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

keras_hub/src/models/xlm_roberta/xlm_roberta_preprocessor.py

@@ -0,0 +1,205 @@
+# 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.multi_segment_packer import (
+    MultiSegmentPacker,
+)
+from keras_hub.src.models.preprocessor import Preprocessor
+from keras_hub.src.models.xlm_roberta.xlm_roberta_tokenizer import (
+    XLMRobertaTokenizer,
+)
+from keras_hub.src.utils.keras_utils import (
+    convert_inputs_to_list_of_tensor_segments,
+)
+
+
+@keras_hub_export("keras_hub.models.XLMRobertaPreprocessor")
+class XLMRobertaPreprocessor(Preprocessor):
+    """An XLM-RoBERTa 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.MultiSegmentPacker`.
+      with the appropriate `"<s>"`, `"</s>"` and `"<pad>"` tokens, i.e., adding
+      a single `"<s>"` at the start of the entire sequence, `"</s></s>"` at the
+      end of each segment, save the last and a `"</s>"` at the end of the
+      entire sequence.
+    3. Construct a dictionary with keys `"token_ids"` and `"padding_mask"`,
+      that can be passed directly to an XLM-RoBERTa model.
+
+    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.tokenizers.XLMRobertaTokenizer` instance.
+        sequence_length: The length of the packed inputs.
+        truncate: The algorithm to truncate a list of batched segments to fit
+            within `sequence_length`. The value can be either `round_robin` or
+            `waterfall`:
+                - `"round_robin"`: Available space is assigned one token at a
+                    time in a round-robin fashion to the inputs that still need
+                    some, until the limit is reached.
+                - `"waterfall"`: The allocation of the budget is done using a
+                    "waterfall" algorithm that allocates quota in a
+                    left-to-right manner and fills up the buckets until we run
+                    out of budget. It supports an arbitrary number of segments.
+
+    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.
+
+    Examples:
+
+    Directly calling the layer on data.
+    ```python
+    preprocessor = keras_hub.models.XLMRobertaPreprocessor.from_preset(
+        "xlm_roberta_base_multi"
+    )
+
+    # Tokenize and pack a single sentence.
+    preprocessor("The quick brown fox jumped.")
+
+    # Tokenize a batch of single sentences.
+    preprocessor(["The quick brown fox jumped.", "اسمي اسماعيل"])
+
+    # Preprocess a batch of sentence pairs.
+    # When handling multiple sequences, always convert to tensors first!
+    first = tf.constant(["The quick brown fox jumped.", "اسمي اسماعيل"])
+    second = tf.constant(["The fox tripped.", "الأسد ملك الغابة"])
+    preprocessor((first, second))
+
+    # Custom vocabulary.
+    def train_sentencepiece(ds, vocab_size):
+        bytes_io = io.BytesIO()
+        sentencepiece.SentencePieceTrainer.train(
+            sentence_iterator=ds.as_numpy_iterator(),
+            model_writer=bytes_io,
+            vocab_size=vocab_size,
+            model_type="WORD",
+            unk_id=0,
+            bos_id=1,
+            eos_id=2,
+        )
+        return bytes_io.getvalue()
+    ds = tf.data.Dataset.from_tensor_slices(
+        ["the quick brown fox", "the earth is round"]
+    )
+    proto = train_sentencepiece(ds, vocab_size=10)
+    tokenizer = keras_hub.models.XLMRobertaTokenizer(proto=proto)
+    preprocessor = keras_hub.models.XLMRobertaPreprocessor(tokenizer)
+    preprocessor("The quick brown fox jumped.")
+    ```
+
+    Mapping with `tf.data.Dataset`.
+    ```python
+    preprocessor = keras_hub.models.XLMRobertaPreprocessor.from_preset(
+        "xlm_roberta_base_multi"
+    )
+
+    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 = XLMRobertaTokenizer
+
+    def __init__(
+        self,
+        tokenizer,
+        sequence_length=512,
+        truncate="round_robin",
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.tokenizer = tokenizer
+        self.packer = None
+        self.truncate = truncate
+        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 = MultiSegmentPacker(
+            start_value=self.tokenizer.start_token_id,
+            end_value=self.tokenizer.end_token_id,
+            sep_value=[self.tokenizer.end_token_id] * 2,
+            pad_value=self.tokenizer.pad_token_id,
+            truncate=self.truncate,
+            sequence_length=self.sequence_length,
+        )
+        self.built = True
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "sequence_length": self.sequence_length,
+                "truncate": self.truncate,
+            }
+        )
+        return config
+
+    def call(self, x, y=None, sample_weight=None):
+        x = convert_inputs_to_list_of_tensor_segments(x)
+        x = [self.tokenizer(segment) for segment in x]
+        token_ids, _ = self.packer(x)
+        x = {
+            "token_ids": token_ids,
+            "padding_mask": token_ids != self.tokenizer.pad_token_id,
+        }
+        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/xlm_roberta/xlm_roberta_presets.py

@@ -0,0 +1,43 @@
+# 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.
+"""XLM-RoBERTa model preset configurations."""
+
+backbone_presets = {
+    "xlm_roberta_base_multi": {
+        "metadata": {
+            "description": (
+                "12-layer XLM-RoBERTa model where case is maintained. "
+                "Trained on CommonCrawl in 100 languages."
+            ),
+            "params": 277450752,
+            "official_name": "XLM-RoBERTa",
+            "path": "xlm_roberta",
+            "model_card": "https://github.com/facebookresearch/fairseq/blob/main/examples/xlmr/README.md",
+        },
+        "kaggle_handle": "kaggle://keras/xlm_roberta/keras/xlm_roberta_base_multi/2",
+    },
+    "xlm_roberta_large_multi": {
+        "metadata": {
+            "description": (
+                "24-layer XLM-RoBERTa model where case is maintained. "
+                "Trained on CommonCrawl in 100 languages."
+            ),
+            "params": 558837760,
+            "official_name": "XLM-RoBERTa",
+            "path": "xlm_roberta",
+            "model_card": "https://github.com/facebookresearch/fairseq/blob/main/examples/xlmr/README.md",
+        },
+        "kaggle_handle": "kaggle://keras/xlm_roberta/keras/xlm_roberta_large_multi/2",
+    },
+}

keras_hub/src/models/xlm_roberta/xlm_roberta_tokenizer.py

@@ -0,0 +1,191 @@
+# 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,
+)
+from keras_hub.src.utils.tensor_utils import tensor_to_list
+
+try:
+    import tensorflow as tf
+except ImportError:
+    tf = None
+
+
+@keras_hub_export("keras_hub.models.XLMRobertaTokenizer")
+class XLMRobertaTokenizer(SentencePieceTokenizer):
+    """An XLM-RoBERTa tokenizer using SentencePiece subword segmentation.
+
+    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
+    XLM-RoBERTa models and provides a `from_preset()` method to automatically
+    download a matching vocabulary for an XLM-RoBERTa preset.
+
+    Note: If you are providing your own custom SentencePiece model, the original
+    fairseq implementation of XLM-RoBERTa re-maps some token indices from the
+    underlying sentencepiece output. To preserve compatibility, we do the same
+    re-mapping here.
+
+    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
+    tokenizer = keras_hub.models.XLMRobertaTokenizer.from_preset(
+        "xlm_roberta_base_multi",
+    )
+
+    # Unbatched inputs.
+    tokenizer("the quick brown fox")
+
+    # Batched inputs.
+    tokenizer(["the quick brown fox", "الأرض كروية"])
+
+    # Detokenization.
+    tokenizer.detokenize(tokenizer("the quick brown fox"))
+
+    # Custom vocabulary
+    def train_sentencepiece(ds, vocab_size):
+        bytes_io = io.BytesIO()
+        sentencepiece.SentencePieceTrainer.train(
+            sentence_iterator=ds.as_numpy_iterator(),
+            model_writer=bytes_io,
+            vocab_size=vocab_size,
+            model_type="WORD",
+            unk_id=0,
+            bos_id=1,
+            eos_id=2,
+        )
+        return bytes_io.getvalue()
+
+    ds = tf.data.Dataset.from_tensor_slices(
+        ["the quick brown fox", "the earth is round"]
+    )
+    proto = train_sentencepiece(ds, vocab_size=10)
+    tokenizer = keras_hub.models.XLMRobertaTokenizer(proto=proto)
+    ```
+    """
+
+    def __init__(self, proto, **kwargs):
+        # List of special tokens.
+        self._vocabulary_prefix = ["<s>", "<pad>", "</s>", "<unk>"]
+
+        # IDs of special tokens.
+        self.start_token_id = 0  # <s>
+        self.pad_token_id = 1  # <pad>
+        self.end_token_id = 2  # </s>
+        self.unk_token_id = 3  # <unk>
+
+        super().__init__(proto=proto, **kwargs)
+
+    def set_proto(self, proto):
+        super().set_proto(proto)
+        if proto is not None:
+            self.mask_token_id = self.vocabulary_size() - 1
+        else:
+            self.mask_token_id = None
+
+    def vocabulary_size(self):
+        """Get the size of the tokenizer vocabulary."""
+        return super().vocabulary_size() + 2
+
+    def get_vocabulary(self):
+        """Get the size of the tokenizer vocabulary."""
+        self._check_vocabulary()
+        vocabulary = tensor_to_list(
+            self._sentence_piece.id_to_string(
+                tf.range(super().vocabulary_size())
+            )
+        )
+        return self._vocabulary_prefix + vocabulary[3:] + ["<mask>"]
+
+    def id_to_token(self, id):
+        """Convert an integer id to a string token."""
+        self._check_vocabulary()
+
+        if id == self.mask_token_id:
+            return "<mask>"
+
+        if id < len(self._vocabulary_prefix) and id >= 0:
+            return self._vocabulary_prefix[id]
+
+        if id - 1 >= super().vocabulary_size() or id - 1 < 0:
+            raise ValueError(
+                f"`id` must be in range [0, {self.vocabulary_size() - 1}]. "
+                f"Received: {id}"
+            )
+
+        return tensor_to_list(self._sentence_piece.id_to_string(id - 1))
+
+    def token_to_id(self, token):
+        """Convert a string token to an integer id."""
+        self._check_vocabulary()
+
+        if token in self._vocabulary_prefix:
+            return self._vocabulary_prefix.index(token)
+
+        spm_token_id = self._sentence_piece.string_to_id(token)
+
+        # OOV token
+        spm_unk_token_id = self._sentence_piece.string_to_id("<unk>")
+        if spm_token_id == spm_unk_token_id:
+            return self.unk_token_id
+
+        return int(spm_token_id.numpy()) + 1
+
+    def tokenize(self, inputs):
+        self._check_vocabulary()
+        tokens = super().tokenize(inputs)
+
+        # Correct `unk_token_id` (0 -> 3). Note that we do not correct
+        # `start_token_id` and `end_token_id`; they are dealt with in
+        # `XLMRobertaPreprocessor`.
+        tokens = tf.where(tf.equal(tokens, 0), self.unk_token_id - 1, tokens)
+
+        # Shift the tokens IDs right by one.
+        return tf.add(tokens, 1)
+
+    def detokenize(self, inputs):
+        self._check_vocabulary()
+        tokens = tf.ragged.boolean_mask(
+            inputs, tf.not_equal(inputs, self.mask_token_id)
+        )
+
+        # Shift the tokens IDs left by one.
+        tokens = tf.subtract(tokens, 1)
+
+        # Correct `unk_token_id`, `end_token_id`, `start_token_id`, respectively.
+        # Note: The `pad_token_id` is taken as 0 (`unk_token_id`) since the
+        # proto does not contain `pad_token_id`. This mapping of the pad token
+        # is done automatically by the above subtraction.
+        tokens = tf.where(tf.equal(tokens, self.unk_token_id - 1), 0, tokens)
+        tokens = tf.where(tf.equal(tokens, self.end_token_id - 1), 2, tokens)
+        tokens = tf.where(tf.equal(tokens, self.start_token_id - 1), 1, tokens)
+
+        # Note: Even though we map `"<s>" and `"</s>"` to the correct IDs,
+        # the `detokenize` method will return empty strings for these tokens.
+        # This is a vagary of the `sentencepiece` library.
+        return super().detokenize(tokens)

keras_hub/src/models/xlnet/__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.