("The quick brown fox jumped."))
+
+    # Custom vocabulary.
+    vocab = ["[UNK]", "[CLS]", "[SEP]", "[PAD]", "[MASK]"]
+    vocab += ["The", "quick", "brown", "fox", "jumped", "."]
+    tokenizer = keras_hub.models.DistilBertTokenizer(vocabulary=vocab)
+    tokenizer("The quick brown fox jumped.")
+    ```
+    """
+
+    def __init__(
+        self,
+        vocabulary,
+        lowercase=False,
+        special_tokens_in_strings=False,
+        **kwargs,
+    ):
+        self.cls_token = "[CLS]"
+        self.sep_token = "[SEP]"
+        self.pad_token = "[PAD]"
+        self.mask_token = "[MASK]"
+        super().__init__(
+            vocabulary=vocabulary,
+            lowercase=lowercase,
+            special_tokens=[
+                self.cls_token,
+                self.sep_token,
+                self.pad_token,
+                self.mask_token,
+            ],
+            special_tokens_in_strings=special_tokens_in_strings,
+            **kwargs,
+        )
+
+    def set_vocabulary(self, vocabulary):
+        super().set_vocabulary(vocabulary)
+
+        if vocabulary is not None:
+            self.cls_token_id = self.token_to_id(self.cls_token)
+            self.sep_token_id = self.token_to_id(self.sep_token)
+            self.pad_token_id = self.token_to_id(self.pad_token)
+            self.mask_token_id = self.token_to_id(self.mask_token)
+        else:
+            self.cls_token_id = None
+            self.sep_token_id = None
+            self.pad_token_id = None
+            self.mask_token_id = None
+
+    def get_config(self):
+        config = super().get_config()
+        del config["special_tokens"]  # Not configurable; set in __init__.
+        return config

keras_hub/src/models/electra/__init__.py

@@ -0,0 +1,20 @@
+# Copyright 2024 The KerasHub 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.
+
+from keras_hub.src.models.electra.electra_backbone import ElectraBackbone
+from keras_hub.src.models.electra.electra_presets import backbone_presets
+from keras_hub.src.models.electra.electra_tokenizer import ElectraTokenizer
+from keras_hub.src.utils.preset_utils import register_presets
+
+register_presets(backbone_presets, (ElectraBackbone, ElectraTokenizer))

keras_hub/src/models/electra/electra_backbone.py

@@ -0,0 +1,247 @@
+# Copyright 2024 The KerasHub 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.
+
+import keras
+
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.layers.modeling.position_embedding import PositionEmbedding
+from keras_hub.src.layers.modeling.reversible_embedding import (
+    ReversibleEmbedding,
+)
+from keras_hub.src.layers.modeling.transformer_encoder import TransformerEncoder
+from keras_hub.src.models.backbone import Backbone
+from keras_hub.src.utils.keras_utils import gelu_approximate
+
+
+def electra_kernel_initializer(stddev=0.02):
+    return keras.initializers.TruncatedNormal(stddev=stddev)
+
+
+@keras_hub_export("keras_hub.models.ElectraBackbone")
+class ElectraBackbone(Backbone):
+    """A Electra encoder network.
+
+    This network implements a bidirectional Transformer-based encoder as
+    described in ["Electra: Pre-training Text Encoders as Discriminators Rather
+    Than Generators"](https://arxiv.org/abs/2003.10555). It includes the
+    embedding lookups and transformer layers, but not the masked language model
+    or classification task networks.
+
+    The default constructor gives a fully customizable, randomly initialized
+    ELECTRA encoder with any number of layers, heads, and embedding
+    dimensions. To load preset architectures and weights, use the
+    `from_preset()` constructor.
+
+    Disclaimer: Pre-trained models are provided on an "as is" basis, without
+    warranties or conditions of any kind. The underlying model is provided by a
+    third party and subject to a separate license, available
+    [here](https://huggingface.co/docs/transformers/model_doc/electra#overview).
+
+    Args:
+        vocabulary_size: int. The size of the token vocabulary.
+        num_layers: int. The number of transformer layers.
+        num_heads: int. The number of attention heads for each transformer.
+            The hidden size must be divisible by the number of attention heads.
+        hidden_dim: int. The size of the transformer encoding and pooler layers.
+        embedding_dim: int. The size of the token embeddings.
+        intermediate_dim: int. The output dimension of the first Dense layer in
+            a two-layer feedforward network for each transformer.
+        dropout: float. Dropout probability for the Transformer encoder.
+        max_sequence_length: int. The maximum sequence length that this encoder
+            can consume. If None, `max_sequence_length` uses the value from
+            sequence length. This determines the variable shape for positional
+            embeddings.
+        dtype: string or `keras.mixed_precision.DTypePolicy`. The dtype to use
+            for model computations and weights. Note that some computations,
+            such as softmax and layer normalization, will always be done at
+            float32 precision regardless of dtype.
+
+    Example:
+        ```python
+        input_data = {
+        "token_ids": np.ones(shape=(1, 12), dtype="int32"),
+        "segment_ids": np.array([[0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0]]),
+        "padding_mask": np.array([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0]]),
+        }
+
+        # Pre-trained ELECTRA encoder.
+        model = keras_hub.models.ElectraBackbone.from_preset(
+            "electra_base_discriminator_en"
+        )
+        model(input_data)
+
+        # Randomly initialized Electra encoder
+        backbone = keras_hub.models.ElectraBackbone(
+            vocabulary_size=1000,
+            num_layers=2,
+            num_heads=2,
+            hidden_dim=32,
+            intermediate_dim=64,
+            dropout=0.1,
+            max_sequence_length=512,
+            )
+        # Returns sequence and pooled outputs.
+        sequence_output, pooled_output = backbone(input_data)
+        ```
+    """
+
+    def __init__(
+        self,
+        vocab_size,
+        num_layers,
+        num_heads,
+        hidden_dim,
+        embedding_dim,
+        intermediate_dim,
+        dropout=0.1,
+        max_sequence_length=512,
+        num_segments=2,
+        dtype=None,
+        **kwargs,
+    ):
+        # === Layers ===
+        self.token_embedding = ReversibleEmbedding(
+            input_dim=vocab_size,
+            output_dim=embedding_dim,
+            embeddings_initializer=electra_kernel_initializer(),
+            dtype=dtype,
+            name="token_embedding",
+        )
+        self.position_embedding = PositionEmbedding(
+            initializer=electra_kernel_initializer(),
+            sequence_length=max_sequence_length,
+            dtype=dtype,
+            name="position_embedding",
+        )
+        self.segment_embedding = keras.layers.Embedding(
+            input_dim=num_segments,
+            output_dim=embedding_dim,
+            embeddings_initializer=electra_kernel_initializer(),
+            dtype=dtype,
+            name="segment_embedding",
+        )
+        self.embeddings_add = keras.layers.Add(
+            dtype=dtype,
+            name="embeddings_add",
+        )
+        self.embeddings_layer_norm = keras.layers.LayerNormalization(
+            axis=-1,
+            epsilon=1e-12,
+            dtype=dtype,
+            name="embeddings_layer_norm",
+        )
+        self.embeddings_dropout = keras.layers.Dropout(
+            dropout,
+            dtype=dtype,
+            name="embeddings_dropout",
+        )
+        if hidden_dim != embedding_dim:
+            self.embeddings_projection = keras.layers.Dense(
+                hidden_dim,
+                kernel_initializer=electra_kernel_initializer(),
+                dtype=dtype,
+                name="embeddings_projection",
+            )
+        self.transformer_layers = []
+        for i in range(num_layers):
+            layer = TransformerEncoder(
+                num_heads=num_heads,
+                intermediate_dim=intermediate_dim,
+                activation=gelu_approximate,
+                dropout=dropout,
+                layer_norm_epsilon=1e-12,
+                kernel_initializer=electra_kernel_initializer(),
+                dtype=dtype,
+                name=f"transformer_layer_{i}",
+            )
+            self.transformer_layers.append(layer)
+        self.pooled_dense = keras.layers.Dense(
+            hidden_dim,
+            kernel_initializer=electra_kernel_initializer(),
+            activation="tanh",
+            dtype=dtype,
+            name="pooled_dense",
+        )
+
+        # === Functional Model ===
+        token_id_input = keras.Input(
+            shape=(None,), dtype="int32", name="token_ids"
+        )
+        segment_id_input = keras.Input(
+            shape=(None,), dtype="int32", name="segment_ids"
+        )
+        padding_mask_input = keras.Input(
+            shape=(None,), dtype="int32", name="padding_mask"
+        )
+        # Embed tokens, positions, and segment ids.
+        tokens = self.token_embedding(token_id_input)
+        positions = self.position_embedding(tokens)
+        segments = self.segment_embedding(segment_id_input)
+        # Add all embeddings together.
+        x = self.embeddings_add((tokens, positions, segments))
+        x = self.embeddings_layer_norm(x)
+        x = self.embeddings_dropout(x)
+        if hidden_dim != embedding_dim:
+            x = self.embeddings_projection(x)
+        # Apply successive transformer encoder blocks.
+        for transformer_layer in self.transformer_layers:
+            x = transformer_layer(x, padding_mask=padding_mask_input)
+        # Index of classification token in the vocabulary
+        cls_token_index = 0
+        sequence_output = x
+        # Construct the two ELECTRA outputs. The pooled output is a dense layer on
+        # top of the [CLS] token.
+        pooled_output = self.pooled_dense(x[:, cls_token_index, :])
+        super().__init__(
+            inputs={
+                "token_ids": token_id_input,
+                "segment_ids": segment_id_input,
+                "padding_mask": padding_mask_input,
+            },
+            outputs={
+                "sequence_output": sequence_output,
+                "pooled_output": pooled_output,
+            },
+            dtype=dtype,
+            **kwargs,
+        )
+
+        # === Config ===
+        self.vocab_size = vocab_size
+        self.num_layers = num_layers
+        self.num_heads = num_heads
+        self.hidden_dim = hidden_dim
+        self.embedding_dim = embedding_dim
+        self.intermediate_dim = intermediate_dim
+        self.dropout = dropout
+        self.max_sequence_length = max_sequence_length
+        self.num_segments = num_segments
+        self.cls_token_index = cls_token_index
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "vocab_size": self.vocab_size,
+                "num_layers": self.num_layers,
+                "num_heads": self.num_heads,
+                "hidden_dim": self.hidden_dim,
+                "embedding_dim": self.embedding_dim,
+                "intermediate_dim": self.intermediate_dim,
+                "dropout": self.dropout,
+                "max_sequence_length": self.max_sequence_length,
+                "num_segments": self.num_segments,
+            }
+        )
+        return config