keras-hub-nightly 0.23.0.dev202508260411__py3-none-any.whl → 0.23.0.dev202508270416__py3-none-any.whl

keras_hub/layers/__init__.py

@@ -75,6 +75,9 @@ from keras_hub.src.models.clip.clip_image_converter import (
 from keras_hub.src.models.cspnet.cspnet_image_converter import (
     CSPNetImageConverter as CSPNetImageConverter,
 )
+from keras_hub.src.models.d_fine.d_fine_image_converter import (
+    DFineImageConverter as DFineImageConverter,
+)
 from keras_hub.src.models.deeplab_v3.deeplab_v3_image_converter import (
     DeepLabV3ImageConverter as DeepLabV3ImageConverter,
 )
@@ -108,6 +111,9 @@ from keras_hub.src.models.moonshine.moonshine_audio_converter import (
 from keras_hub.src.models.pali_gemma.pali_gemma_image_converter import (
     PaliGemmaImageConverter as PaliGemmaImageConverter,
 )
+from keras_hub.src.models.parseq.parseq_image_converter import (
+    PARSeqImageConverter as PARSeqImageConverter,
+)
 from keras_hub.src.models.resnet.resnet_image_converter import (
     ResNetImageConverter as ResNetImageConverter,
 )

keras_hub/models/__init__.py

@@ -108,6 +108,15 @@ from keras_hub.src.models.cspnet.cspnet_image_classifier import (
 from keras_hub.src.models.cspnet.cspnet_image_classifier_preprocessor import (
     CSPNetImageClassifierPreprocessor as CSPNetImageClassifierPreprocessor,
 )
+from keras_hub.src.models.d_fine.d_fine_backbone import (
+    DFineBackbone as DFineBackbone,
+)
+from keras_hub.src.models.d_fine.d_fine_object_detector import (
+    DFineObjectDetector as DFineObjectDetector,
+)
+from keras_hub.src.models.d_fine.d_fine_object_detector_preprocessor import (
+    DFineObjectDetectorPreprocessor as DFineObjectDetectorPreprocessor,
+)
 from keras_hub.src.models.deberta_v3.deberta_v3_backbone import (
     DebertaV3Backbone as DebertaV3Backbone,
 )
@@ -446,6 +455,18 @@ from keras_hub.src.models.pali_gemma.pali_gemma_causal_lm_preprocessor import (
 from keras_hub.src.models.pali_gemma.pali_gemma_tokenizer import (
     PaliGemmaTokenizer as PaliGemmaTokenizer,
 )
+from keras_hub.src.models.parseq.parseq_backbone import (
+    PARSeqBackbone as PARSeqBackbone,
+)
+from keras_hub.src.models.parseq.parseq_causal_lm import (
+    PARSeqCausalLM as PARSeqCausalLM,
+)
+from keras_hub.src.models.parseq.parseq_causal_lm_preprocessor import (
+    PARSeqCausalLMPreprocessor as PARSeqCausalLMPreprocessor,
+)
+from keras_hub.src.models.parseq.parseq_tokenizer import (
+    PARSeqTokenizer as PARSeqTokenizer,
+)
 from keras_hub.src.models.phi3.phi3_backbone import Phi3Backbone as Phi3Backbone
 from keras_hub.src.models.phi3.phi3_causal_lm import (
     Phi3CausalLM as Phi3CausalLM,

keras_hub/src/layers/modeling/position_embedding.py

@@ -31,6 +31,11 @@ class PositionEmbedding(keras.layers.Layer):
         start_index: An integer or integer tensor. The starting position to
             compute the position embedding from. This is useful during cached
             decoding, where each position is predicted separately in a loop.
+        positions: Tensor of shape `(sequence_length,)` or
+            `(batch_size, sequence_length)`. Custom positions for the input
+            sequence. If specified, this tensor will be used to
+            compute the position embedding, and the `start_index` argument will
+            be ignored. This is useful for cases with non-standard positions.
 
     Example:
 
@@ -91,18 +96,28 @@ class PositionEmbedding(keras.layers.Layer):
         )
         self.built = True
 
-    def call(self, inputs, start_index=0):
+    def call(self, inputs, start_index=0, positions=None):
         shape = ops.shape(inputs)
         feature_length = shape[-1]
         sequence_length = shape[-2]
         # trim to match the length of the input sequence, which might be less
         # than the sequence_length of the layer.
         position_embeddings = ops.convert_to_tensor(self.position_embeddings)
-        position_embeddings = ops.slice(
-            position_embeddings,
-            (start_index, 0),
-            (sequence_length, feature_length),
-        )
+        if positions is None:
+            position_embeddings = ops.slice(
+                position_embeddings,
+                (start_index, 0),
+                (sequence_length, feature_length),
+            )
+        else:
+            # Take care of unbatched `positions`.
+            if len(ops.shape(positions)) == 1:
+                positions = ops.expand_dims(positions, axis=0)
+
+            position_embeddings = ops.take(
+                position_embeddings, positions, axis=0
+            )
+
         return ops.broadcast_to(position_embeddings, shape)
 
     def compute_output_shape(self, input_shape):

keras_hub/src/layers/modeling/rotary_embedding.py

@@ -37,6 +37,11 @@ class RotaryEmbedding(keras.layers.Layer):
         start_index: An integer or integer tensor. The starting position to
             compute the rotary embedding from. This is useful during cached
             decoding, where each position is predicted separately in a loop.
+        positions: Tensor of shape `(sequence_length,)` or
+            `(batch_size, sequence_length)`. Custom positions for the input
+            sequence. If specified, this tensor will be used to
+            compute the rotary embedding, and the `start_index` argument will
+            be ignored. This is useful for cases with non-standard positions.
 
     Examples:
 
@@ -76,6 +81,11 @@ class RotaryEmbedding(keras.layers.Layer):
         self.built = True
 
     def call(self, inputs, start_index=0, positions=None):
+        # Take care of unbatched `positions`.
+        if positions is not None:
+            if len(ops.shape(positions)) == 1:
+                positions = ops.expand_dims(positions, axis=0)
+
         inputs = ops.moveaxis(
             inputs, (self.feature_axis, self.sequence_axis), (-1, 1)
         )
@@ -103,6 +113,7 @@ class RotaryEmbedding(keras.layers.Layer):
         return positions + ops.cast(start_index, dtype="float32")
 
     def _compute_cos_sin_embedding(self, inputs, start_index=0, positions=None):
+        batch_axis = 0
         feature_axis = len(inputs.shape) - 1
         sequence_axis = 1
 
@@ -111,21 +122,20 @@ class RotaryEmbedding(keras.layers.Layer):
 
         if positions is None:
             positions = self._compute_positions(inputs, start_index)
+            positions = ops.expand_dims(positions, axis=batch_axis)
         else:
             positions = ops.cast(positions, "float32")
-
         positions = positions / ops.cast(self.scaling_factor, "float32")
-        freq = ops.einsum("i,j->ij", positions, inverse_freq)
+
+        freq = ops.einsum("bi,j->bij", positions, inverse_freq)
+
         embedding = ops.stack((freq, freq), axis=-2)
         embedding = ops.reshape(
             embedding, (*ops.shape(freq)[:-1], ops.shape(freq)[-1] * 2)
         )
 
-        # Reshape the embedding to be broadcastable with input shape.
-        if feature_axis < sequence_axis:
-            embedding = ops.transpose(embedding)
         for axis in range(len(inputs.shape)):
-            if axis != sequence_axis and axis != feature_axis:
+            if axis not in (batch_axis, sequence_axis, feature_axis):
                 embedding = ops.expand_dims(embedding, axis)
 
         cos_emb = ops.cast(ops.cos(embedding), self.compute_dtype)

keras_hub/src/layers/modeling/sine_position_encoding.py

@@ -30,6 +30,11 @@ class SinePositionEncoding(keras.layers.Layer):
         start_index: An integer or integer tensor. The starting position to
             compute the encoding from. This is useful during cached decoding,
             where each position is predicted separately in a loop.
+        positions: Tensor of shape `(sequence_length,)` or
+            `(batch_size, sequence_length)`. Custom positions for the input
+            sequence. If specified, this tensor will be used to
+            compute the position embedding, and the `start_index` argument will
+            be ignored. This is useful for cases with non-standard positions.
 
     Example:
     ```python
@@ -58,27 +63,35 @@ class SinePositionEncoding(keras.layers.Layer):
         self.max_wavelength = max_wavelength
         self.built = True
 
-    def call(self, inputs, start_index=0):
+    def call(self, inputs, start_index=0, positions=None):
         shape = ops.shape(inputs)
         seq_length = shape[-2]
         hidden_size = shape[-1]
-        positions = ops.arange(seq_length)
-        positions = ops.cast(positions + start_index, self.compute_dtype)
+
+        if positions is None:
+            positions = ops.arange(seq_length)
+            positions = ops.cast(positions + start_index, self.compute_dtype)
+
+        # Take care of unbatched `positions`.
+        if len(ops.shape(positions)) == 1:
+            positions = ops.expand_dims(positions, axis=0)
+
         min_freq = ops.cast(1 / self.max_wavelength, dtype=self.compute_dtype)
         timescales = ops.power(
             min_freq,
             ops.cast(2 * (ops.arange(hidden_size) // 2), self.compute_dtype)
             / ops.cast(hidden_size, self.compute_dtype),
         )
-        angles = ops.expand_dims(positions, 1) * ops.expand_dims(timescales, 0)
+        angles = ops.einsum("bi,j->bij", positions, timescales)
+
         # even indices are sine, odd are cosine
         cos_mask = ops.cast(ops.arange(hidden_size) % 2, self.compute_dtype)
         sin_mask = 1 - cos_mask
-        # embedding shape is [seq_length, hidden_size]
-        positional_encodings = (
-            ops.sin(angles) * sin_mask + ops.cos(angles) * cos_mask
-        )
 
+        # embedding shape is `[bsz (or 1), seq_length, hidden_size]`.
+        positional_encodings = ops.einsum(
+            "bij,j->bij", ops.sin(angles), sin_mask
+        ) + ops.einsum("bij,j->bij", ops.cos(angles), cos_mask)
         return ops.broadcast_to(positional_encodings, shape)
 
     def get_config(self):

keras_hub/src/layers/modeling/token_and_position_embedding.py

@@ -120,11 +120,12 @@ class TokenAndPositionEmbedding(keras.layers.Layer):
         )
         return config
 
-    def call(self, inputs, start_index=0):
+    def call(self, inputs, start_index=0, positions=None):
         embedded_tokens = self.token_embedding(inputs)
         embedded_positions = self.position_embedding(
             embedded_tokens,
             start_index=start_index,
+            positions=positions,
         )
         outputs = embedded_tokens + embedded_positions
         return outputs

keras_hub/src/models/backbone.py

@@ -91,21 +91,16 @@ class Backbone(keras.Model):
         }
 
         # Add quantization support by utilizing `DTypePolicyMap`
-        try:
-            if isinstance(
-                self.dtype_policy, keras.dtype_policies.DTypePolicyMap
-            ):
-                config.update({"dtype": self.dtype_policy})
-            else:
-                policy_map = keras.dtype_policies.DTypePolicyMap()
-                for layer in self._flatten_layers():
-                    if layer.quantization_mode is not None:
-                        policy_map[layer.path] = layer.dtype_policy
-                if len(policy_map) > 0:
-                    config.update({"dtype": policy_map})
-        # Before Keras 3.2, there is no `keras.dtype_policies.get`.
-        except AttributeError:
-            pass
+        dtype = self.dtype_policy
+        if not isinstance(dtype, keras.dtype_policies.DTypePolicyMap):
+            policy_map = keras.dtype_policies.DTypePolicyMap()
+            for layer in self._flatten_layers():
+                if layer.quantization_mode is not None:
+                    policy_map[layer.path] = layer.dtype_policy
+            if len(policy_map) > 0:
+                dtype = policy_map
+
+        config.update({"dtype": keras.dtype_policies.serialize(dtype)})
         return config
 
     @classmethod