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

keras_hub/src/models/bart/bart_seq_2_seq_lm_preprocessor.py

@@ -0,0 +1,262 @@
+# 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.bart.bart_preprocessor import BartPreprocessor
+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
+
+try:
+    import tensorflow as tf
+except ImportError:
+    tf = None
+
+
+@keras_hub_export("keras_hub.models.BartSeq2SeqLMPreprocessor")
+class BartSeq2SeqLMPreprocessor(BartPreprocessor):
+    """BART Seq2Seq LM preprocessor.
+
+    This layer is used as preprocessor for seq2seq tasks using the BART model.
+    This class subclasses `keras_hub.models.BartPreprocessor` and keeps most of
+    its functionality. It has two changes from the superclass:
+
+     1. Sets the `y` (label) and `sample_weights` fields by shifting the
+        decoder input sequence one step towards the left. Both these fields are
+        inferred internally, and any passed values will be ignored.
+     2. Drops the last token from the decoder input sequence as it does not have
+        a successor.
+
+    Args:
+        tokenizer: A `keras_hub.models.BartTokenizer` instance.
+        encoder_sequence_length: The length of the packed encoder inputs.
+        decoder_sequence_length: The length of the packed decoder inputs.
+
+    Call arguments:
+        x: A dictionary with `encoder_text` and `decoder_text` as its keys.
+            Each value in the dictionary should be a tensor of single string
+            sequences. Inputs may be batched or unbatched. Raw python inputs
+            will be converted to tensors.
+        y: Label data. Should always be `None` as the layer generates labels by
+            shifting the decoder input sequence one step to the left.
+        sample_weight: Label weights. Should always be `None` as the layer
+            generates label weights by shifting the padding mask one step to the
+            left.
+
+    Examples:
+
+    Directly calling the layer on data
+    ```python
+    preprocessor = keras_hub.models.BartPreprocessor.from_preset("bart_base_en")
+
+    # Preprocess unbatched inputs.
+    inputs = {
+        "encoder_text": "The fox was sleeping.",
+        "decoder_text": "The fox was awake."
+    }
+    preprocessor(inputs)
+
+    # Preprocess batched inputs.
+    inputs = {
+        "encoder_text": ["The fox was sleeping.", "The lion was quiet."],
+        "decoder_text": ["The fox was awake.", "The lion was roaring."]
+    }
+    preprocessor(inputs)
+
+    # Custom vocabulary.
+    vocab = {
+        "<s>": 0,
+        "<pad>": 1,
+        "</s>": 2,
+        "Ġafter": 5,
+        "noon": 6,
+        "Ġsun": 7,
+    }
+    merges = ["Ġ a", "Ġ s", "Ġ n", "e r", "n o", "o n", "Ġs u", "Ġa f", "no on"]
+    merges += ["Ġsu n", "Ġaf t", "Ġaft er"]
+
+    tokenizer = keras_hub.models.BartTokenizer(
+        vocabulary=vocab,
+        merges=merges,
+    )
+    preprocessor = keras_hub.models.BartPreprocessor(
+        tokenizer=tokenizer,
+        encoder_sequence_length=20,
+        decoder_sequence_length=10,
+    )
+    inputs = {
+        "encoder_text": "The fox was sleeping.",
+        "decoder_text": "The fox was awake."
+    }
+    preprocessor(inputs)
+    ```
+
+    Mapping with `tf.data.Dataset`.
+    ```python
+    preprocessor = keras_hub.models.BartPreprocessor.from_preset("bart_base_en")
+
+    # Map single sentences.
+    features = {
+        "encoder_text": tf.constant(
+            ["The fox was sleeping.", "The lion was quiet."]
+        ),
+        "decoder_text": tf.constant(
+            ["The fox was awake.", "The lion was roaring."]
+        )
+    }
+    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,
+        *,
+        encoder_sequence_length=None,
+        decoder_sequence_length=None,
+        # `sequence_length` is an alias for `decoder_sequence_length`
+        sequence_length=None,
+    ):
+        if y is not None or sample_weight is not None:
+            logging.warning(
+                "`BartSeq2SeqLMPreprocessor` infers `y` and `sample_weight` "
+                "from the provided input data, i.e., `x`. However, non-`None`"
+                "values have been passed for `y` or `sample_weight` or both. "
+                "These values will be ignored."
+            )
+
+        if encoder_sequence_length is None:
+            encoder_sequence_length = self.encoder_sequence_length
+        decoder_sequence_length = decoder_sequence_length or sequence_length
+        if decoder_sequence_length is None:
+            decoder_sequence_length = self.decoder_sequence_length
+
+        x = super().call(
+            x,
+            encoder_sequence_length=encoder_sequence_length,
+            decoder_sequence_length=decoder_sequence_length + 1,
+        )
+        decoder_token_ids = x.pop("decoder_token_ids")
+        decoder_padding_mask = x.pop("decoder_padding_mask")
+
+        # The last token does not have a next token. Hence, we truncate it.
+        x = {
+            **x,
+            "decoder_token_ids": decoder_token_ids[..., :-1],
+            "decoder_padding_mask": decoder_padding_mask[..., :-1],
+        }
+        # Target `y` will be the decoder input sequence shifted one step to the
+        # left (i.e., the next token).
+        y = decoder_token_ids[..., 1:]
+        sample_weight = decoder_padding_mask[..., 1:]
+        return keras.utils.pack_x_y_sample_weight(x, y, sample_weight)
+
+    def generate_preprocess(
+        self,
+        x,
+        *,
+        encoder_sequence_length=None,
+        # `sequence_length` is an alias for `decoder_sequence_length`
+        decoder_sequence_length=None,
+        sequence_length=None,
+    ):
+        """Convert encoder and decoder input strings to integer token inputs for generation.
+
+        Similar to calling the layer for training, this method takes in a dict
+        containing `"encoder_text"` and `"decoder_text"`, with strings or tensor
+        strings for values, 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 decoder sequence (as generation is expected to continue at the end
+        of the inputted decoder prompt).
+        """
+        if not self.built:
+            self.build(None)
+
+        if isinstance(x, dict):
+            encoder_text = x["encoder_text"]
+            decoder_text = x["decoder_text"]
+        else:
+            encoder_text = x
+            # Initialize empty prompt for the decoder.
+            decoder_text = tf.fill((tf.shape(encoder_text)[0],), "")
+
+        if encoder_sequence_length is None:
+            encoder_sequence_length = self.encoder_sequence_length
+        decoder_sequence_length = decoder_sequence_length or sequence_length
+        if decoder_sequence_length is None:
+            decoder_sequence_length = self.decoder_sequence_length
+
+        # Tokenize and pack the encoder inputs.
+        # TODO: Remove `[0]` once we have shifted to `MultiSegmentPacker`.
+        encoder_text = convert_inputs_to_list_of_tensor_segments(encoder_text)[
+            0
+        ]
+        encoder_token_ids = self.tokenizer(encoder_text)
+        encoder_token_ids, encoder_padding_mask = self.encoder_packer(
+            encoder_token_ids,
+            sequence_length=encoder_sequence_length,
+        )
+
+        # Tokenize and pack the decoder inputs.
+        decoder_text = convert_inputs_to_list_of_tensor_segments(decoder_text)[
+            0
+        ]
+        decoder_token_ids = self.tokenizer(decoder_text)
+        decoder_token_ids, decoder_padding_mask = self.decoder_packer(
+            decoder_token_ids,
+            sequence_length=decoder_sequence_length,
+            add_end_value=False,
+        )
+
+        return {
+            "encoder_token_ids": encoder_token_ids,
+            "encoder_padding_mask": encoder_padding_mask,
+            "decoder_token_ids": decoder_token_ids,
+            "decoder_padding_mask": decoder_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.
+        """
+        if not self.built:
+            self.build(None)
+
+        token_ids, padding_mask = (
+            x["decoder_token_ids"],
+            x["decoder_padding_mask"],
+        )
+        ids_to_strip = (
+            self.tokenizer.start_token_id,
+            self.tokenizer.end_token_id,
+        )
+        token_ids = strip_to_ragged(token_ids, padding_mask, ids_to_strip)
+        return self.tokenizer.detokenize(token_ids)

keras_hub/src/models/bart/bart_tokenizer.py

@@ -0,0 +1,124 @@
+# 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.BartTokenizer")
+class BartTokenizer(BytePairTokenizer):
+    """A BART 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 BART
+    models and provides a `from_preset()` method to automatically download
+    a matching vocabulary for a BART preset.
+
+    This tokenizer does not provide truncation or padding of inputs. It can be
+    combined with a `keras_hub.models.BartPreprocessor` layer for input
+    packing.
+
+    If input is a batch of strings (rank > 0), the layer will output a
+    `tf.RaggedTensor` where the last dimension of the output is ragged.
+
+    If input is a scalar string (rank == 0), the layer will output a dense
+    `tf.Tensor` with static shape `[None]`.
+
+    Args:
+        vocabulary: 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.BartTokenizer.from_preset(
+        "bart_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 = {"<s>": 0, "<pad>": 1, "</s>": 2, "<mask>": 3}
+    vocab = {**vocab, "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.BartTokenizer(
+        vocabulary=vocab,
+        merges=merges,
+    )
+    tokenizer("The quick brown fox jumped.")
+    ```
+    """
+
+    def __init__(
+        self,
+        vocabulary=None,
+        merges=None,
+        **kwargs,
+    ):
+        self.start_token = "<s>"
+        self.pad_token = "<pad>"
+        self.end_token = "</s>"
+
+        super().__init__(
+            vocabulary=vocabulary,
+            merges=merges,
+            unsplittable_tokens=[
+                self.start_token,
+                self.pad_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.
+            for token in [self.start_token, self.pad_token, self.end_token]:
+                if token not in self.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.pad_token_id = self.token_to_id(self.pad_token)
+            self.end_token_id = self.token_to_id(self.end_token)
+        else:
+            self.start_token_id = None
+            self.pad_token_id = None
+            self.end_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/bert/__init__.py

@@ -0,0 +1,23 @@
+# 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.bert.bert_backbone import BertBackbone
+from keras_hub.src.models.bert.bert_classifier import BertClassifier
+from keras_hub.src.models.bert.bert_presets import backbone_presets
+from keras_hub.src.models.bert.bert_presets import classifier_presets
+from keras_hub.src.models.bert.bert_tokenizer import BertTokenizer
+from keras_hub.src.utils.preset_utils import register_presets
+
+register_presets(backbone_presets, (BertBackbone, BertTokenizer))
+register_presets(classifier_presets, (BertClassifier, BertTokenizer))

keras_hub/src/models/bert/bert_backbone.py

@@ -0,0 +1,227 @@
+# Copyright 2024 The KerasHub Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import keras
+
+from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.layers.modeling.position_embedding import PositionEmbedding
+from keras_hub.src.layers.modeling.reversible_embedding import (
+    ReversibleEmbedding,
+)
+from keras_hub.src.layers.modeling.transformer_encoder import TransformerEncoder
+from keras_hub.src.models.backbone import Backbone
+from keras_hub.src.utils.keras_utils import gelu_approximate
+
+
+def bert_kernel_initializer(stddev=0.02):
+    return keras.initializers.TruncatedNormal(stddev=stddev)
+
+
+@keras_hub_export("keras_hub.models.BertBackbone")
+class BertBackbone(Backbone):
+    """A BERT encoder network.
+
+    This class implements a bi-directional Transformer-based encoder as
+    described in ["BERT: Pre-training of Deep Bidirectional Transformers for
+    Language Understanding"](https://arxiv.org/abs/1810.04805). It includes the
+    embedding lookups and transformer layers, but not the masked language model
+    or next sentence prediction heads.
+
+    The default constructor gives a fully customizable, randomly initialized
+    BERT encoder with any number of layers, heads, and embedding dimensions. To
+    load preset architectures and weights, use the `from_preset()` constructor.
+
+    Disclaimer: Pre-trained models are provided on an "as is" basis, without
+    warranties or conditions of any kind.
+
+    Args:
+        vocabulary_size: int. The size of the token vocabulary.
+        num_layers: int. The number of transformer layers.
+        num_heads: int. The number of attention heads for each transformer.
+            The hidden size must be divisible by the number of attention heads.
+        hidden_dim: int. The size of the transformer encoding and pooler layers.
+        intermediate_dim: int. The output dimension of the first Dense layer in
+            a two-layer feedforward network for each transformer.
+        dropout: float. Dropout probability for the Transformer encoder.
+        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.
+        num_segments: int. The number of types that the 'segment_ids' input can
+            take.
+        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"),
+        "segment_ids": np.array([[0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0]]),
+        "padding_mask": np.array([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0]]),
+    }
+
+    # Pretrained BERT encoder.
+    model = keras_hub.models.BertBackbone.from_preset("bert_base_en_uncased")
+    model(input_data)
+
+    # Randomly initialized BERT encoder with a custom config.
+    model = keras_hub.models.BertBackbone(
+        vocabulary_size=30552,
+        num_layers=4,
+        num_heads=4,
+        hidden_dim=256,
+        intermediate_dim=512,
+        max_sequence_length=128,
+    )
+    model(input_data)
+    ```
+    """
+
+    def __init__(
+        self,
+        vocabulary_size,
+        num_layers,
+        num_heads,
+        hidden_dim,
+        intermediate_dim,
+        dropout=0.1,
+        max_sequence_length=512,
+        num_segments=2,
+        dtype=None,
+        **kwargs,
+    ):
+        # === Layers ===
+        self.token_embedding = ReversibleEmbedding(
+            input_dim=vocabulary_size,
+            output_dim=hidden_dim,
+            embeddings_initializer=bert_kernel_initializer(),
+            dtype=dtype,
+            name="token_embedding",
+        )
+        self.position_embedding = PositionEmbedding(
+            initializer=bert_kernel_initializer(),
+            sequence_length=max_sequence_length,
+            dtype=dtype,
+            name="position_embedding",
+        )
+        self.segment_embedding = keras.layers.Embedding(
+            input_dim=num_segments,
+            output_dim=hidden_dim,
+            embeddings_initializer=bert_kernel_initializer(),
+            dtype=dtype,
+            name="segment_embedding",
+        )
+        self.embeddings_add = keras.layers.Add(
+            dtype=dtype,
+            name="embeddings_add",
+        )
+        self.embeddings_layer_norm = keras.layers.LayerNormalization(
+            axis=-1,
+            epsilon=1e-12,
+            dtype=dtype,
+            name="embeddings_layer_norm",
+        )
+        self.embeddings_dropout = keras.layers.Dropout(
+            dropout,
+            dtype=dtype,
+            name="embeddings_dropout",
+        )
+        self.transformer_layers = []
+        for i in range(num_layers):
+            layer = TransformerEncoder(
+                num_heads=num_heads,
+                intermediate_dim=intermediate_dim,
+                activation=gelu_approximate,
+                dropout=dropout,
+                layer_norm_epsilon=1e-12,
+                kernel_initializer=bert_kernel_initializer(),
+                dtype=dtype,
+                name=f"transformer_layer_{i}",
+            )
+            self.transformer_layers.append(layer)
+        self.pooled_dense = keras.layers.Dense(
+            hidden_dim,
+            kernel_initializer=bert_kernel_initializer(),
+            activation="tanh",
+            dtype=dtype,
+            name="pooled_dense",
+        )
+
+        # === Functional Model ===
+        token_id_input = keras.Input(
+            shape=(None,), dtype="int32", name="token_ids"
+        )
+        segment_id_input = keras.Input(
+            shape=(None,), dtype="int32", name="segment_ids"
+        )
+        padding_mask_input = keras.Input(
+            shape=(None,), dtype="int32", name="padding_mask"
+        )
+        # Embed tokens, positions, and segment ids.
+        tokens = self.token_embedding(token_id_input)
+        positions = self.position_embedding(tokens)
+        segments = self.segment_embedding(segment_id_input)
+        # Sum, normalize and apply dropout to embeddings.
+        x = self.embeddings_add((tokens, positions, segments))
+        x = self.embeddings_layer_norm(x)
+        x = self.embeddings_dropout(x)
+        for transformer_layer in self.transformer_layers:
+            x = transformer_layer(x, padding_mask=padding_mask_input)
+        # Construct the two BERT outputs. The pooled output is a dense layer on
+        # top of the [CLS] token.
+        sequence_output = x
+        cls_token_index = 0
+        pooled_output = self.pooled_dense(x[:, cls_token_index, :])
+        super().__init__(
+            inputs={
+                "token_ids": token_id_input,
+                "segment_ids": segment_id_input,
+                "padding_mask": padding_mask_input,
+            },
+            outputs={
+                "sequence_output": sequence_output,
+                "pooled_output": pooled_output,
+            },
+            dtype=dtype,
+            **kwargs,
+        )
+
+        # === Config ===
+        self.vocabulary_size = vocabulary_size
+        self.num_layers = num_layers
+        self.num_heads = num_heads
+        self.hidden_dim = hidden_dim
+        self.intermediate_dim = intermediate_dim
+        self.dropout = dropout
+        self.max_sequence_length = max_sequence_length
+        self.num_segments = num_segments
+        self.cls_token_index = cls_token_index
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "vocabulary_size": self.vocabulary_size,
+                "num_layers": self.num_layers,
+                "num_heads": self.num_heads,
+                "hidden_dim": self.hidden_dim,
+                "intermediate_dim": self.intermediate_dim,
+                "dropout": self.dropout,
+                "max_sequence_length": self.max_sequence_length,
+                "num_segments": self.num_segments,
+            }
+        )
+        return config