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

keras_hub/src/models/gpt_neo_x/gpt_neo_x_decoder.py

@@ -0,0 +1,258 @@
+# 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.transformer_layer_utils import (
+    compute_causal_mask,
+)
+from keras_hub.src.layers.modeling.transformer_layer_utils import (
+    merge_padding_and_attention_mask,
+)
+from keras_hub.src.models.gpt_neo_x.gpt_neo_x_attention import GPTNeoXAttention
+from keras_hub.src.utils.keras_utils import clone_initializer
+
+
+class GPTNeoXDecoder(keras.layers.Layer):
+    """GPTNeoX decoder.
+
+    This class follows the architecture of the GPT-NeoX decoder layer in the
+    paper [GPT-NeoX-20B: An Open-Source Autoregressive Language Model](https://arxiv.org/abs/2204.06745).
+    Users can instantiate multiple instances of this class to stack up a decoder.
+
+    This layer will always apply a causal mask to the decoder attention layer.
+
+    Args:
+        intermediate_dim: int, the hidden size of feedforward network.
+        num_heads: int, the number of heads for multi-head attention.
+        dropout: float. the dropout value, shared by
+            the multi-head attention and feedforward layers.
+        activation: string or `keras.activations`. the activation function of
+            feedforward network.
+        layer_norm_epsilon: float. The epsilon value in layer
+            normalization components.
+        kernel_initializer: string or `keras.initializers` initializer. The
+            kernel initializer for the dense and multi-head  attention layers.
+        bias_initializer: string or `keras.initializers` initializer. The bias
+            initializer for the dense and multi-head  attention layers.
+        rotary_max_wavelength: int. The maximum angular wavelength of the
+            sine/cosine curves, for rotary embeddings.
+        rotary_percentage: float. The percentage by which query, key, value
+            matrices are to be rotated.
+        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.
+        name: string. The name of the layer.
+    """
+
+    def __init__(
+        self,
+        intermediate_dim,
+        num_heads,
+        dropout=0.0,
+        activation="relu",
+        layer_norm_epsilon=1e-5,
+        kernel_initializer="glorot_uniform",
+        bias_initializer="zeros",
+        rotary_percentage=0.25,
+        rotary_max_wavelength=10000,
+        max_sequence_length=512,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.intermediate_dim = intermediate_dim
+        self.num_heads = num_heads
+        self.dropout = dropout
+        self.rotary_percentage = rotary_percentage
+        self.rotary_max_wavelength = rotary_max_wavelength
+        self.max_sequence_length = max_sequence_length
+        self.activation = keras.activations.get(activation)
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.kernel_initializer = keras.initializers.get(kernel_initializer)
+        self.bias_initializer = keras.initializers.get(bias_initializer)
+        self.supports_masking = True
+        self.rotary_percentage = rotary_percentage
+        self._decoder_sequence_shape = None
+
+    def build(self, decoder_sequence_shape):
+        self._decoder_sequence_shape = decoder_sequence_shape
+        hidden_dim = decoder_sequence_shape[-1]
+        # Self attention layers.
+        self._self_attention_layer = GPTNeoXAttention(
+            num_heads=self.num_heads,
+            hidden_dim=hidden_dim,
+            dropout=self.dropout,
+            rotary_percentage=self.rotary_percentage,
+            rotary_max_wavelength=self.rotary_max_wavelength,
+            max_sequence_length=self.max_sequence_length,
+            kernel_initializer=clone_initializer(self.kernel_initializer),
+            bias_initializer=clone_initializer(self.bias_initializer),
+            dtype=self.dtype_policy,
+            name="self_attention",
+        )
+        self._self_attention_layer.build(decoder_sequence_shape)
+
+        self._self_attention_layer_norm = keras.layers.LayerNormalization(
+            epsilon=self.layer_norm_epsilon,
+            dtype=self.dtype_policy,
+            name="self_attention_layer_norm",
+        )
+        self._self_attention_layer_norm.build(decoder_sequence_shape)
+
+        self._self_attention_dropout = keras.layers.Dropout(
+            rate=self.dropout,
+            dtype=self.dtype_policy,
+            name="self_attention_dropout",
+        )
+
+        # Feedforward layers.
+        self._feedforward_intermediate_dense = keras.layers.Dense(
+            self.intermediate_dim,
+            activation=self.activation,
+            kernel_initializer=clone_initializer(self.kernel_initializer),
+            bias_initializer=clone_initializer(self.bias_initializer),
+            dtype=self.dtype_policy,
+            name="feedforward_intermediate_dense",
+        )
+        self._feedforward_intermediate_dense.build(decoder_sequence_shape)
+
+        self._feedforward_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="feedforward_output_dense",
+        )
+
+        intermediate_shape = list(decoder_sequence_shape)
+        intermediate_shape[-1] = self.intermediate_dim
+        self._feedforward_output_dense.build(tuple(intermediate_shape))
+
+        self._feedforward_layer_norm = keras.layers.LayerNormalization(
+            epsilon=self.layer_norm_epsilon,
+            dtype=self.dtype_policy,
+            name="feedforward_layer_norm",
+        )
+        self._feedforward_layer_norm.build(decoder_sequence_shape)
+
+        self._feedforward_dropout = keras.layers.Dropout(
+            rate=self.dropout,
+            dtype=self.dtype_policy,
+            name="feedforward_dropout",
+        )
+        self.built = True
+
+    def call(
+        self,
+        decoder_sequence,
+        decoder_padding_mask=None,
+        decoder_attention_mask=None,
+        self_attention_cache=None,
+        self_attention_cache_update_index=None,
+    ):
+        self_attention_mask = self._compute_self_attention_mask(
+            decoder_sequence=decoder_sequence,
+            decoder_padding_mask=decoder_padding_mask,
+            decoder_attention_mask=decoder_attention_mask,
+            self_attention_cache=self_attention_cache,
+            self_attention_cache_update_index=self_attention_cache_update_index,
+        )
+
+        residual = decoder_sequence
+
+        x = self._self_attention_layer_norm(decoder_sequence)
+
+        # Self attention block.
+        x, self_attention_cache = self._self_attention_layer(
+            hidden_states=x,
+            attention_mask=self_attention_mask,
+            cache=self_attention_cache,
+            cache_update_index=self_attention_cache_update_index,
+        )
+        x = self._self_attention_dropout(x)
+        attention_output = x
+
+        x = self._feedforward_layer_norm(decoder_sequence)
+        x = self._feedforward_intermediate_dense(x)
+        x = self._feedforward_output_dense(x)
+        feedforward_output = x
+        x = feedforward_output + attention_output + residual
+
+        if self_attention_cache is not None:
+            return (x, self_attention_cache)
+        else:
+            return x
+
+    def _compute_self_attention_mask(
+        self,
+        decoder_sequence,
+        decoder_padding_mask,
+        decoder_attention_mask,
+        self_attention_cache=None,
+        self_attention_cache_update_index=None,
+    ):
+        decoder_mask = merge_padding_and_attention_mask(
+            decoder_sequence, decoder_padding_mask, decoder_attention_mask
+        )
+        batch_size = ops.shape(decoder_sequence)[0]
+        input_length = output_length = ops.shape(decoder_sequence)[1]
+        # We need to handle a rectangular causal mask when doing cached
+        # decoding. For generative inference, `decoder_sequence` will
+        # generally be length 1, and `cache` will be the full generation length.
+        if self_attention_cache is not None:
+            input_length = ops.shape(self_attention_cache)[2]
+
+        causal_mask = compute_causal_mask(
+            batch_size,
+            input_length,
+            output_length,
+            (
+                0
+                if self_attention_cache_update_index is None
+                else self_attention_cache_update_index
+            ),
+        )
+        return (
+            ops.minimum(decoder_mask, causal_mask)
+            if decoder_mask is not None
+            else causal_mask
+        )
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "intermediate_dim": self.intermediate_dim,
+                "num_heads": self.num_heads,
+                "dropout": self.dropout,
+                "rotary_percentage": self.rotary_percentage,
+                "rotary_max_wavelength": self.rotary_max_wavelength,
+                "max_sequence_length": self.max_sequence_length,
+                "activation": keras.activations.serialize(self.activation),
+                "layer_norm_epsilon": self.layer_norm_epsilon,
+                "kernel_initializer": keras.initializers.serialize(
+                    self.kernel_initializer
+                ),
+                "bias_initializer": keras.initializers.serialize(
+                    self.bias_initializer
+                ),
+                "decoder_sequence_shape": self._decoder_sequence_shape,
+            }
+        )
+        return config
+
+    def compute_output_shape(self, decoder_sequence_shape):
+        return decoder_sequence_shape

keras_hub/src/models/gpt_neo_x/gpt_neo_x_preprocessor.py

@@ -0,0 +1,145 @@
+# 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.gpt_neo_x.gpt_neo_x_tokenizer import GPTNeoXTokenizer
+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.GPTNeoXPreprocessor")
+class GPTNeoXPreprocessor(Preprocessor):
+    """GPTNeoX 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.GPTNeoXBackbone`.
+
+    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.
+
+    `GPTNeoXPreprocessor` forces the input to have only one segment, as GPTNeoX 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.GPTNeoXTokenizer` 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.
+    """
+
+    tokenizer_cls = GPTNeoXTokenizer
+
+    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(
+                "GPTNeoX requires each input feature to contain only "
+                f"one segment, but received {len(x)}. If you are using GPTNeoX "
+                "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/gpt_neo_x/gpt_neo_x_tokenizer.py

@@ -0,0 +1,88 @@
+# 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.GPTNeoXTokenizer")
+class GPTNeoXTokenizer(BytePairTokenizer):
+    """A GPTNeoX 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 GPTNeoX
+    models and provides a `from_preset()` method to automatically download
+    a matching vocabulary for a GPTNeoX 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.
+    """
+
+    def __init__(
+        self,
+        vocabulary=None,
+        merges=None,
+        **kwargs,
+    ):
+        # GPTNeoX 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/image_classifier.py

@@ -0,0 +1,90 @@
+# Copyright 2023 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.models.task import Task
+
+
+@keras_hub_export("keras_hub.models.ImageClassifier")
+class ImageClassifier(Task):
+    """Base class for all image classification tasks.
+
+    `ImageClassifier` tasks wrap a `keras_hub.models.Backbone` and
+    a `keras_hub.models.Preprocessor` to create a model that can be used for
+    image classification. `ImageClassifier` tasks take an additional
+    `num_classes` argument, controlling the number of predicted output classes.
+
+    To fine-tune with `fit()`, pass a dataset containing tuples of `(x, y)`
+    labels where `x` is a string and `y` is a integer from `[0, num_classes)`.
+
+    All `ImageClassifier` tasks include a `from_preset()` constructor which can be
+    used to load a pre-trained config and weights.
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        # Default compilation.
+        self.compile()
+
+    def compile(
+        self,
+        optimizer="auto",
+        loss="auto",
+        *,
+        metrics="auto",
+        **kwargs,
+    ):
+        """Configures the `ImageClassifier` task for training.
+
+        The `ImageClassifier` task extends the default compilation signature of
+        `keras.Model.compile` with defaults for `optimizer`, `loss`, and
+        `metrics`. To override these defaults, pass any value
+        to these arguments during compilation.
+
+        Args:
+            optimizer: `"auto"`, an optimizer name, or a `keras.Optimizer`
+                instance. Defaults to `"auto"`, which uses the default optimizer
+                for the given model and task. See `keras.Model.compile` and
+                `keras.optimizers` for more info on possible `optimizer` values.
+            loss: `"auto"`, a loss name, or a `keras.losses.Loss` instance.
+                Defaults to `"auto"`, where a
+                `keras.losses.SparseCategoricalCrossentropy` loss will be
+                applied for the classification task. See
+                `keras.Model.compile` and `keras.losses` for more info on
+                possible `loss` values.
+            metrics: `"auto"`, or a list of metrics to be evaluated by
+                the model during training and testing. Defaults to `"auto"`,
+                where a `keras.metrics.SparseCategoricalAccuracy` will be
+                applied to track the accuracy of the model during training.
+                See `keras.Model.compile` and `keras.metrics` for
+                more info on possible `metrics` values.
+            **kwargs: See `keras.Model.compile` for a full list of arguments
+                supported by the compile method.
+        """
+        if optimizer == "auto":
+            optimizer = keras.optimizers.Adam(5e-5)
+        if loss == "auto":
+            activation = getattr(self, "activation", None)
+            activation = keras.activations.get(activation)
+            from_logits = activation != keras.activations.softmax
+            loss = keras.losses.SparseCategoricalCrossentropy(from_logits)
+        if metrics == "auto":
+            metrics = [keras.metrics.SparseCategoricalAccuracy()]
+        super().compile(
+            optimizer=optimizer,
+            loss=loss,
+            metrics=metrics,
+            **kwargs,
+        )

keras_hub/src/models/llama/__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.llama.llama_backbone import LlamaBackbone
+from keras_hub.src.models.llama.llama_presets import backbone_presets
+from keras_hub.src.models.llama.llama_tokenizer import LlamaTokenizer
+from keras_hub.src.utils.preset_utils import register_presets
+
+register_presets(backbone_presets, (LlamaBackbone, LlamaTokenizer))