keras-hub-nightly 0.15.0.dev20240823171555__py3-none-any.whl → 0.16.0.dev2024092017__py3-none-any.whl

keras_hub/src/models/llama/llama_causal_lm_preprocessor.py

@@ -12,19 +12,15 @@
 # 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.llama.llama_preprocessor import LlamaPreprocessor
-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
+from keras_hub.src.models.causal_lm_preprocessor import CausalLMPreprocessor
+from keras_hub.src.models.llama.llama_backbone import LlamaBackbone
+from keras_hub.src.models.llama.llama_tokenizer import LlamaTokenizer
 
 
 @keras_hub_export("keras_hub.models.LlamaCausalLMPreprocessor")
-class LlamaCausalLMPreprocessor(LlamaPreprocessor):
+class LlamaCausalLMPreprocessor(CausalLMPreprocessor):
     """Llama Causal LM preprocessor.
 
     This preprocessing layer is meant for use with
@@ -91,80 +87,5 @@ class LlamaCausalLMPreprocessor(LlamaPreprocessor):
     ```
     """
 
-    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(
-                "`LlamaCausalLMPreprocessor` 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,)
-        token_ids = strip_to_ragged(token_ids, padding_mask, ids_to_strip)
-        return self.tokenizer.detokenize(token_ids)
+    backbone_cls = LlamaBackbone
+    tokenizer_cls = LlamaTokenizer

keras_hub/src/models/llama/llama_tokenizer.py

@@ -13,12 +13,18 @@
 # limitations under the License.
 
 from keras_hub.src.api_export import keras_hub_export
+from keras_hub.src.models.llama.llama_backbone import LlamaBackbone
 from keras_hub.src.tokenizers.sentence_piece_tokenizer import (
     SentencePieceTokenizer,
 )
 
 
-@keras_hub_export("keras_hub.models.LlamaTokenizer")
+@keras_hub_export(
+    [
+        "keras_hub.tokenizers.LlamaTokenizer",
+        "keras_hub.models.LlamaTokenizer",
+    ]
+)
 class LlamaTokenizer(SentencePieceTokenizer):
     """Llama tokenizer layer based on SentencePiece.
 
@@ -28,10 +34,6 @@ class LlamaTokenizer(SentencePieceTokenizer):
     Llama models and provides a `from_preset()` method to automatically
     download a matching vocabulary for a Llama preset.
 
-    This tokenizer does not provide truncation or padding of inputs. It can be
-    combined with a `keras_hub.models.LlamaPreprocessor` 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.
 
@@ -60,25 +62,10 @@ class LlamaTokenizer(SentencePieceTokenizer):
     ```
     """
 
+    backbone_cls = LlamaBackbone
+
     def __init__(self, proto, **kwargs):
-        self.start_token = "<s>"
-        self.end_token = "</s>"
+        self._add_special_token("<s>", "start_token")
+        self._add_special_token("</s>", "end_token")
+        self.pad_token_id = 0
         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)
-            self.pad_token_id = 0
-        else:
-            self.start_token_id = None
-            self.end_token_id = None
-            self.pad_token_id = None

keras_hub/src/models/llama3/__init__.py

@@ -14,7 +14,6 @@
 
 from keras_hub.src.models.llama3.llama3_backbone import Llama3Backbone
 from keras_hub.src.models.llama3.llama3_presets import backbone_presets
-from keras_hub.src.models.llama3.llama3_tokenizer import Llama3Tokenizer
 from keras_hub.src.utils.preset_utils import register_presets
 
-register_presets(backbone_presets, (Llama3Backbone, Llama3Tokenizer))
+register_presets(backbone_presets, Llama3Backbone)

keras_hub/src/models/llama3/llama3_causal_lm_preprocessor.py

@@ -12,19 +12,14 @@
 # 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
+from keras_hub.src.models.causal_lm_preprocessor import CausalLMPreprocessor
+from keras_hub.src.models.llama3.llama3_backbone import Llama3Backbone
+from keras_hub.src.models.llama3.llama3_tokenizer import Llama3Tokenizer
 
 
 @keras_hub_export("keras_hub.models.Llama3CausalLMPreprocessor")
-class Llama3CausalLMPreprocessor(Llama3Preprocessor):
+class Llama3CausalLMPreprocessor(CausalLMPreprocessor):
     """Llama 3 Causal LM preprocessor.
 
     This preprocessing layer is meant for use with
@@ -91,83 +86,5 @@ class Llama3CausalLMPreprocessor(Llama3Preprocessor):
     ```
     """
 
-    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)
+    backbone_cls = Llama3Backbone
+    tokenizer_cls = Llama3Tokenizer

keras_hub/src/models/llama3/llama3_tokenizer.py

@@ -13,51 +13,30 @@
 # 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.tokenizers.byte_pair_tokenizer import BytePairTokenizer
 
 
-@keras_hub_export("keras_hub.models.Llama3Tokenizer")
+@keras_hub_export(
+    [
+        "keras_hub.tokenizers.Llama3Tokenizer",
+        "keras_hub.models.Llama3Tokenizer",
+    ]
+)
 class Llama3Tokenizer(BytePairTokenizer):
+    backbone_cls = Llama3Backbone
+
     def __init__(
         self,
         vocabulary=None,
         merges=None,
         **kwargs,
     ):
-        self.start_token = "<|begin_of_text|>"
-        self.end_token = "<|end_of_text|>"
-
+        self._add_special_token("<|begin_of_text|>", "start_token")
+        self._add_special_token("<|end_of_text|>", "end_token")
+        self.pad_token_id = 0
         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

@@ -45,8 +45,6 @@ class MaskedLM(Task):
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        # Default compilation.
-        self.compile()
 
     def compile(
         self,

keras_hub/src/models/masked_lm_preprocessor.py

@@ -0,0 +1,156 @@
+# 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.masked_lm_mask_generator import (
+    MaskedLMMaskGenerator,
+)
+from keras_hub.src.layers.preprocessing.multi_segment_packer import (
+    MultiSegmentPacker,
+)
+from keras_hub.src.models.preprocessor import Preprocessor
+from keras_hub.src.utils.tensor_utils import preprocessing_function
+
+
+@keras_hub_export("keras_hub.models.MaskedLMPreprocessor")
+class MaskedLMPreprocessor(Preprocessor):
+    """Base class for masked language modeling preprocessing layers.
+
+    `MaskedLMPreprocessor` tasks wrap a `keras_hub.tokenizer.Tokenizer` to
+    create a preprocessing layer for masked language modeling tasks. It is
+    intended to be paired with a `keras.models.MaskedLM` task.
+
+    All `MaskedLMPreprocessor` take inputs a single input. This can be a single
+    string, a batch of strings, or a tuple of batches of string segments that
+    should be combined into a single sequence. See examples below. These inputs
+    will be tokenized, combined, and masked randomly along the sequence.
+
+    This layer will always output a `(x, y, sample_weight)` tuple, where `x`
+    is a dictionary with the masked, tokenized inputs, `y` contains the tokens
+    that were masked in `x`, and `sample_weight` marks where `y` contains padded
+    values. The exact contents of `x` will vary depending on the model being
+    used.
+
+    All `MaskedLMPreprocessor` tasks include a `from_preset()` constructor
+    which can be used to load a pre-trained config and vocabularies. You can
+    call the `from_preset()` constructor directly on this base class, in which
+    case the correct class for you model will be automatically instantiated.
+
+    Examples.
+    ```python
+    preprocessor = keras_hub.models.MaskedLMPreprocessor.from_preset(
+        "bert_base_en_uncased",
+        sequence_length=256, # Optional.
+    )
+
+    # Tokenize, mask and pack a single sentence.
+    x = "The quick brown fox jumped."
+    x, y, sample_weight = preprocessor(x)
+
+    # Preprocess a batch of labeled sentence pairs.
+    first = ["The quick brown fox jumped.", "Call me Ishmael."]
+    second = ["The fox tripped.", "Oh look, a whale."]
+    x, y, sample_weight = preprocessor((first, second))
+
+    # With a `tf.data.Dataset`.
+    ds = tf.data.Dataset.from_tensor_slices((first, second))
+    ds = ds.map(preprocessor, num_parallel_calls=tf.data.AUTOTUNE)
+    ```
+    """
+
+    def __init__(
+        self,
+        tokenizer,
+        sequence_length=512,
+        truncate="round_robin",
+        mask_selection_rate=0.15,
+        mask_selection_length=96,
+        mask_token_rate=0.8,
+        random_token_rate=0.1,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.tokenizer = tokenizer
+        self.packer = None
+        self.sequence_length = sequence_length
+        self.truncate = truncate
+        self.mask_selection_rate = mask_selection_rate
+        self.mask_selection_length = mask_selection_length
+        self.mask_token_rate = mask_token_rate
+        self.random_token_rate = random_token_rate
+        self.masker = None
+
+    def build(self, input_shape):
+        super().build(input_shape)
+        # Defer masker 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,
+            pad_value=self.tokenizer.pad_token_id,
+            truncate=self.truncate,
+            sequence_length=self.sequence_length,
+        )
+        self.masker = MaskedLMMaskGenerator(
+            mask_selection_rate=self.mask_selection_rate,
+            mask_selection_length=self.mask_selection_length,
+            mask_token_rate=self.mask_token_rate,
+            random_token_rate=self.random_token_rate,
+            vocabulary_size=self.tokenizer.vocabulary_size(),
+            mask_token_id=self.tokenizer.mask_token_id,
+            unselectable_token_ids=self.tokenizer.special_token_ids,
+        )
+
+    @preprocessing_function
+    def call(self, x, y=None, sample_weight=None):
+        x = x if isinstance(x, tuple) else (x,)
+        x = tuple(self.tokenizer(segment) for segment in x)
+        token_ids, segment_ids = self.packer(x)
+        padding_mask = token_ids != self.tokenizer.pad_token_id
+        masker_outputs = self.masker(token_ids)
+        x = {
+            "token_ids": masker_outputs["token_ids"],
+            "padding_mask": padding_mask,
+            "segment_ids": segment_ids,
+            "mask_positions": masker_outputs["mask_positions"],
+        }
+        y = masker_outputs["mask_ids"]
+        sample_weight = masker_outputs["mask_weights"]
+        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,
+                "truncate": self.truncate,
+                "mask_selection_rate": self.mask_selection_rate,
+                "mask_selection_length": self.mask_selection_length,
+                "mask_token_rate": self.mask_token_rate,
+                "random_token_rate": self.random_token_rate,
+            }
+        )
+        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/mistral/__init__.py

@@ -14,7 +14,6 @@
 
 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))
+register_presets(backbone_presets, MistralBackbone)

keras_hub/src/models/mistral/mistral_causal_lm_preprocessor.py

@@ -12,21 +12,14 @@
 # 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.mistral.mistral_preprocessor import (
-    MistralPreprocessor,
-)
-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
+from keras_hub.src.models.causal_lm_preprocessor import CausalLMPreprocessor
+from keras_hub.src.models.mistral.mistral_backbone import MistralBackbone
+from keras_hub.src.models.mistral.mistral_tokenizer import MistralTokenizer
 
 
 @keras_hub_export("keras_hub.models.MistralCausalLMPreprocessor")
-class MistralCausalLMPreprocessor(MistralPreprocessor):
+class MistralCausalLMPreprocessor(CausalLMPreprocessor):
     """Mistral Causal LM preprocessor.
 
     This preprocessing layer is meant for use with
@@ -93,83 +86,5 @@ class MistralCausalLMPreprocessor(MistralPreprocessor):
     ```
     """
 
-    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(
-                "`MistralCausalLMPreprocessor` 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.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)
+    backbone_cls = MistralBackbone
+    tokenizer_cls = MistralTokenizer