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

keras_hub/src/models/gpt2/gpt2_preprocessor.py

@@ -0,0 +1,187 @@
+# 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.start_end_packer import StartEndPacker
+from keras_hub.src.models.gpt2.gpt2_tokenizer import GPT2Tokenizer
+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.GPT2Preprocessor")
+class GPT2Preprocessor(Preprocessor):
+    """GPT2 preprocessing layer which tokenizes and packs inputs.
+
+    This preprocessing layer will do 2 things:
+
+    - Tokenize the inputs using the `tokenizer`.
+    - Construct a dictionary with keys `"token_ids"`, `"padding_mask"`, that can
+        be passed directly to a `keras_hub.models.GPT2Backbone`.
+
+    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`.
+
+    The call method of this layer accepts three arguments, `x`, `y`, and
+    `sample_weight`. `x` can be a python string or tensor representing a single
+    segment, a list of python strings representing a batch of single segments,
+    or a list of tensors representing multiple segments to be packed together.
+    `y` and `sample_weight` are both optional, can have any format, and will be
+    passed through unaltered.
+
+    `GPT2Preprocessor` forces the input to have only one segment, as GPT2 is
+    mainly used for generation tasks. For tasks having multi-segment inputs
+    like "glue/mnli", please use a model designed for classification purposes
+    such as BERT or RoBERTa.
+
+    Args:
+        tokenizer: A `keras_hub.models.GPT2Tokenizer` instance.
+        sequence_length: The length of the packed inputs.
+        add_start_token: If `True`, the preprocessor will prepend the tokenizer
+            start token to each input sequence.
+        add_end_token: If `True`, the preprocessor will append the tokenizer
+            end token to each input sequence.
+
+    Call arguments:
+        x: A string, `tf.Tensor` or list of python strings.
+        y: Any label data. Will be passed through unaltered.
+        sample_weight: Any label weight data. Will be passed through unaltered.
+        sequence_length: Pass to override the configured `sequence_length` of
+            the layer.
+
+    Examples:
+
+    Directly calling the layer on data.
+    ```python
+    preprocessor = keras_hub.models.GPT2Preprocessor.from_preset("gpt2_base_en")
+
+    # Tokenize and pack a single sentence.
+    preprocessor("The quick brown fox jumped.")
+
+    # Tokenize a batch of single sentences.
+    preprocessor(["The quick brown fox jumped.", "Call me Ishmael."])
+
+    # Custom vocabulary.
+    features = ["a quick fox.", "a fox quick."]
+    vocab = {"<|endoftext|>": 0, "a": 4, "Ġquick": 5, "Ġfox": 6}
+    merges = ["Ġ q", "u i", "c k", "ui ck", "Ġq uick"]
+    merges += ["Ġ f", "o x", "Ġf ox"]
+    tokenizer = keras_hub.models.GPT2Tokenizer(
+        vocabulary=vocab,
+        merges=merges,
+    )
+    preprocessor = keras_hub.models.GPT2Preprocessor(tokenizer=tokenizer)
+    preprocessor("The quick brown fox jumped.")
+    ```
+
+    Mapping with `tf.data.Dataset`.
+    ```python
+    preprocessor = keras_hub.models.GPT2Preprocessor.from_preset("gpt2_base_en")
+
+    text = tf.constant(["The quick brown fox jumped.", "Call me Ishmael."])
+    label = tf.constant([1, 1])
+
+    # Map labeled single sentences.
+    ds = tf.data.Dataset.from_tensor_slices((text, label))
+    ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
+
+    # Map unlabeled single sentences.
+    ds = tf.data.Dataset.from_tensor_slices(text)
+    ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
+    ```
+    """
+
+    tokenizer_cls = GPT2Tokenizer
+
+    def __init__(
+        self,
+        tokenizer,
+        sequence_length=1024,
+        add_start_token=True,
+        add_end_token=True,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.tokenizer = tokenizer
+        self.packer = None
+        self.sequence_length = sequence_length
+        self.add_start_token = add_start_token
+        self.add_end_token = add_end_token
+
+    def build(self, input_shape):
+        # Defer packer creation to `build()` so that we can be sure tokenizer
+        # assets have loaded when restoring a saved model.
+        self.packer = StartEndPacker(
+            start_value=self.tokenizer.start_token_id,
+            end_value=self.tokenizer.end_token_id,
+            pad_value=self.tokenizer.pad_token_id,
+            sequence_length=self.sequence_length,
+            return_padding_mask=True,
+        )
+        self.built = True
+
+    def call(
+        self,
+        x,
+        y=None,
+        sample_weight=None,
+        sequence_length=None,
+    ):
+        x = convert_inputs_to_list_of_tensor_segments(x)
+        if len(x) != 1:
+            raise ValueError(
+                "GPT2 requires each input feature to contain only "
+                f"one segment, but received {len(x)}. If you are using GPT2 "
+                "for a multi-segment classification task, please refer to "
+                "classification models like BERT or RoBERTa."
+            )
+        sequence_length = sequence_length or self.sequence_length
+        token_ids, padding_mask = self.packer(
+            self.tokenizer(x[0]),
+            sequence_length=sequence_length,
+            add_start_value=self.add_start_token,
+            add_end_value=self.add_end_token,
+        )
+        x = {
+            "token_ids": token_ids,
+            "padding_mask": padding_mask,
+        }
+        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,
+                "add_start_token": self.add_start_token,
+                "add_end_token": self.add_end_token,
+            }
+        )
+        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/gpt2/gpt2_presets.py

@@ -0,0 +1,82 @@
+# 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.
+"""GPT-2 model preset configurations."""
+
+# Metadata for loading pretrained model weights.
+backbone_presets = {
+    "gpt2_base_en": {
+        "metadata": {
+            "description": (
+                "12-layer GPT-2 model where case is maintained. "
+                "Trained on WebText."
+            ),
+            "params": 124439808,
+            "official_name": "GPT-2",
+            "path": "gpt2",
+            "model_card": "https://github.com/openai/gpt-2/blob/master/model_card.md",
+        },
+        "kaggle_handle": "kaggle://keras/gpt2/keras/gpt2_base_en/2",
+    },
+    "gpt2_medium_en": {
+        "metadata": {
+            "description": (
+                "24-layer GPT-2 model where case is maintained. "
+                "Trained on WebText."
+            ),
+            "params": 354823168,
+            "official_name": "GPT-2",
+            "path": "gpt2",
+            "model_card": "https://github.com/openai/gpt-2/blob/master/model_card.md",
+        },
+        "kaggle_handle": "kaggle://keras/gpt2/keras/gpt2_medium_en/2",
+    },
+    "gpt2_large_en": {
+        "metadata": {
+            "description": (
+                "36-layer GPT-2 model where case is maintained. "
+                "Trained on WebText."
+            ),
+            "params": 774030080,
+            "official_name": "GPT-2",
+            "path": "gpt2",
+            "model_card": "https://github.com/openai/gpt-2/blob/master/model_card.md",
+        },
+        "kaggle_handle": "kaggle://keras/gpt2/keras/gpt2_large_en/2",
+    },
+    "gpt2_extra_large_en": {
+        "metadata": {
+            "description": (
+                "48-layer GPT-2 model where case is maintained. "
+                "Trained on WebText."
+            ),
+            "params": 1557611200,
+            "official_name": "GPT-2",
+            "path": "gpt2",
+            "model_card": "https://github.com/openai/gpt-2/blob/master/model_card.md",
+        },
+        "kaggle_handle": "kaggle://keras/gpt2/keras/gpt2_extra_large_en/2",
+    },
+    "gpt2_base_en_cnn_dailymail": {
+        "metadata": {
+            "description": (
+                "12-layer GPT-2 model where case is maintained. "
+                "Finetuned on the CNN/DailyMail summarization dataset."
+            ),
+            "params": 124439808,
+            "official_name": "GPT-2",
+            "path": "gpt2",
+        },
+        "kaggle_handle": "kaggle://keras/gpt2/keras/gpt2_base_en_cnn_dailymail/2",
+    },
+}

keras_hub/src/models/gpt2/gpt2_tokenizer.py

@@ -0,0 +1,110 @@
+# 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.byte_pair_tokenizer import BytePairTokenizer
+
+
+@keras_hub_export("keras_hub.models.GPT2Tokenizer")
+class GPT2Tokenizer(BytePairTokenizer):
+    """A GPT-2 tokenizer using Byte-Pair Encoding subword segmentation.
+
+    This tokenizer class will tokenize raw strings into integer sequences and
+    is based on `keras_hub.tokenizers.BytePairTokenizer`. Unlike the
+    underlying tokenizer, it will check for all special tokens needed by GPT-2
+    models and provides a `from_preset()` method to automatically download
+    a matching vocabulary for a GPT-2 preset.
+
+    This tokenizer does not provide truncation or padding of inputs.
+
+    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: string or dict, maps token to integer ids. If it is a
+            string, it should be the file path to a json file.
+        merges: string or list, contains the merge rule. If it is a string,
+            it should be the file path to merge rules. The merge rule file
+            should have one merge rule per line. Every merge rule contains
+            merge entities separated by a space.
+
+    Examples:
+
+    ```python
+    # Unbatched input.
+    tokenizer = keras_hub.models.GPT2Tokenizer.from_preset("gpt2_base_en")
+    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 = {"<|endoftext|>": 0, "a": 4, "Ġquick": 5, "Ġfox": 6}
+    merges = ["Ġ q", "u i", "c k", "ui ck", "Ġq uick"]
+    merges += ["Ġ f", "o x", "Ġf ox"]
+    tokenizer = keras_hub.models.GPT2Tokenizer(vocabulary=vocab, merges=merges)
+    tokenizer("a quick fox.")
+    ```
+    """
+
+    def __init__(
+        self,
+        vocabulary=None,
+        merges=None,
+        **kwargs,
+    ):
+        # GPT2 uses the same start as end token, i.e., "<|endoftext|>".
+        self.end_token = self.start_token = "<|endoftext|>"
+
+        super().__init__(
+            vocabulary=vocabulary,
+            merges=merges,
+            unsplittable_tokens=[self.end_token],
+            **kwargs,
+        )
+
+    def set_vocabulary_and_merges(self, vocabulary, merges):
+        super().set_vocabulary_and_merges(vocabulary, merges)
+
+        if vocabulary is not None:
+            # Check for necessary special tokens.
+            if self.end_token not in self.get_vocabulary():
+                raise ValueError(
+                    f"Cannot find token `'{self.end_token}'` in the provided "
+                    f"`vocabulary`. Please provide `'{self.end_token}'` in "
+                    "your `vocabulary` or use a pretrained `vocabulary` name."
+                )
+
+            self.end_token_id = self.token_to_id(self.end_token)
+            self.start_token_id = self.end_token_id
+            self.pad_token_id = 0
+        else:
+            self.end_token_id = None
+            self.start_token_id = None
+            self.pad_token_id = None
+
+    def get_config(self):
+        config = super().get_config()
+        # In the constructor, we pass the list of special tokens to the
+        # `unsplittable_tokens` arg of the superclass' constructor. Hence, we
+        # delete it from the config here.
+        del config["unsplittable_tokens"]
+        return config

keras_hub/src/models/gpt_neo_x/__init__.py

@@ -0,0 +1,13 @@
+# 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.

keras_hub/src/models/gpt_neo_x/gpt_neo_x_attention.py

@@ -0,0 +1,251 @@
+# 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 import ops
+
+from keras_hub.src.layers.modeling.rotary_embedding import RotaryEmbedding
+from keras_hub.src.utils.keras_utils import clone_initializer
+
+
+class GPTNeoXAttention(keras.layers.Layer):
+    """GPTNeoXAttention layer.
+
+    This is an implementation of attention layer as described in the
+    paper ["GPT-NeoX-20B: An Open-Source Autoregressive Language Model"](https://arxiv.org/abs/2204.06745).
+    Effectively, this layer implements Multi-Head Self Attention with a rotary
+    embedding for encoding position information.
+
+    Args:
+        num_heads: int. Number of attention heads.
+        hidden_dim: int. Hidden dimension of the input, i.e., `hidden_states`.
+        bucket_size: int. The size of the relative position
+            buckets. Generally equal to `max_sequence_length // 2`.
+        dropout: float. Dropout probability.
+        kernel_initializer: string or `keras.initializers` initializer.
+            The kernel initializer for the dense layers.
+        bias_initializer: string or `keras.initializers` initializer.
+            The bias initializer for the dense layers.
+        rotary_percentage: float. The percentage by which query, key, value
+            matrices are to be rotated.
+        rotary_max_wavelength: int. The maximum angular wavelength of the
+            sine/cosine curves, for rotary embeddings.
+        max_sequence_length: int. The maximum input sequence length.
+    """
+
+    def __init__(
+        self,
+        num_heads,
+        hidden_dim,
+        dropout=0.0,
+        kernel_initializer="glorot_uniform",
+        bias_initializer="zeros",
+        rotary_percentage=0.25,
+        rotary_max_wavelength=10000,
+        max_sequence_length=512,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.num_heads = num_heads
+        self.hidden_dim = hidden_dim
+        self.rotary_percentage = rotary_percentage
+        self.dropout = dropout
+        self.attn_head_size = hidden_dim // num_heads
+        self.rotary_max_wavelength = rotary_max_wavelength
+        self.rotary_dim = int(self.attn_head_size * rotary_percentage)
+        self.rotary_embedding_layer = RotaryEmbedding(
+            max_wavelength=rotary_max_wavelength,
+            dtype=self.dtype_policy,
+        )
+        self.kernel_initializer = keras.initializers.get(kernel_initializer)
+        self.bias_initializer = keras.initializers.get(bias_initializer)
+        self.max_sequence_length = max_sequence_length
+
+    def build(self, input_shape):
+        self._qkv_dense = keras.layers.EinsumDense(
+            equation="abc,cde->abde",
+            output_shape=(None, self.num_heads, 3 * self.attn_head_size),
+            bias_axes="de",
+            **self._get_common_kwargs_for_sublayer(use_bias=True),
+            dtype=self.dtype_policy,
+            name="query_key_value",
+        )
+        self._qkv_dense.build(input_shape)
+
+        self._attn_dropout_layer = keras.layers.Dropout(
+            self.dropout,
+            dtype=self.dtype_policy,
+            name="attention_dropout",
+        )
+
+        self._softmax = keras.layers.Softmax(
+            axis=-1,
+            dtype="float32",
+            name="attention_softmax",
+        )
+
+        # Output.
+        self._output_dense = keras.layers.EinsumDense(
+            equation="abc,cd->abd",
+            output_shape=(None, self.hidden_dim),
+            bias_axes="d",
+            **self._get_common_kwargs_for_sublayer(use_bias=True),
+            dtype=self.dtype_policy,
+            name="attention_output",
+        )
+
+        self._output_dense.build(input_shape)
+        self.built = True
+
+    def _get_common_kwargs_for_sublayer(self, use_bias=True):
+        common_kwargs = {}
+
+        kernel_initializer = clone_initializer(self.kernel_initializer)
+        bias_initializer = clone_initializer(self.bias_initializer)
+
+        common_kwargs["kernel_initializer"] = kernel_initializer
+        if use_bias:
+            common_kwargs["bias_initializer"] = bias_initializer
+
+        return common_kwargs
+
+    def _masked_softmax(self, attention_scores, attention_mask=None):
+        if attention_mask is not None:
+            mask_expansion_axis = -3
+            for _ in range(
+                len(attention_scores.shape) - len(attention_mask.shape)
+            ):
+                attention_mask = ops.expand_dims(
+                    attention_mask, axis=mask_expansion_axis
+                )
+        return self._softmax(attention_scores, attention_mask)
+
+    def _compute_attention(
+        self, query, key, value, attention_mask=None, training=None
+    ):
+        attention_scores = ops.einsum("aecd,abcd->acbe", key, query)
+
+        norm_factor = ops.sqrt(
+            ops.convert_to_tensor(self.attn_head_size, self.compute_dtype)
+        )
+
+        attention_scores /= norm_factor
+
+        attention_scores = self._masked_softmax(
+            attention_scores, attention_mask
+        )
+        attention_scores = self._attn_dropout_layer(
+            attention_scores, training=training
+        )
+        attention_output = ops.einsum(
+            "acbe,aecd->abcd", attention_scores, value
+        )
+
+        return attention_output
+
+    def call(
+        self,
+        hidden_states,
+        attention_mask=None,
+        cache=None,
+        cache_update_index=None,
+        training=None,
+    ):
+        query_key_value = self._qkv_dense(hidden_states)
+
+        query = query_key_value[..., : self.attn_head_size]
+
+        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:
+                key_update = query_key_value[
+                    ..., self.attn_head_size : 2 * self.attn_head_size
+                ]
+                value_update = query_key_value[..., 2 * self.attn_head_size :]
+                start = [0, cache_update_index, 0, 0]
+                key = ops.slice_update(key_cache, start, key_update)
+                value = ops.slice_update(value_cache, start, value_update)
+                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}"
+                )
+            key = query_key_value[
+                ..., self.attn_head_size : 2 * self.attn_head_size
+            ]
+            value = query_key_value[..., 2 * self.attn_head_size :]
+
+        query_rot, query_pass = (
+            query[..., : self.rotary_dim],
+            query[..., self.rotary_dim :],
+        )
+        key_rot, key_pass = (
+            key[..., : self.rotary_dim],
+            key[..., self.rotary_dim :],
+        )
+
+        query_rot = self.rotary_embedding_layer(query_rot)
+        key_rot = self.rotary_embedding_layer(key_rot)
+
+        query = ops.concatenate((query_rot, query_pass), axis=-1)
+        key = ops.concatenate((key_rot, key_pass), axis=-1)
+
+        attention_output = self._compute_attention(
+            query=query,
+            key=key,
+            value=value,
+            attention_mask=attention_mask,
+            training=training,
+        )
+
+        # Reshape `attention_output` to `(batch_size, sequence_length, hidden_dim)`.
+        attention_output = ops.reshape(
+            attention_output,
+            [
+                ops.shape(attention_output)[0],
+                ops.shape(attention_output)[1],
+                self.hidden_dim,
+            ],
+        )
+
+        attention_output = self._output_dense(attention_output)
+
+        return attention_output, cache
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "num_heads": self.num_heads,
+                "hidden_dim": self.hidden_dim,
+                "dropout": self.dropout,
+                "kernel_initializer": keras.initializers.serialize(
+                    self.kernel_initializer
+                ),
+                "bias_initializer": keras.initializers.serialize(
+                    self.bias_initializer
+                ),
+                "rotary_percentage": self.rotary_percentage,
+                "rotary_max_wavelength": self.rotary_max_wavelength,
+                "max_sequence_length": self.max_sequence_length,
+            }
+        )
+        return config