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

keras_hub/src/models/llama3/llama3_backbone.py

@@ -0,0 +1,84 @@
+# 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.models.llama.llama_backbone import LlamaBackbone
+
+
+# LLaMA 3 shares the same architecture as its predecessors
+# So, we simply create an alias for API consistency
+@keras_hub_export("keras_hub.models.Llama3Backbone")
+class Llama3Backbone(LlamaBackbone):
+    """
+    The Llama Transformer core architecture with hyperparameters.
+
+    This network implements a Transformer-based decoder network,
+    Llama, as described in
+    ["Llama 7B"](https://arxiv.org/pdf/2310.06825.pdf).
+    It includes the embedding lookups and transformer layers.
+
+    The default constructor gives a fully customizable, randomly initialized
+    Llama model with any number of layers, heads, and embedding
+    dimensions. To load preset architectures and weights, use the `from_preset`
+    constructor.
+
+    Args:
+        vocabulary_size (int): The size of the token vocabulary.
+        num_layers (int): The number of transformer layers.
+        num_query_heads (int): The number of query attention heads for
+            each transformer.
+        hidden_dim (int): The size of the transformer encoding and pooling layers.
+        intermediate_dim (int): The output dimension of the first Dense layer in a
+            three-layer feedforward network for each transformer.
+        num_key_value_heads (int): The number of key and value attention heads for
+            each transformer.
+        rope_max_wavelength (int, optional): The maximum angular wavelength of the
+            sine/cosine curves, for rotary embeddings. Defaults to `10000`.
+        rope_scaling_factor (float, optional): The scaling factor for calculation
+            of roatary embedding. Defaults to `1.0`.
+        layer_norm_epsilon (float, optional): Epsilon for the layer normalization
+            layers in the transformer decoder. Defaults to `1e-6`.
+        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.
+
+    Examples:
+
+    ```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 Llama decoder.
+    model = keras_hub.models.Llama3Backbone.from_preset("llama3_8b_en")
+    model(input_data)
+
+    # Randomly initialized Llama decoder with custom config.
+    model = keras_hub.models.Llama3Backbone(
+        vocabulary_size=10,
+        hidden_dim=512,
+        num_layers=2,
+        num_query_heads=32,
+        num_key_value_heads=8,
+        intermediate_dim=1024,
+        layer_norm_epsilon=1e-6,
+        dtype="float32"
+    )
+    model(input_data)
+    ```
+    """
+
+    pass

keras_hub/src/models/llama3/llama3_causal_lm.py

@@ -0,0 +1,46 @@
+# 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.models.llama3.llama3_backbone import Llama3Backbone
+from keras_hub.src.models.llama3.llama3_causal_lm_preprocessor import (
+    Llama3CausalLMPreprocessor,
+)
+from keras_hub.src.models.llama.llama_causal_lm import LlamaCausalLM
+
+
+@keras_hub_export("keras_hub.models.Llama3CausalLM")
+class Llama3CausalLM(LlamaCausalLM):
+    """An end-to-end Llama 3 model for causal language modeling.
+
+    A causal language model (LM) predicts the next token based on previous
+    tokens. This task setup can be used to train the model unsupervised on
+    plain text input, or to autoregressively generate plain text similar to
+    the data used for training. This task can be used for pre-training or
+    fine-tuning a LLaMA 3 model, simply by calling `fit()`.
+
+    This model has a `generate()` method, which generates text based on a
+    prompt. The generation strategy used is controlled by an additional
+    `sampler` argument on `compile()`. You can recompile the model with
+    different `keras_hub.samplers` objects to control the generation. By
+    default, `"top_k"` sampling will be used.
+
+    Args:
+        backbone: A `keras_hub.models.Llama3Backbone` instance.
+        preprocessor: A `keras_hub.models.Llama3CausalLMPreprocessor` or `None`.
+            If `None`, this model will not apply preprocessing, and inputs
+            should be preprocessed before calling the model.
+    """
+
+    backbone_cls = Llama3Backbone
+    preprocessor_cls = Llama3CausalLMPreprocessor

keras_hub/src/models/llama3/llama3_causal_lm_preprocessor.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 absl import logging
+
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.models.llama3.llama3_preprocessor import Llama3Preprocessor
+from keras_hub.src.utils.keras_utils import (
+    convert_inputs_to_list_of_tensor_segments,
+)
+from keras_hub.src.utils.tensor_utils import strip_to_ragged
+
+
+@keras_hub_export("keras_hub.models.Llama3CausalLMPreprocessor")
+class Llama3CausalLMPreprocessor(Llama3Preprocessor):
+    """Llama 3 Causal LM preprocessor.
+
+    This preprocessing layer is meant for use with
+    `keras_hub.models.Llama3CausalLM`. By default, it will take in batches of
+    strings, and return outputs in a `(x, y, sample_weight)` format, where the
+    `y` label is the next token id in the `x` sequence.
+
+    For use with generation, the layer also exposes two methods
+    `generate_preprocess()` and `generate_postprocess()`. When this preprocessor
+    is attached to a `keras_hub.models.Llama3CausalLM` instance, these methods
+    will be called implicitly in `generate()`. They can also be called
+    standalone (e.g. to precompute preprocessing inputs for generation in a
+    separate process).
+
+    Args:
+        tokenizer: A `keras_hub.models.Llama3Tokenizer` 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 `False`.
+        add_end_token: If `True`, the preprocessor will append the tokenizer
+            end token to each input sequence. Default is `False`.
+
+    Call arguments:
+        x: A string, `tf.Tensor` or list of python strings.
+        y: Label data. Should always be `None` as the layer generates labels.
+        sample_weight: Label weights. Should always be `None` as the layer
+            generates label weights.
+        sequence_length: Pass to override the configured `sequence_length` of
+            the layer.
+
+    Examples:
+    ```python
+    # Load the preprocessor from a preset.
+    preprocessor = keras_hub.models.Llama3CausalLMPreprocessor.from_preset(
+        "llama_base_en"
+    )
+
+    # Tokenize and pack a single sentence.
+    sentence = tf.constant("League of legends")
+    preprocessor(sentence)
+    # Same output.
+    preprocessor("League of legends")
+
+    # Tokenize a batch of sentences.
+    sentences = tf.constant(["Taco tuesday", "Fish taco please!"])
+    preprocessor(sentences)
+    # Same output.
+    preprocessor(["Taco tuesday", "Fish taco please!"])
+
+    # Map a dataset to preprocess a single sentence.
+    features = tf.constant(
+        [
+            "Avatar 2 is amazing!",
+            "Well, I am not sure.",
+        ]
+    )
+    labels = tf.constant([1, 0])
+    ds = tf.data.Dataset.from_tensor_slices((features, labels))
+    ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
+
+    # Map a dataset to preprocess unlabled sentences.
+    ds = tf.data.Dataset.from_tensor_slices(features)
+    ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
+    ```
+    """
+
+    def call(
+        self,
+        x,
+        y=None,
+        sample_weight=None,
+        sequence_length=None,
+    ):
+        if y is not None or sample_weight is not None:
+            logging.warning(
+                "`Llama3CausalLMPreprocessor` generates `y` and "
+                "`sample_weight` based on your input data, but your data "
+                "already contains `y` or `sample_weight`. Your `y` and "
+                "`sample_weight` will be ignored."
+            )
+        sequence_length = sequence_length or self.sequence_length
+
+        x = convert_inputs_to_list_of_tensor_segments(x)[0]
+        x = self.tokenizer(x)
+        # Pad with one extra token to account for the truncation below.
+        token_ids, padding_mask = self.packer(
+            x,
+            sequence_length=sequence_length + 1,
+            add_start_value=self.add_start_token,
+            add_end_value=self.add_end_token,
+        )
+        # The last token does not have a next token, so we truncate it out.
+        x = {
+            "token_ids": token_ids[..., :-1],
+            "padding_mask": padding_mask[..., :-1],
+        }
+        # Target `y` will be the next token.
+        y, sample_weight = token_ids[..., 1:], padding_mask[..., 1:]
+        return keras.utils.pack_x_y_sample_weight(x, y, sample_weight)
+
+    def generate_preprocess(
+        self,
+        x,
+        sequence_length=None,
+    ):
+        """Convert strings to integer token input for generation.
+
+        Similar to calling the layer for training, this method takes in strings
+        or tensor strings, tokenizes and packs the input, and computes a padding
+        mask masking all inputs not filled in with a padded value.
+
+        Unlike calling the layer for training, this method does not compute
+        labels and will never append a `tokenizer.end_token_id` to the end of
+        the sequence (as generation is expected to continue at the end of the
+        inputted prompt).
+        """
+        if not self.built:
+            self.build(None)
+
+        x = convert_inputs_to_list_of_tensor_segments(x)[0]
+        x = self.tokenizer(x)
+        token_ids, padding_mask = self.packer(
+            x, sequence_length=sequence_length, add_end_value=False
+        )
+        return {
+            "token_ids": token_ids,
+            "padding_mask": padding_mask,
+        }
+
+    def generate_postprocess(
+        self,
+        x,
+    ):
+        """Convert integer token output to strings for generation.
+
+        This method reverses `generate_preprocess()`, by first removing all
+        padding and start/end tokens, and then converting the integer sequence
+        back to a string.
+        """
+        token_ids, padding_mask = x["token_ids"], x["padding_mask"]
+        ids_to_strip = (
+            self.tokenizer.end_token_id,
+            self.tokenizer.start_token_id,
+        )
+        token_ids = strip_to_ragged(token_ids, padding_mask, ids_to_strip)
+        return self.tokenizer.detokenize(token_ids)

keras_hub/src/models/llama3/llama3_preprocessor.py

@@ -0,0 +1,21 @@
+# 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.models.llama3.llama3_tokenizer import Llama3Tokenizer
+from keras_hub.src.models.llama.llama_preprocessor import LlamaPreprocessor
+
+
+@keras_hub_export("keras_hub.models.Llama3Preprocessor")
+class Llama3Preprocessor(LlamaPreprocessor):
+    tokenizer_cls = Llama3Tokenizer

keras_hub/src/models/llama3/llama3_presets.py

@@ -0,0 +1,69 @@
+# 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.
+"""Llama 3 model preset configurations."""
+
+# Metadata for loading pretrained model weights.
+backbone_presets = {
+    "llama3_8b_en": {
+        "metadata": {
+            "description": "8 billion parameter, 32-layer, base LLaMA 3 model.",
+            "params": 8030261248,
+            "official_name": "LLaMA 3",
+            "path": "llama3",
+            "model_card": "https://github.com/meta-llama/llama3",
+        },
+        "kaggle_handle": "kaggle://keras/llama3/keras/llama3_8b_en/3",
+    },
+    "llama3_8b_en_int8": {
+        "metadata": {
+            "description": (
+                "8 billion parameter, 32-layer, base LLaMA 3 model with "
+                "activation and weights quantized to int8."
+            ),
+            "params": 8031894016,
+            "official_name": "LLaMA 3",
+            "path": "llama3",
+            "model_card": "https://github.com/meta-llama/llama3",
+        },
+        "kaggle_handle": "kaggle://keras/llama3/keras/llama3_8b_en_int8/1",
+    },
+    "llama3_instruct_8b_en": {
+        "metadata": {
+            "description": (
+                "8 billion parameter, 32-layer, instruction tuned LLaMA 3 "
+                "model."
+            ),
+            "params": 8030261248,
+            "official_name": "LLaMA 3",
+            "path": "llama3",
+            "model_card": "https://github.com/meta-llama/llama3",
+        },
+        "kaggle_handle": "kaggle://keras/llama3/keras/llama3_instruct_8b_en/3",
+    },
+    "llama3_instruct_8b_en_int8": {
+        "metadata": {
+            "description": (
+                "8 billion parameter, 32-layer, instruction tuned LLaMA 3 "
+                "model with activation and weights quantized to int8."
+            ),
+            "params": 8031894016,
+            "official_name": "LLaMA 3",
+            "path": "llama3",
+            "model_card": "https://github.com/meta-llama/llama3",
+        },
+        "kaggle_handle": (
+            "kaggle://keras/llama3/keras/llama3_instruct_8b_en_int8/1"
+        ),
+    },
+}

keras_hub/src/models/llama3/llama3_tokenizer.py

@@ -0,0 +1,63 @@
+# 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.Llama3Tokenizer")
+class Llama3Tokenizer(BytePairTokenizer):
+    def __init__(
+        self,
+        vocabulary=None,
+        merges=None,
+        **kwargs,
+    ):
+        self.start_token = "<|begin_of_text|>"
+        self.end_token = "<|end_of_text|>"
+
+        super().__init__(
+            vocabulary=vocabulary,
+            merges=merges,
+            unsplittable_tokens=[self.start_token, 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.start_token_id = self.token_to_id(self.start_token)
+            self.end_token_id = self.token_to_id(self.end_token)
+            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/masked_lm.py

@@ -0,0 +1,101 @@
+# 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.models.task import Task
+
+
+@keras_hub_export("keras_hub.models.MaskedLM")
+class MaskedLM(Task):
+    """Base class for masked language modeling tasks.
+
+    `MaskedLM` tasks wrap a `keras_hub.models.Backbone` and
+    a `keras_hub.models.Preprocessor` to create a model that can be used for
+    unsupervised fine-tuning with a masked language modeling loss.
+
+    When calling `fit()`, all input will be tokenized, and random tokens in
+    the input sequence will be masked. These positions of these masked tokens
+    will be fed as an additional model input, and the original value of the
+    tokens predicted by the model outputs.
+
+    All `MaskedLM` tasks include a `from_preset()` constructor which can be used
+    to load a pre-trained config and weights.
+
+    Example:
+    ```python
+    # Load a Bert MaskedLM with pre-trained weights.
+    masked_lm = keras_hub.models.MaskedLM.from_preset(
+        "bert_base_en",
+    )
+    masked_lm.fit(train_ds)
+    ```
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        # Default compilation.
+        self.compile()
+
+    def compile(
+        self,
+        optimizer="auto",
+        loss="auto",
+        *,
+        weighted_metrics="auto",
+        **kwargs,
+    ):
+        """Configures the `MaskedLM` task for training.
+
+        The `MaskedLM` task extends the default compilation signature of
+        `keras.Model.compile` with defaults for `optimizer`, `loss`, and
+        `weighted_metrics`. To override these defaults, pass any value
+        to these arguments during compilation.
+
+        Note that because training inputs include padded tokens which are
+        excluded from the loss, it is almost always a good idea to compile with
+        `weighted_metrics` and not `metrics`.
+
+        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 token classification `MaskedLM` task. See
+                `keras.Model.compile` and `keras.losses` for more info on
+                possible `loss` values.
+            weighted_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 at guessing masked
+                token values. See `keras.Model.compile` and `keras.metrics` for
+                more info on possible `weighted_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":
+            loss = keras.losses.SparseCategoricalCrossentropy(from_logits=True)
+        if weighted_metrics == "auto":
+            weighted_metrics = [keras.metrics.SparseCategoricalAccuracy()]
+        super().compile(
+            optimizer=optimizer,
+            loss=loss,
+            weighted_metrics=weighted_metrics,
+            **kwargs,
+        )

keras_hub/src/models/mistral/__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.mistral.mistral_backbone import MistralBackbone
+from keras_hub.src.models.mistral.mistral_presets import backbone_presets
+from keras_hub.src.models.mistral.mistral_tokenizer import MistralTokenizer
+from keras_hub.src.utils.preset_utils import register_presets
+
+register_presets(backbone_presets, (MistralBackbone, MistralTokenizer))