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

keras_hub/src/models/stable_diffusion_v3/__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.

keras_hub/src/models/stable_diffusion_v3/clip_encoder_block.py

@@ -0,0 +1,103 @@
+# 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 import layers
+from keras import ops
+
+
+def quick_gelu(x):
+    return x * ops.sigmoid(1.702 * x)
+
+
+class CLIPEncoderBlock(layers.Layer):
+    def __init__(
+        self,
+        hidden_dim,
+        num_heads,
+        intermediate_dim,
+        intermediate_activation="quick_gelu",
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        if hidden_dim % num_heads != 0:
+            raise ValueError(
+                "`hidden_dim` must be divisible by `num_heads`. "
+                f"Received: hidden_dim={hidden_dim}, num_heads={num_heads}"
+            )
+        self.hidden_dim = hidden_dim
+        self.num_heads = num_heads
+        self.intermediate_dim = intermediate_dim
+        self.intermediate_activation = intermediate_activation
+
+        if intermediate_activation == "quick_gelu":
+            intermediate_activation = quick_gelu
+
+        self.layer_norm_1 = layers.LayerNormalization(
+            epsilon=0.00001, dtype=self.dtype_policy, name="layer_norm_1"
+        )
+        self.attention = layers.MultiHeadAttention(
+            num_heads,
+            hidden_dim // num_heads,
+            dtype=self.dtype_policy,
+            name="attention",
+        )
+        self.layer_norm_2 = layers.LayerNormalization(
+            epsilon=0.00001, dtype=self.dtype_policy, name="layer_norm_2"
+        )
+        self.dense_1 = layers.Dense(
+            self.intermediate_dim, dtype=self.dtype_policy, name="dense_1"
+        )
+        self.activation = layers.Activation(
+            intermediate_activation, dtype=self.dtype_policy, name="activation"
+        )
+        self.dense_2 = layers.Dense(
+            self.hidden_dim, dtype=self.dtype_policy, name="dense_2"
+        )
+
+    def build(self, input_shape):
+        self.layer_norm_1.build(input_shape)
+        self.attention.build(input_shape, input_shape, input_shape)
+        self.layer_norm_2.build(input_shape)
+        self.dense_1.build(input_shape)
+        input_shape = self.dense_1.compute_output_shape(input_shape)
+        self.dense_2.build(input_shape)
+
+    def compute_output_shape(self, inputs_shape):
+        outputs_shape = list(inputs_shape)
+        outputs_shape[-1] = self.hidden_dim
+        return outputs_shape
+
+    def call(self, x, training=None):
+        residual = x
+        x = self.layer_norm_1(x)
+        x = self.attention(x, x, x, training=training, use_causal_mask=True)
+        x = ops.add(residual, x)
+
+        residual = x
+        x = self.dense_1(self.layer_norm_2(residual))
+        x = self.activation(x)
+        x = self.dense_2(x)
+        x = ops.add(residual, x)
+        return x
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "hidden_dim": self.hidden_dim,
+                "num_heads": self.num_heads,
+                "intermediate_dim": self.intermediate_dim,
+                "intermediate_activation": self.intermediate_activation,
+            }
+        )
+        return config

keras_hub/src/models/stable_diffusion_v3/clip_preprocessor.py

@@ -0,0 +1,93 @@
+# 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.layers.preprocessing.start_end_packer import StartEndPacker
+from keras_hub.src.models.preprocessor import Preprocessor
+from keras_hub.src.models.stable_diffusion_v3.clip_tokenizer import (
+    CLIPTokenizer,
+)
+from keras_hub.src.utils.tensor_utils import preprocessing_function
+
+try:
+    import tensorflow as tf
+except ImportError:
+    tf = None
+
+
+class CLIPPreprocessor(Preprocessor):
+    tokenizer_cls = CLIPTokenizer
+
+    def __init__(
+        self,
+        tokenizer,
+        sequence_length=77,
+        add_start_token=True,
+        add_end_token=False,
+        to_lower=True,
+        pad_with_end_token=True,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.tokenizer = tokenizer
+        self.sequence_length = sequence_length
+        self.add_start_token = add_start_token
+        self.add_end_token = add_end_token
+        self.to_lower = to_lower
+        self.pad_with_end_token = pad_with_end_token
+
+    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.
+        pad_value = self.tokenizer.pad_token_id
+        if self.pad_with_end_token:
+            pad_value = self.tokenizer.end_token_id
+
+        self.packer = StartEndPacker(
+            start_value=self.tokenizer.start_token_id,
+            end_value=self.tokenizer.end_token_id,
+            pad_value=pad_value,
+            sequence_length=self.sequence_length,
+            return_padding_mask=True,
+        )
+        self.built = True
+
+    @preprocessing_function
+    def call(self, x, y=None, sample_weight=None, sequence_length=None):
+        if self.to_lower:
+            x = tf.strings.lower(x)
+        token_ids, padding_mask = self.packer(
+            self.tokenizer(x),
+            sequence_length=sequence_length or self.sequence_length,
+            add_start_value=self.add_start_token,
+            add_end_value=self.add_end_token,
+        )
+        x = {
+            "token_ids": token_ids,
+            "padding_mask": padding_mask,
+        }
+        return keras.utils.pack_x_y_sample_weight(x, y, sample_weight)
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "sequence_length": self.sequence_length,
+                "add_start_token": self.add_start_token,
+                "add_end_token": self.add_end_token,
+                "to_lower": self.to_lower,
+                "pad_with_end_token": self.pad_with_end_token,
+            }
+        )
+        return config

keras_hub/src/models/stable_diffusion_v3/clip_text_encoder.py

@@ -0,0 +1,149 @@
+# Copyright 2024 The KerasHub Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import keras
+from keras import layers
+from keras import ops
+
+from keras_hub.src.layers.modeling.token_and_position_embedding import (
+    TokenAndPositionEmbedding,
+)
+from keras_hub.src.models.stable_diffusion_v3.clip_encoder_block import (
+    CLIPEncoderBlock,
+)
+
+
+class CLIPTextEncoder(keras.Model):
+    def __init__(
+        self,
+        embedding_dim,
+        hidden_dim,
+        num_layers,
+        num_heads,
+        intermediate_dim,
+        intermediate_activation="quick_gelu",
+        intermediate_output_index=None,
+        vocabulary_size=49408,
+        sequence_length=77,
+        dtype=None,
+        **kwargs,
+    ):
+        if (
+            intermediate_output_index is not None
+            and intermediate_output_index < 0
+        ):
+            intermediate_output_index += num_layers
+
+        # === Layers ===
+        self.embedding = TokenAndPositionEmbedding(
+            vocabulary_size=vocabulary_size,
+            sequence_length=sequence_length,
+            embedding_dim=embedding_dim,
+            dtype=dtype,
+            name="embedding",
+        )
+        self.encoder_layers = [
+            CLIPEncoderBlock(
+                hidden_dim,
+                num_heads,
+                intermediate_dim,
+                intermediate_activation,
+                dtype=dtype,
+            )
+            for _ in range(num_layers)
+        ]
+        self.layer_norm = layers.LayerNormalization(
+            epsilon=0.00001, dtype=dtype, name="layer_norm"
+        )
+        self.text_projection = layers.Dense(
+            hidden_dim,
+            use_bias=False,
+            dtype=dtype,
+            name="text_projection",
+        )
+
+        # === Functional Model ===
+        encoder_token_ids = layers.Input(
+            shape=(sequence_length,), dtype="int32", name="encoder_token_ids"
+        )
+        x = self.embedding(encoder_token_ids)
+        encoder_intermediate_output = None
+        # Encoder.
+        for i, block in enumerate(self.encoder_layers):
+            x = block(x)
+            if i == intermediate_output_index:
+                encoder_intermediate_output = x
+        x = self.layer_norm(x)
+        encoder_output = x
+        if encoder_intermediate_output is not None:
+            encoder_intermediate_output = self.layer_norm(
+                encoder_intermediate_output
+            )
+        # Projection.
+        indices = ops.expand_dims(
+            ops.cast(ops.argmax(encoder_token_ids, axis=-1), "int32"), axis=-1
+        )
+        pooled_output = ops.take_along_axis(x, indices[:, :, None], axis=1)
+        pooled_output = ops.squeeze(pooled_output, axis=1)
+        projection_output = self.text_projection(pooled_output)
+
+        outputs = {
+            "encoder_sequence_output": encoder_output,
+            "encoder_pooled_output": pooled_output,
+            "encoder_projection_output": projection_output,
+        }
+        if intermediate_output_index is not None:
+            outputs["encoder_intermediate_output"] = encoder_intermediate_output
+
+        super().__init__(
+            inputs={"encoder_token_ids": encoder_token_ids},
+            outputs=outputs,
+            **kwargs,
+        )
+
+        # === Config ===
+        self.embedding_dim = embedding_dim
+        self.hidden_dim = hidden_dim
+        self.num_layers = num_layers
+        self.num_heads = num_heads
+        self.intermediate_dim = intermediate_dim
+        self.intermediate_activation = intermediate_activation
+        self.intermediate_output_index = intermediate_output_index
+        self.vocabulary_size = vocabulary_size
+        self.sequence_length = sequence_length
+
+        if dtype is not None:
+            try:
+                self.dtype_policy = keras.dtype_policies.get(dtype)
+            # Before Keras 3.2, there is no `keras.dtype_policies.get`.
+            except AttributeError:
+                if isinstance(dtype, keras.DTypePolicy):
+                    dtype = dtype.name
+                self.dtype_policy = keras.DTypePolicy(dtype)
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "embedding_dim": self.embedding_dim,
+                "hidden_dim": self.hidden_dim,
+                "num_layers": self.num_layers,
+                "num_heads": self.num_heads,
+                "intermediate_dim": self.intermediate_dim,
+                "intermediate_activation": self.intermediate_activation,
+                "intermediate_output_index": self.intermediate_output_index,
+                "vocabulary_size": self.vocabulary_size,
+                "sequence_length": self.sequence_length,
+            }
+        )
+        return config

keras_hub/src/models/stable_diffusion_v3/clip_tokenizer.py

@@ -0,0 +1,167 @@
+# 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.tokenizers.byte_pair_tokenizer import BytePairTokenizer
+from keras_hub.src.tokenizers.byte_pair_tokenizer import convert_to_ragged_batch
+from keras_hub.src.tokenizers.byte_pair_tokenizer import split_strings_for_bpe
+
+try:
+    import tensorflow as tf
+except ImportError:
+    tf = None
+
+
+class CLIPTokenizer(BytePairTokenizer):
+    def __init__(self, vocabulary=None, merges=None, **kwargs):
+        self.start_token = "<|startoftext|>"
+        self.end_token = "<|endoftext|>"
+
+        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 _bpe_merge_and_update_cache(self, tokens):
+        """Process unseen tokens and add to cache."""
+        words = self._transform_bytes(tokens)
+
+        # In StableDiffusionV3, we need to add `</w>` to the last word.
+        words = tf.strings.reduce_join(words, axis=1, separator=" ")
+        words = tf.strings.join([words, "</w>"])
+        words = tf.strings.split(words, sep=" ")
+
+        tokenized_words = self._bpe_merge(words)
+
+        # For each word, join all its token by a whitespace,
+        # e.g., ["dragon", "fly"] => "dragon fly" for hash purpose.
+        tokenized_words = tf.strings.reduce_join(
+            tokenized_words, axis=1, separator=" "
+        )
+        self.cache.insert(tokens, tokenized_words)
+
+    def tokenize(self, inputs):
+        self._check_vocabulary()
+        if not isinstance(inputs, (tf.Tensor, tf.RaggedTensor)):
+            inputs = tf.convert_to_tensor(inputs)
+
+        if self.add_prefix_space:
+            inputs = tf.strings.join([" ", inputs])
+
+        scalar_input = inputs.shape.rank == 0
+        if scalar_input:
+            inputs = tf.expand_dims(inputs, 0)
+
+        raw_tokens = split_strings_for_bpe(inputs, self.unsplittable_tokens)
+
+        # Strip and remove empty tokens.
+        raw_tokens = tf.strings.strip(raw_tokens)
+        raw_tokens = tf.ragged.boolean_mask(raw_tokens, raw_tokens != "")
+
+        token_row_splits = raw_tokens.row_splits
+        flat_tokens = raw_tokens.flat_values
+
+        # Check cache.
+        cache_lookup = self.cache.lookup(flat_tokens)
+        cache_mask = cache_lookup == ""
+
+        has_unseen_words = tf.math.reduce_any(
+            (cache_lookup == "") & (flat_tokens != "")
+        )
+
+        def process_unseen_tokens():
+            unseen_tokens = tf.boolean_mask(flat_tokens, cache_mask)
+            self._bpe_merge_and_update_cache(unseen_tokens)
+            return self.cache.lookup(flat_tokens)
+
+        # If `has_unseen_words == True`, it means not all tokens are in cache,
+        # we will process the unseen tokens. Otherwise return the cache lookup.
+        tokenized_words = tf.cond(
+            has_unseen_words,
+            process_unseen_tokens,
+            lambda: cache_lookup,
+        )
+
+        tokens = tf.strings.split(tokenized_words, sep=" ")
+        if self.compute_dtype != tf.string:
+            # Encode merged tokens.
+            tokens = self.token_to_id_map.lookup(tokens)
+
+        # Unflatten to match input.
+        tokens = tf.RaggedTensor.from_row_splits(
+            tokens.flat_values,
+            tf.gather(tokens.row_splits, token_row_splits),
+        )
+
+        # Convert to a dense output if `sequence_length` is set.
+        if self.sequence_length:
+            output_shape = tokens.shape.as_list()
+            output_shape[-1] = self.sequence_length
+            tokens = tokens.to_tensor(shape=output_shape)
+
+        # Convert to a dense output if input in scalar
+        if scalar_input:
+            tokens = tf.squeeze(tokens, 0)
+            tf.ensure_shape(tokens, shape=[self.sequence_length])
+
+        return tokens
+
+    def detokenize(self, inputs):
+        self._check_vocabulary()
+        inputs, unbatched, _ = convert_to_ragged_batch(inputs)
+        inputs = tf.cast(inputs, self.dtype)
+        unicode_text = tf.strings.reduce_join(
+            self.id_to_token_map.lookup(inputs), axis=-1
+        )
+
+        # When detokenizing, we need to remove </w> and extra whitespace.
+        unicode_text = tf.strings.regex_replace(unicode_text, r"</w>", " ")
+        unicode_text = tf.strings.strip(unicode_text)
+
+        split_unicode_text = tf.strings.unicode_split(unicode_text, "UTF-8")
+        outputs = tf.strings.reduce_join(
+            self.unicode2byte.lookup(split_unicode_text), axis=-1
+        )
+
+        if unbatched:
+            outputs = tf.squeeze(outputs, 0)
+        return outputs
+
+    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