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

keras_hub/src/models/mistral/mistral_preprocessor.py

@@ -0,0 +1,190 @@
+# 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.mistral.mistral_tokenizer import MistralTokenizer
+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.MistralPreprocessor")
+class MistralPreprocessor(Preprocessor):
+    """A Mistral 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.StartEndPacker`.
+       with the appropriate tokens.
+     3. Construct a dictionary with keys `"token_ids"`, and `"padding_mask"`
+       that can be passed directly to `keras_hub.models.MistralBackbone`.
+
+    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.MistralTokenizer` 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. Default is `True`.
+        add_end_token: If `True`, the preprocessor will append the tokenizer
+            end token to each input sequence. Default is `False`.
+
+    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.
+        sequence_length: Pass to override the configured `sequence_length` of
+            the layer.
+
+    Examples:
+
+    Directly calling the from_preset().
+    ```python
+    preprocessor = keras_hub.models.MistralPreprocessor.from_preset(
+        "mistral_base_en"
+    )
+
+    # Tokenize and pack a single sentence.
+    preprocessor("The quick brown fox jumped.")
+
+    # Tokenize and a batch of single sentences.
+    preprocessor(["The quick brown fox jumped.", "Call me Ishmael."])
+
+    # Preprocess a batch of sentence pairs.
+    # When handling multiple sequences, always convert to tensors first!
+    first = tf.constant(["The quick brown fox jumped.", "Call me Ishmael."])
+    second = tf.constant(["The fox tripped.", "Oh look, a whale."])
+    preprocessor((first, second))
+    ```
+
+    Mapping with `tf.data.Dataset`.
+    ```python
+    preprocessor = keras_hub.models.MistralPreprocessor.from_preset(
+        "mistral_base_en"
+    )
+    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 = MistralTokenizer
+
+    def __init__(
+        self,
+        tokenizer,
+        sequence_length=1024,
+        add_start_token=True,
+        add_end_token=False,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.tokenizer = tokenizer
+        self.packer = None
+        self.add_start_token = add_start_token
+        self.add_end_token = add_end_token
+        self.sequence_length = sequence_length
+
+    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,
+            sequence_length=self.sequence_length,
+            return_padding_mask=True,
+        )
+        self.built = True
+
+    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
+
+    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(
+                "Mistral requires each input feature to contain only "
+                f"one segment, but received {len(x)}. If you are using Mistral"
+                " 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)
+
+    @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/mistral/mistral_presets.py

@@ -0,0 +1,48 @@
+# 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.
+"""Mistral model preset configurations."""
+
+# Metadata for loading pretrained model weights.
+backbone_presets = {
+    "mistral_7b_en": {
+        "metadata": {
+            "description": "Mistral 7B base model",
+            "params": 7241732096,
+            "official_name": "Mistral",
+            "path": "mistral",
+            "model_card": "https://github.com/mistralai/mistral-src/blob/main/README.md",
+        },
+        "kaggle_handle": "kaggle://keras/mistral/keras/mistral_7b_en/6",
+    },
+    "mistral_instruct_7b_en": {
+        "metadata": {
+            "description": "Mistral 7B instruct model",
+            "params": 7241732096,
+            "official_name": "Mistral",
+            "path": "mistral",
+            "model_card": "https://github.com/mistralai/mistral-src/blob/main/README.md",
+        },
+        "kaggle_handle": "kaggle://keras/mistral/keras/mistral_instruct_7b_en/6",
+    },
+    "mistral_0.2_instruct_7b_en": {
+        "metadata": {
+            "description": "Mistral 7B instruct Version 0.2 model",
+            "params": 7241732096,
+            "official_name": "Mistral",
+            "path": "mistral",
+            "model_card": "https://github.com/mistralai/mistral-src/blob/main/README.md",
+        },
+        "kaggle_handle": "kaggle://keras/mistral/keras/mistral_0.2_instruct_7b_en/1",
+    },
+}

keras_hub/src/models/mistral/mistral_tokenizer.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.
+
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.tokenizers.sentence_piece_tokenizer import (
+    SentencePieceTokenizer,
+)
+
+
+@keras_hub_export("keras_hub.models.MistralTokenizer")
+class MistralTokenizer(SentencePieceTokenizer):
+    """Mistral tokenizer layer based on SentencePiece.
+
+    This tokenizer class will tokenize raw strings into integer sequences and
+    is based on `keras_hub.tokenizers.SentencePieceTokenizer`. Unlike the
+    underlying tokenizer, it will check for all special tokens needed by
+    Mistral models and provides a `from_preset()` method to automatically
+    download a matching vocabulary for a Mistral preset.
+
+    This tokenizer does not provide truncation or padding of inputs. It can be
+    combined with a `keras_hub.models.MistralPreprocessor` 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:
+        proto: Either a `string` path to a SentencePiece proto file, or a
+            `bytes` object with a serialized SentencePiece proto. See the
+            [SentencePiece repository](https://github.com/google/sentencepiece)
+            for more details on the format.
+
+    Examples:
+    ```python
+    # Unbatched input.
+    tokenizer = keras_hub.models.MistralTokenizer.from_preset(
+        "mistral_7b_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."))
+    ```
+    """
+
+    def __init__(self, proto, **kwargs):
+        self.start_token = "<s>"
+        self.end_token = "</s>"
+        super().__init__(proto=proto, **kwargs)
+
+    def set_proto(self, proto):
+        super().set_proto(proto)
+        if proto is not None:
+            for token in [self.start_token, self.end_token]:
+                if token not in self.get_vocabulary():
+                    raise ValueError(
+                        f"Cannot find token `'{token}'` in the provided "
+                        f"`vocabulary`. Please provide `'{token}'` in your "
+                        "`vocabulary` or use a pretrained `vocabulary` name."
+                    )
+            self.start_token_id = self.token_to_id(self.start_token)
+            self.end_token_id = self.token_to_id(self.end_token)
+        else:
+            self.start_token_id = None
+            self.end_token_id = None

keras_hub/src/models/mistral/mistral_transformer_decoder.py

@@ -0,0 +1,265 @@
+# 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.mistral.mistral_attention import (
+    CachedMistralAttention,
+)
+from keras_hub.src.models.mistral.mistral_layer_norm import (
+    MistralLayerNormalization,
+)
+from keras_hub.src.utils.keras_utils import clone_initializer
+
+
+class MistralTransformerDecoder(keras.layers.Layer):
+    """A Transformer decoder layer for the Mistral backbone."""
+
+    def __init__(
+        self,
+        intermediate_dim,
+        num_query_heads,
+        num_key_value_heads,
+        rope_max_wavelength=10000,
+        rope_scaling_factor=1.0,
+        activation="silu",
+        layer_norm_epsilon=1e-5,
+        kernel_initializer="glorot_uniform",
+        sliding_window=512,
+        dropout=0,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.intermediate_dim = intermediate_dim
+        self.num_query_heads = num_query_heads
+        self.num_key_value_heads = num_key_value_heads
+
+        self.rope_max_wavelength = rope_max_wavelength
+        self.rope_scaling_factor = rope_scaling_factor
+
+        self.dropout = dropout
+
+        self.sliding_window = sliding_window
+        self.activation = keras.activations.get(activation)
+        self.layer_norm_epsilon = layer_norm_epsilon
+        self.kernel_initializer = keras.initializers.get(kernel_initializer)
+
+        self.supports_masking = True
+
+    def build(self, decoder_sequence_shape):
+        self._decoder_sequence_shape = decoder_sequence_shape
+        self.hidden_dim = decoder_sequence_shape[-1]
+
+        # Self attention layer.
+        self._self_attention_layer = CachedMistralAttention(
+            num_query_heads=self.num_query_heads,
+            num_key_value_heads=self.num_key_value_heads,
+            rope_max_wavelength=self.rope_max_wavelength,
+            rope_scaling_factor=self.rope_scaling_factor,
+            sliding_window=self.sliding_window,
+            kernel_initializer=clone_initializer(self.kernel_initializer),
+            dropout=self.dropout,
+            dtype=self.dtype_policy,
+            name="self_attention",
+        )
+        self._self_attention_layer.build(decoder_sequence_shape)
+
+        self._self_attention_layernorm = MistralLayerNormalization(
+            epsilon=self.layer_norm_epsilon,
+            dtype=self.dtype_policy,
+            name="self_attention_layernorm",
+        )
+        self._self_attention_layernorm.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,
+            kernel_initializer=clone_initializer(self.kernel_initializer),
+            use_bias=False,
+            dtype=self.dtype_policy,
+            name="feedforward_intermediate_dense",
+        )
+        self._feedforward_intermediate_dense.build(decoder_sequence_shape)
+
+        self._feedforward_gate_dense = keras.layers.Dense(
+            self.intermediate_dim,
+            kernel_initializer=clone_initializer(self.kernel_initializer),
+            use_bias=False,
+            dtype=self.dtype_policy,
+            name="feedforward_gate_dense",
+        )
+        self._feedforward_gate_dense.build(decoder_sequence_shape)
+
+        self._feedforward_output_dense = keras.layers.Dense(
+            self.hidden_dim,
+            kernel_initializer=clone_initializer(self.kernel_initializer),
+            use_bias=False,
+            dtype=self.dtype_policy,
+            name="feedforward_output_dense",
+        )
+
+        self._feedforward_output_dense.build(
+            self._feedforward_gate_dense.compute_output_shape(
+                decoder_sequence_shape
+            )
+        )
+
+        self._feedforward_layernorm = MistralLayerNormalization(
+            epsilon=self.layer_norm_epsilon,
+            dtype=self.dtype_policy,
+            name="feedforward_layernorm",
+        )
+        self._feedforward_layernorm.build(decoder_sequence_shape)
+
+        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,
+        training=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_layernorm(decoder_sequence)
+
+        # Self attention block.
+        x = self._self_attention_layer(
+            hidden_states=x,
+            attention_mask=self_attention_mask,
+            cache=self_attention_cache,
+            cache_update_index=self_attention_cache_update_index,
+        )
+
+        if self_attention_cache is not None:
+            x, self_attention_cache = x
+
+        x = self._self_attention_dropout(x, training=training)
+
+        x = x + residual
+        residual = x
+
+        x = self._feedforward_layernorm(x)
+        gate_output = self._feedforward_gate_dense(x)
+
+        # Note that we run the activation function in full 32-bit
+        # precision since this is what `torch.nn.functional.silu`
+        # does. Internally, `torch.nn.functional.silu` converts the
+        # inputs to float32, computes SiLU, and converts the outputs
+        # back to compute dtype.
+        # CPU Kernel: https://github.com/pytorch/pytorch/blob/35c493f2cf9b623bfdc7e6b34dc1cb39690a7919/aten/src/ATen/native/cpu/Activation.cpp#L1221-L1235  # noqa: E501
+        # CUDA Kernel: https://github.com/pytorch/pytorch/blob/35c493f2cf9b623bfdc7e6b34dc1cb39690a7919/aten/src/ATen/native/cuda/ActivationSiluKernel.cu  # noqa: E501
+        gate_output = ops.cast(gate_output, "float32")
+        gate_output = self.activation(gate_output)
+        gate_output = ops.cast(gate_output, self.compute_dtype)
+
+        x = self._feedforward_intermediate_dense(x)
+
+        x = self._feedforward_output_dense(ops.multiply(x, gate_output))
+
+        decoder_output = x + residual
+
+        if self_attention_cache is not None:
+            return decoder_output, self_attention_cache
+        return decoder_output
+
+    def _compute_self_attention_mask(
+        self,
+        decoder_sequence,
+        decoder_padding_mask,
+        decoder_attention_mask,
+        self_attention_cache,
+        self_attention_cache_update_index,
+    ):
+        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]
+
+        cache_update_index = (
+            0
+            if self_attention_cache_update_index is None
+            else self_attention_cache_update_index
+        )
+
+        # The lower traingular attention mask
+        causal_mask = compute_causal_mask(
+            batch_size, input_length, output_length, cache_update_index
+        )
+
+        # Mistral uses a banded attention mask if sliding window is not None
+        if self.sliding_window is not None:
+            # Below is a workaround for `ops.triu` for Keras 2.
+            # TODO(tirthasheshpatel): Use `ops.triu` once Keras 2 support is removed.
+            # causal_mask = ops.triu(causal_mask, k=-self.sliding_window)
+            i = ops.arange(output_length)[:, None] + cache_update_index
+            j = ops.arange(input_length)[None, :]
+            causal_mask_upper = ops.cast(i < j + self.sliding_window, "int32")
+            causal_mask = ops.minimum(causal_mask, causal_mask_upper)
+
+        return (
+            ops.minimum(decoder_mask, causal_mask)
+            if decoder_mask is not None
+            else causal_mask
+        )
+
+    def compute_output_shape(self, decoder_sequence_shape):
+        return decoder_sequence_shape
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "intermediate_dim": self.intermediate_dim,
+                "num_query_heads": self.num_query_heads,
+                "rope_max_wavelength": self.rope_max_wavelength,
+                "rope_scaling_factor": self.rope_scaling_factor,
+                "num_key_value_heads": self.num_key_value_heads,
+                "sliding_window": self.sliding_window,
+                "activation": keras.activations.serialize(self.activation),
+                "layer_norm_epsilon": self.layer_norm_epsilon,
+                "kernel_initializer": keras.initializers.serialize(
+                    self.kernel_initializer
+                ),
+                "dropout": self.dropout,
+            }
+        )
+        return config

keras_hub/src/models/mix_transformer/__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.