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

keras_hub/src/models/bert/bert_presets.py

@@ -0,0 +1,147 @@
+# 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.
+"""BERT model preset configurations."""
+
+backbone_presets = {
+    "bert_tiny_en_uncased": {
+        "metadata": {
+            "description": (
+                "2-layer BERT model where all input is lowercased. "
+                "Trained on English Wikipedia + BooksCorpus."
+            ),
+            "params": 4385920,
+            "official_name": "BERT",
+            "path": "bert",
+            "model_card": "https://github.com/google-research/bert/blob/master/README.md",
+        },
+        "kaggle_handle": "kaggle://keras/bert/keras/bert_tiny_en_uncased/2",
+    },
+    "bert_small_en_uncased": {
+        "metadata": {
+            "description": (
+                "4-layer BERT model where all input is lowercased. "
+                "Trained on English Wikipedia + BooksCorpus."
+            ),
+            "params": 28763648,
+            "official_name": "BERT",
+            "path": "bert",
+            "model_card": "https://github.com/google-research/bert/blob/master/README.md",
+        },
+        "kaggle_handle": "kaggle://keras/bert/keras/bert_small_en_uncased/2",
+    },
+    "bert_medium_en_uncased": {
+        "metadata": {
+            "description": (
+                "8-layer BERT model where all input is lowercased. "
+                "Trained on English Wikipedia + BooksCorpus."
+            ),
+            "params": 41373184,
+            "official_name": "BERT",
+            "path": "bert",
+            "model_card": "https://github.com/google-research/bert/blob/master/README.md",
+        },
+        "kaggle_handle": "kaggle://keras/bert/keras/bert_medium_en_uncased/2",
+    },
+    "bert_base_en_uncased": {
+        "metadata": {
+            "description": (
+                "12-layer BERT model where all input is lowercased. "
+                "Trained on English Wikipedia + BooksCorpus."
+            ),
+            "params": 109482240,
+            "official_name": "BERT",
+            "path": "bert",
+            "model_card": "https://github.com/google-research/bert/blob/master/README.md",
+        },
+        "kaggle_handle": "kaggle://keras/bert/keras/bert_base_en_uncased/2",
+    },
+    "bert_base_en": {
+        "metadata": {
+            "description": (
+                "12-layer BERT model where case is maintained. "
+                "Trained on English Wikipedia + BooksCorpus."
+            ),
+            "params": 108310272,
+            "official_name": "BERT",
+            "path": "bert",
+            "model_card": "https://github.com/google-research/bert/blob/master/README.md",
+        },
+        "kaggle_handle": "kaggle://keras/bert/keras/bert_base_en/2",
+    },
+    "bert_base_zh": {
+        "metadata": {
+            "description": (
+                "12-layer BERT model. Trained on Chinese Wikipedia."
+            ),
+            "params": 102267648,
+            "official_name": "BERT",
+            "path": "bert",
+            "model_card": "https://github.com/google-research/bert/blob/master/README.md",
+        },
+        "kaggle_handle": "kaggle://keras/bert/keras/bert_base_zh/2",
+    },
+    "bert_base_multi": {
+        "metadata": {
+            "description": (
+                "12-layer BERT model where case is maintained. Trained on trained on Wikipedias of 104 languages"
+            ),
+            "params": 177853440,
+            "official_name": "BERT",
+            "path": "bert",
+            "model_card": "https://github.com/google-research/bert/blob/master/README.md",
+        },
+        "kaggle_handle": "kaggle://keras/bert/keras/bert_base_multi/2",
+    },
+    "bert_large_en_uncased": {
+        "metadata": {
+            "description": (
+                "24-layer BERT model where all input is lowercased. "
+                "Trained on English Wikipedia + BooksCorpus."
+            ),
+            "params": 335141888,
+            "official_name": "BERT",
+            "path": "bert",
+            "model_card": "https://github.com/google-research/bert/blob/master/README.md",
+        },
+        "kaggle_handle": "kaggle://keras/bert/keras/bert_large_en_uncased/2",
+    },
+    "bert_large_en": {
+        "metadata": {
+            "description": (
+                "24-layer BERT model where case is maintained. "
+                "Trained on English Wikipedia + BooksCorpus."
+            ),
+            "params": 333579264,
+            "official_name": "BERT",
+            "path": "bert",
+            "model_card": "https://github.com/google-research/bert/blob/master/README.md",
+        },
+        "kaggle_handle": "kaggle://keras/bert/keras/bert_large_en/2",
+    },
+}
+
+classifier_presets = {
+    "bert_tiny_en_uncased_sst2": {
+        "metadata": {
+            "description": (
+                "The bert_tiny_en_uncased backbone model fine-tuned on the SST-2 sentiment analysis dataset."
+            ),
+            "params": 4385920,
+            "official_name": "BERT",
+            "path": "bert",
+            "model_card": "https://github.com/google-research/bert/blob/master/README.md",
+        },
+        "kaggle_handle": "kaggle://keras/bert/keras/bert_tiny_en_uncased_sst2/4",
+    }
+}

keras_hub/src/models/bert/bert_tokenizer.py

@@ -0,0 +1,112 @@
+# 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.BertTokenizer")
+class BertTokenizer(WordPieceTokenizer):
+    """A BERT 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 BERT
+    models and provides a `from_preset()` method to automatically download
+    a matching vocabulary for a BERT preset.
+
+    This tokenizer does not provide truncation or padding of inputs. It can be
+    combined with a `keras_hub.models.BertPreprocessor` 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.BertTokenizer.from_preset(
+        "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.BertTokenizer(vocabulary=vocab)
+    tokenizer("The quick brown fox jumped.")
+    ```
+    """
+
+    def __init__(
+        self,
+        vocabulary=None,
+        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/bloom/__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.bloom.bloom_backbone import BloomBackbone
+from keras_hub.src.models.bloom.bloom_presets import backbone_presets
+from keras_hub.src.models.bloom.bloom_tokenizer import BloomTokenizer
+from keras_hub.src.utils.preset_utils import register_presets
+
+register_presets(backbone_presets, (BloomBackbone, BloomTokenizer))

keras_hub/src/models/bloom/bloom_attention.py

@@ -0,0 +1,186 @@
+# 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 math
+
+import keras
+from keras import ops
+
+from keras_hub.src.layers.modeling.alibi_bias import AlibiBias
+from keras_hub.src.utils.keras_utils import clone_initializer
+
+
+class BloomAttention(keras.layers.Layer):
+    def __init__(
+        self,
+        num_heads,
+        dropout=0.0,
+        kernel_initializer="glorot_uniform",
+        bias_initializer="zeros",
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.kernel_initializer = keras.initializers.get(kernel_initializer)
+        self.bias_initializer = keras.initializers.get(bias_initializer)
+
+    def build(self, inputs_shape):
+        batch_size, seq_length, hidden_dim = inputs_shape
+
+        self.head_dim = hidden_dim // self.num_heads
+
+        # Layer-wise attention scaling
+        self.inv_norm_factor = 1.0 / math.sqrt(self.head_dim)
+
+        self._query_dense = keras.layers.EinsumDense(
+            equation="btm,mnh->btnh",
+            output_shape=(None, self.num_heads, self.head_dim),
+            bias_axes="nh",
+            kernel_initializer=clone_initializer(self.kernel_initializer),
+            bias_initializer=clone_initializer(self.bias_initializer),
+            dtype=self.dtype_policy,
+            name="query_dense",
+        )
+        self._query_dense.build(inputs_shape)
+
+        self._key_dense = keras.layers.EinsumDense(
+            equation="bsm,mnh->bsnh",
+            output_shape=(None, self.num_heads, self.head_dim),
+            bias_axes="nh",
+            kernel_initializer=clone_initializer(self.kernel_initializer),
+            bias_initializer=clone_initializer(self.bias_initializer),
+            dtype=self.dtype_policy,
+            name="key_dense",
+        )
+        self._key_dense.build(inputs_shape)
+
+        self._value_dense = keras.layers.EinsumDense(
+            equation="bsm,mnh->bsnh",
+            output_shape=(None, self.num_heads, self.head_dim),
+            bias_axes="nh",
+            kernel_initializer=clone_initializer(self.kernel_initializer),
+            bias_initializer=clone_initializer(self.bias_initializer),
+            dtype=self.dtype_policy,
+            name="value_dense",
+        )
+        self._value_dense.build(inputs_shape)
+
+        self._alibi_layer = AlibiBias(
+            dtype=self.dtype_policy,
+        )
+
+        self._output_dense = keras.layers.Dense(
+            hidden_dim,
+            kernel_initializer=clone_initializer(self.kernel_initializer),
+            bias_initializer=clone_initializer(self.bias_initializer),
+            dtype=self.dtype_policy,
+            name="output_dense",
+        )
+        self._output_dense.build(inputs_shape)
+
+        self._dropout_layer = keras.layers.Dropout(
+            rate=self.dropout,
+            dtype=self.dtype_policy,
+            name="dropout",
+        )
+        self._softmax = keras.layers.Softmax(
+            dtype="float32",
+            name="softmax",
+        )
+
+        self.built = True
+
+    def call(
+        self,
+        hidden_states,
+        attention_mask=None,
+        cache=None,
+        cache_update_index=None,
+    ):
+        batch_size, seq_length, hidden_dim = ops.shape(hidden_states)
+
+        query = self._query_dense(hidden_states)
+        key = self._key_dense(hidden_states)
+        value = self._value_dense(hidden_states)
+
+        if cache is not None:
+            key_cache = cache[:, 0, ...]
+            value_cache = cache[:, 1, ...]
+            if cache_update_index is None:
+                key = key_cache
+                value = value_cache
+            else:
+                start = [0, cache_update_index, 0, 0]
+                key = ops.slice_update(key_cache, start, key)
+                value = ops.slice_update(value_cache, start, value)
+                cache = ops.stack((key, value), axis=1)
+        else:
+            if cache_update_index is not None:
+                raise ValueError(
+                    "`cache_update_index` should not be set if `cache` is "
+                    f"`None`. Received: cache={cache}, "
+                    f"cache_update_index={cache_update_index}"
+                )
+
+        # query (batch_size, num_heads, query_length, head_dim)
+        query = ops.transpose(query, [0, 2, 1, 3])
+        # value (batch_size, num_heads, kv_length, head_dim)
+        value = ops.transpose(value, [0, 2, 1, 3])
+        # key   (batch_size, num_heads, head_dim, kv_length)
+        key = ops.transpose(key, [0, 2, 3, 1])
+
+        attention_scores = (
+            ops.matmul(query, key) * self.inv_norm_factor
+        )  # [batch_size, num_heads, query_length, kv_length]
+        attention_scores = self._alibi_layer(attention_scores)
+        attention_scores = self._softmax(
+            attention_scores, ops.expand_dims(attention_mask, 1)
+        )
+        attention_scores = self._dropout_layer(attention_scores)
+
+        attention_output = ops.matmul(
+            attention_scores, value
+        )  # [batch_size, num_heads, query_length, head_dim]
+
+        attention_output = ops.transpose(
+            attention_output, [0, 2, 1, 3]
+        )  # [batch_size, query_length, num_heads, head_dim]
+        attention_output = ops.reshape(
+            attention_output,
+            [batch_size, seq_length, self.num_heads * self.head_dim],
+        )  # [batch_size, query_length, hidden_dim]
+
+        attention_output = self._output_dense(attention_output)
+        attention_output = self._dropout_layer(attention_output)
+
+        if cache is not None:
+            return attention_output, cache
+
+        return attention_output
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "num_heads": self.num_heads,
+                "dropout": self.dropout,
+                "kernel_initializer": keras.initializers.serialize(
+                    self.kernel_initializer
+                ),
+                "bias_initializer": keras.initializers.serialize(
+                    self.bias_initializer
+                ),
+            }
+        )
+        return config

keras_hub/src/models/bloom/bloom_backbone.py

@@ -0,0 +1,173 @@
+# 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.reversible_embedding import (
+    ReversibleEmbedding,
+)
+from keras_hub.src.models.backbone import Backbone
+from keras_hub.src.models.bloom.bloom_decoder import BloomDecoder
+
+
+def _bloom_kernel_initializer(stddev=0.02):
+    return keras.initializers.RandomNormal(stddev=stddev)
+
+
+@keras_hub_export("keras_hub.models.BloomBackbone")
+class BloomBackbone(Backbone):
+    """A BLOOM decoder network.
+
+    This network implements a Transformer-based decoder network, BigScience
+    Language Open-science Open-access Multilingual (BLOOM), as descriped in
+    ["BLOOM: A 176B-Parameter Open-Access Multilingual Language Model"](https://arxiv.org/pdf/2211.05100.pdf).
+
+    The default constructor gives a fully customizable, randomly initialized
+    Bloom model 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/spaces/bigscience/license).
+
+    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 dimensionality of the embeddings and hidden states.
+        intermediate_dim: int. The output dimension of the first Dense layer in
+            the MLP network of each transformer.
+        dropout: float. Dropout probability for the Transformer decoder.
+        layer_norm_epsilon: float. Epsilon for the layer normalization layers in
+            the transformer decoder.
+        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"),
+        "padding_mask": np.array([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0]]),
+    }
+
+    # Pretrained BLOOM decoder.
+    model = keras_hub.models.BloomBackbone.from_preset("bloom_560m_multi")
+    model(input_data)
+
+    # Randomly initialized BLOOM decoder with a custom config.
+    model = keras_hub.models.BloomBackbone(
+        vocabulary_size=10,
+        num_layers=2,
+        num_heads=2,
+        hidden_dim=32,
+        intermediate_dim=32*4,
+        dropout=0.0,
+        layer_norm_epsilon=1e-5,
+    )
+    model(input_data)
+    ```
+
+    """
+
+    def __init__(
+        self,
+        vocabulary_size,
+        num_layers,
+        num_heads,
+        hidden_dim,
+        intermediate_dim,
+        dropout=0.0,
+        layer_norm_epsilon=1e-5,
+        dtype=None,
+        **kwargs,
+    ):
+        # === Layers ===
+        self.token_embedding = ReversibleEmbedding(
+            input_dim=vocabulary_size,
+            output_dim=hidden_dim,
+            embeddings_initializer=_bloom_kernel_initializer(stddev=0.02),
+            dtype=dtype,
+            name="token_embedding",
+        )
+        self.embeddings_layer_norm = keras.layers.LayerNormalization(
+            epsilon=layer_norm_epsilon,
+            dtype=dtype,
+            name="embedding_layernorm",
+        )
+        self.transformer_layers = []
+        for i in range(num_layers):
+            layer = BloomDecoder(
+                num_heads=num_heads,
+                intermediate_dim=intermediate_dim,
+                dropout=dropout,
+                layer_norm_epsilon=layer_norm_epsilon,
+                dtype=dtype,
+                name=f"transformer_layer_{i}",
+            )
+            self.transformer_layers.append(layer)
+        self.layer_norm = keras.layers.LayerNormalization(
+            epsilon=layer_norm_epsilon,
+            dtype=dtype,
+            name="final_layernorm",
+        )
+
+        # === Functional Model ===
+        token_id_input = keras.Input(
+            shape=(None,), dtype="int32", name="token_ids"
+        )
+        padding_mask_input = keras.Input(
+            shape=(None,), dtype="int32", name="padding_mask"
+        )
+        x = self.token_embedding(token_id_input)
+        x = self.embeddings_layer_norm(x)
+        for transformer_layer in self.transformer_layers:
+            x = transformer_layer(x, decoder_padding_mask=padding_mask_input)
+        sequence_output = self.layer_norm(x)
+        super().__init__(
+            inputs={
+                "token_ids": token_id_input,
+                "padding_mask": padding_mask_input,
+            },
+            outputs=sequence_output,
+            dtype=dtype,
+            **kwargs,
+        )
+
+        # === Config ===
+        self.vocabulary_size = vocabulary_size
+        self.num_layers = num_layers
+        self.num_heads = num_heads
+        self.hidden_dim = hidden_dim
+        self.intermediate_dim = intermediate_dim
+        self.dropout = dropout
+        self.layer_norm_epsilon = layer_norm_epsilon
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "vocabulary_size": self.vocabulary_size,
+                "num_layers": self.num_layers,
+                "num_heads": self.num_heads,
+                "hidden_dim": self.hidden_dim,
+                "intermediate_dim": self.intermediate_dim,
+                "dropout": self.dropout,
+                "layer_norm_epsilon": self.layer_norm_epsilon,
+            }
+        )
+        return config