keras-hub-nightly 0.21.0.dev202505050407__py3-none-any.whl → 0.21.0.dev202505070407__py3-none-any.whl

keras_hub/src/models/retinanet/retinanet_image_converter.py

@@ -6,16 +6,3 @@ from keras_hub.src.models.retinanet.retinanet_backbone import RetinaNetBackbone
 @keras_hub_export("keras_hub.layers.RetinaNetImageConverter")
 class RetinaNetImageConverter(ImageConverter):
     backbone_cls = RetinaNetBackbone
-
-    def __init__(
-        self,
-        *args,
-        **kwargs,
-    ):
-        # TODO: update presets and remove these old config options. They were
-        # never needed.
-        if "norm_mean" in kwargs:
-            kwargs["offset"] = [-x for x in kwargs.pop("norm_mean")]
-        if "norm_std" in kwargs:
-            kwargs["scale"] = [1.0 / x for x in kwargs.pop("norm_std")]
-        super().__init__(*args, **kwargs)

keras_hub/src/models/retinanet/retinanet_presets.py

@@ -11,7 +11,7 @@ backbone_presets = {
             "params": 34121239,
             "path": "retinanet",
         },
-        "kaggle_handle": "kaggle://keras/retinanet/keras/retinanet_resnet50_fpn_coco/3",
+        "kaggle_handle": "kaggle://keras/retinanet/keras/retinanet_resnet50_fpn_coco/4",
     },
     "retinanet_resnet50_fpn_v2_coco": {
         "metadata": {
@@ -22,6 +22,6 @@ backbone_presets = {
             "params": 31558592,
             "path": "retinanet",
         },
-        "kaggle_handle": "kaggle://keras/retinanet/keras/retinanet_resnet50_fpn_v2_coco/2",
+        "kaggle_handle": "kaggle://keras/retinanet/keras/retinanet_resnet50_fpn_v2_coco/3",
     },
 }

keras_hub/src/models/segformer/segformer_image_segmenter_preprocessor.py

@@ -1,5 +1,3 @@
-import keras
-
 from keras_hub.src.api_export import keras_hub_export
 from keras_hub.src.models.image_segmenter_preprocessor import (
     ImageSegmenterPreprocessor,
@@ -8,25 +6,9 @@ from keras_hub.src.models.segformer.segformer_backbone import SegFormerBackbone
 from keras_hub.src.models.segformer.segformer_image_converter import (
     SegFormerImageConverter,
 )
-from keras_hub.src.utils.tensor_utils import preprocessing_function
-
-IMAGENET_DEFAULT_MEAN = [0.485, 0.456, 0.406]
-IMAGENET_DEFAULT_STD = [0.229, 0.224, 0.225]
 
 
 @keras_hub_export("keras_hub.models.SegFormerImageSegmenterPreprocessor")
 class SegFormerImageSegmenterPreprocessor(ImageSegmenterPreprocessor):
     backbone_cls = SegFormerBackbone
     image_converter_cls = SegFormerImageConverter
-
-    @preprocessing_function
-    def call(self, x, y=None, sample_weight=None):
-        if self.image_converter:
-            x = self.image_converter(x)
-            if y is not None:
-                y = self.image_converter(y)
-
-        x = x / 255
-        x = (x - IMAGENET_DEFAULT_MEAN) / IMAGENET_DEFAULT_STD
-
-        return keras.utils.pack_x_y_sample_weight(x, y, sample_weight)

keras_hub/src/models/segformer/segformer_presets.py

@@ -10,7 +10,7 @@ presets = {
             "params": 3719027,
             "path": "segformer_b0",
         },
-        "kaggle_handle": "kaggle://keras/segformer/keras/segformer_b0_ade20k_512/2",
+        "kaggle_handle": "kaggle://keras/segformer/keras/segformer_b0_ade20k_512/3",
     },
     "segformer_b1_ade20k_512": {
         "metadata": {
@@ -21,7 +21,7 @@ presets = {
             "params": 13682643,
             "path": "segformer_b1",
         },
-        "kaggle_handle": "kaggle://keras/segformer/keras/segformer_b1_ade20k_512/2",
+        "kaggle_handle": "kaggle://keras/segformer/keras/segformer_b1_ade20k_512/5",
     },
     "segformer_b2_ade20k_512": {
         "metadata": {
@@ -32,7 +32,7 @@ presets = {
             "params": 24727507,
             "path": "segformer_b2",
         },
-        "kaggle_handle": "kaggle://keras/segformer/keras/segformer_b2_ade20k_512/2",
+        "kaggle_handle": "kaggle://keras/segformer/keras/segformer_b2_ade20k_512/3",
     },
     "segformer_b3_ade20k_512": {
         "metadata": {
@@ -43,7 +43,7 @@ presets = {
             "params": 44603347,
             "path": "segformer_b3",
         },
-        "kaggle_handle": "kaggle://keras/segformer/keras/segformer_b3_ade20k_512/2",
+        "kaggle_handle": "kaggle://keras/segformer/keras/segformer_b3_ade20k_512/3",
     },
     "segformer_b4_ade20k_512": {
         "metadata": {
@@ -54,7 +54,7 @@ presets = {
             "params": 61373907,
             "path": "segformer_b4",
         },
-        "kaggle_handle": "kaggle://keras/segformer/keras/segformer_b4_ade20k_512/2",
+        "kaggle_handle": "kaggle://keras/segformer/keras/segformer_b4_ade20k_512/3",
     },
     "segformer_b5_ade20k_640": {
         "metadata": {
@@ -65,7 +65,7 @@ presets = {
             "params": 81974227,
             "path": "segformer_b5",
         },
-        "kaggle_handle": "kaggle://keras/segformer/keras/segformer_b5_ade20k_640/2",
+        "kaggle_handle": "kaggle://keras/segformer/keras/segformer_b5_ade20k_640/3",
     },
     "segformer_b0_cityscapes_1024": {
         "metadata": {
@@ -76,7 +76,7 @@ presets = {
             "params": 3719027,
             "path": "segformer_b0",
         },
-        "kaggle_handle": "kaggle://keras/segformer/keras/segformer_b0_cityscapes_1024/2",
+        "kaggle_handle": "kaggle://keras/segformer/keras/segformer_b0_cityscapes_1024/3",
     },
     "segformer_b1_cityscapes_1024": {
         "metadata": {
@@ -87,7 +87,7 @@ presets = {
             "params": 13682643,
             "path": "segformer_b1",
         },
-        "kaggle_handle": "kaggle://keras/segformer/keras/segformer_b1_ade20k_512/2",
+        "kaggle_handle": "kaggle://keras/segformer/keras/segformer_b1_ade20k_512/1",
     },
     "segformer_b2_cityscapes_1024": {
         "metadata": {
@@ -98,7 +98,7 @@ presets = {
             "params": 24727507,
             "path": "segformer_b2",
         },
-        "kaggle_handle": "kaggle://keras/segformer/keras/segformer_b2_cityscapes_1024/2",
+        "kaggle_handle": "kaggle://keras/segformer/keras/segformer_b2_cityscapes_1024/3",
     },
     "segformer_b3_cityscapes_1024": {
         "metadata": {
@@ -109,7 +109,7 @@ presets = {
             "params": 44603347,
             "path": "segformer_b3",
         },
-        "kaggle_handle": "kaggle://keras/segformer/keras/segformer_b3_cityscapes_1024/2",
+        "kaggle_handle": "kaggle://keras/segformer/keras/segformer_b3_cityscapes_1024/3",
     },
     "segformer_b4_cityscapes_1024": {
         "metadata": {
@@ -120,7 +120,7 @@ presets = {
             "params": 61373907,
             "path": "segformer_b4",
         },
-        "kaggle_handle": "kaggle://keras/segformer/keras/segformer_b4_cityscapes_1024/2",
+        "kaggle_handle": "kaggle://keras/segformer/keras/segformer_b4_cityscapes_1024/3",
     },
     "segformer_b5_cityscapes_1024": {
         "metadata": {
@@ -131,6 +131,6 @@ presets = {
             "params": 81974227,
             "path": "segformer_b5",
         },
-        "kaggle_handle": "kaggle://keras/segformer/keras/segformer_b5_cityscapes_1024/2",
+        "kaggle_handle": "kaggle://keras/segformer/keras/segformer_b5_cityscapes_1024/3",
     },
 }

keras_hub/src/models/task.py

@@ -236,14 +236,17 @@ class Task(PipelineModel):
             objects_to_skip=backbone_layer_ids,
         )
 
-    def save_to_preset(self, preset_dir):
+    def save_to_preset(self, preset_dir, max_shard_size=10):
         """Save task to a preset directory.
 
         Args:
             preset_dir: The path to the local model preset directory.
+            max_shard_size: `int` or `float`. Maximum size in GB for each
+                sharded file. If `None`, no sharding will be done. Defaults to
+                `10`.
         """
         saver = get_preset_saver(preset_dir)
-        saver.save_task(self)
+        saver.save_task(self, max_shard_size=max_shard_size)
 
     @property
     def layers(self):

keras_hub/src/utils/keras_utils.py

@@ -1,3 +1,4 @@
+import inspect
 import sys
 
 import keras
@@ -147,3 +148,13 @@ def get_gpu_names():
         ]
     else:
         return [""]
+
+
+def sharded_weights_available():
+    """Whether sharded weights serialization is available.
+
+    Returns:
+        `True` if sharded weights are available, `False` otherwise.
+    """
+    save_weights_signature = inspect.signature(keras.saving.save_weights)
+    return "max_shard_size" in save_weights_signature.parameters

keras_hub/src/utils/preset_utils.py

@@ -10,6 +10,8 @@ from absl import logging
 
 from keras_hub.src.api_export import keras_hub_export
 from keras_hub.src.utils.keras_utils import print_msg
+from keras_hub.src.utils.keras_utils import sharded_weights_available
+from keras_hub.src.utils.tensor_utils import get_tensor_size_in_bits
 
 try:
     import kagglehub
@@ -48,6 +50,7 @@ METADATA_FILE = "metadata.json"
 # Weight file names.
 MODEL_WEIGHTS_FILE = "model.weights.h5"
 TASK_WEIGHTS_FILE = "task.weights.h5"
+SHARDED_MODEL_WEIGHTS_CONFIG_FILE = "model.weights.json"
 
 # HuggingFace filenames.
 README_FILE = "README.md"
@@ -647,7 +650,7 @@ class KerasPresetLoader(PresetLoader):
         backbone = self._load_serialized_object(self.config, **kwargs)
         if load_weights:
             jax_memory_cleanup(backbone)
-            backbone.load_weights(get_file(self.preset, MODEL_WEIGHTS_FILE))
+            self._load_backbone_weights(backbone)
         return backbone
 
     def load_tokenizer(self, cls, config_file=TOKENIZER_CONFIG_FILE, **kwargs):
@@ -697,8 +700,7 @@ class KerasPresetLoader(PresetLoader):
                 task.load_task_weights(task_weights)
             else:
                 jax_memory_cleanup(task.backbone)
-            backbone_weights = get_file(self.preset, MODEL_WEIGHTS_FILE)
-            task.backbone.load_weights(backbone_weights)
+            self._load_backbone_weights(task.backbone)
         return task
 
     def load_preprocessor(
@@ -726,18 +728,64 @@ class KerasPresetLoader(PresetLoader):
         config["config"] = {**config["config"], **kwargs}
         return keras.saving.deserialize_keras_object(config)
 
+    def _get_sharded_filenames(self, config_path):
+        with open(config_path, encoding="utf-8") as config_file:
+            config = json.load(config_file)
+        weight_map = config["weight_map"]
+        return sorted(set(weight_map.values()))
+
+    def _load_backbone_weights(self, backbone):
+        # Detect if the backbone is sharded or not.
+        has_single_file_weights = check_file_exists(
+            self.preset, MODEL_WEIGHTS_FILE
+        )
+        if has_single_file_weights:
+            filepath = get_file(self.preset, MODEL_WEIGHTS_FILE)
+        else:
+            if not sharded_weights_available():
+                raise RuntimeError(
+                    "Sharded weights loading is not supported in the current "
+                    f"Keras version {keras.__version__}. "
+                    "Please update to a newer version."
+                )
+            filepath = get_file(self.preset, SHARDED_MODEL_WEIGHTS_CONFIG_FILE)
+            sharded_filenames = self._get_sharded_filenames(filepath)
+            for sharded_filename in sharded_filenames:
+                # Download the sharded weights.
+                _ = get_file(self.preset, sharded_filename)
+        backbone.load_weights(filepath)
+
 
 class KerasPresetSaver:
     def __init__(self, preset_dir):
         os.makedirs(preset_dir, exist_ok=True)
         self.preset_dir = preset_dir
 
-    def save_backbone(self, backbone):
+    def save_backbone(self, backbone, max_shard_size=10):
         self._save_serialized_object(backbone, config_file=CONFIG_FILE)
-        backbone_weight_path = os.path.join(self.preset_dir, MODEL_WEIGHTS_FILE)
-        backbone.save_weights(backbone_weight_path)
         self._save_metadata(backbone)
 
+        # Save the weights.
+        backbone_size_in_bytes = self._get_variables_size_in_bytes(
+            backbone.variables
+        )
+        backbone_size_in_gb = backbone_size_in_bytes / (1024**3)
+        # If the size of the backbone is larger than `max_shard_size`, save
+        # sharded weights.
+        if sharded_weights_available() and backbone_size_in_gb > max_shard_size:
+            backbone_sharded_weights_config_path = os.path.join(
+                self.preset_dir, SHARDED_MODEL_WEIGHTS_CONFIG_FILE
+            )
+            backbone.save_weights(
+                backbone_sharded_weights_config_path,
+                max_shard_size=max_shard_size,
+            )
+        else:
+            backbone_weight_path = os.path.join(
+                self.preset_dir, MODEL_WEIGHTS_FILE
+            )
+            backbone.save_weights(backbone_weight_path)
+
     def save_tokenizer(self, tokenizer):
         config_file = TOKENIZER_CONFIG_FILE
         if hasattr(tokenizer, "config_file"):
@@ -755,7 +803,7 @@ class KerasPresetSaver:
     def save_image_converter(self, converter):
         self._save_serialized_object(converter, IMAGE_CONVERTER_CONFIG_FILE)
 
-    def save_task(self, task):
+    def save_task(self, task, max_shard_size=10):
         # Save task specific config and weights.
         self._save_serialized_object(task, TASK_CONFIG_FILE)
         if task.has_task_weights():
@@ -763,10 +811,12 @@ class KerasPresetSaver:
             task.save_task_weights(task_weight_path)
         # Save backbone.
         if hasattr(task.backbone, "save_to_preset"):
-            task.backbone.save_to_preset(self.preset_dir)
+            task.backbone.save_to_preset(
+                self.preset_dir, max_shard_size=max_shard_size
+            )
         else:
             # Allow saving a `keras.Model` that is not a backbone subclass.
-            self.save_backbone(task.backbone)
+            self.save_backbone(task.backbone, max_shard_size=max_shard_size)
         # Save preprocessor.
         if task.preprocessor and hasattr(task.preprocessor, "save_to_preset"):
             task.preprocessor.save_to_preset(self.preset_dir)
@@ -823,3 +873,13 @@ class KerasPresetSaver:
         metadata_path = os.path.join(self.preset_dir, METADATA_FILE)
         with open(metadata_path, "w") as metadata_file:
             metadata_file.write(json.dumps(metadata, indent=4))
+
+    def _get_variables_size_in_bytes(self, variables):
+        unique_variables = {}
+        for v in variables:
+            if id(v) not in unique_variables:
+                unique_variables[id(v)] = (v.shape, v.dtype)
+        total_memory_size = 0
+        for shape, dtype in unique_variables.values():
+            total_memory_size += get_tensor_size_in_bits(shape, dtype)
+        return total_memory_size / 8

keras_hub/src/utils/tensor_utils.py

@@ -1,6 +1,8 @@
 import contextlib
 import functools
 import inspect
+import math
+import re
 import threading
 
 import keras
@@ -305,6 +307,29 @@ def is_string_dtype(dtype):
     return "string" in keras.backend.standardize_dtype(dtype)
 
 
+def get_dtype_size_in_bits(dtype):
+    """Get the size of a given dtype in bits."""
+    dtype = keras.backend.standardize_dtype(dtype)
+    # If dtype is bool, return 1 immediately.
+    if dtype == "bool":
+        return 1
+    # Else, we extract the bit size from the string.
+    return int(re.sub(r"bfloat|float|uint|int", "", dtype))
+
+
+def get_tensor_size_in_bits(shape, dtype):
+    """Calculate the size given dtype and shape in bits.
+
+    Args:
+        dtype: The dtype of the tensor.
+        shape: List of iterables representing the shape of the tensor.
+
+    Returns:
+        The size of the tensor in bytes.
+    """
+    return math.prod(shape) * get_dtype_size_in_bits(dtype)
+
+
 def any_equal(inputs, values, padding_mask):
     """Return a mask that is True anywhere `inputs` has a value in `values`.
 
@@ -320,7 +345,8 @@ def any_equal(inputs, values, padding_mask):
     Returns:
         A tensor with `inputs` shape where each position is True if it contains
             a value from any `values`. Padding mask will be applied before
-            returning."""
+            returning.
+    """
     output = ops.equal(inputs, values[0])
     for value in values[1:]:
         value_equality = ops.equal(inputs, value)

keras_hub/src/utils/timm/convert_cspnet.py

@@ -17,10 +17,69 @@ def convert_backbone_config(timm_config):
         bottle_ratio = (0.5,) + (1.0,)
         block_ratio = (1.0,) + (0.5,)
         expand_ratio = (2.0,) + (1.0,)
+        stem_padding = "same"
+        stem_pooling = None
         stage_type = "csp"
+        groups = 1
         block_type = "dark_block"
         down_growth = True
-        stackwise_strides = 2
+        stackwise_strides = [2, 2, 2, 2, 2]
+        avg_down = False
+        cross_linear = False
+    elif timm_architecture == "cspresnet50":
+        stem_filters = 64
+        stem_kernel_size = 7
+        stem_strides = 4
+        stackwise_depth = [3, 3, 5, 2]
+        stackwise_strides = [1, 2, 2, 2]
+        stackwise_num_filters = [128, 256, 512, 1024]
+        block_type = "bottleneck_block"
+        stage_type = "csp"
+        bottle_ratio = [0.5]
+        block_ratio = [1.0]
+        expand_ratio = [2.0]
+        stem_padding = "valid"
+        stem_pooling = "max"
+        avg_down = False
+        groups = 1
+        down_growth = False
+        cross_linear = True
+    elif timm_architecture == "cspresnext50":
+        stem_filters = 64
+        stem_kernel_size = 7
+        stem_strides = 4
+        stackwise_depth = [3, 3, 5, 2]
+        stackwise_num_filters = [256, 512, 1024, 2048]
+        bottle_ratio = [1.0]
+        block_ratio = [0.5]
+        expand_ratio = [1.0]
+        stage_type = "csp"
+        block_type = "bottleneck_block"
+        stem_pooling = "max"
+        stackwise_strides = [1, 2, 2, 2]
+        groups = 32
+        stem_padding = "valid"
+        avg_down = False
+        down_growth = False
+        cross_linear = True
+    elif timm_architecture == "darknet53":
+        stem_filters = 32
+        stem_kernel_size = 3
+        stem_strides = 1
+        stackwise_depth = [1, 2, 8, 8, 4]
+        stackwise_num_filters = [64, 128, 256, 512, 1024]
+        bottle_ratio = [0.5]
+        block_ratio = [1.0]
+        groups = 1
+        expand_ratio = [1.0]
+        stage_type = "dark"
+        block_type = "dark_block"
+        stem_pooling = None
+        stackwise_strides = [2, 2, 2, 2, 2]
+        stem_padding = "same"
+        avg_down = False
+        down_growth = False
+        cross_linear = False
     else:
         raise ValueError(
             f"Currently, the architecture {timm_architecture} is not supported."
@@ -38,6 +97,11 @@ def convert_backbone_config(timm_config):
         block_type=block_type,
         stackwise_strides=stackwise_strides,
         down_growth=down_growth,
+        stem_pooling=stem_pooling,
+        stem_padding=stem_padding,
+        avg_down=avg_down,
+        cross_linear=cross_linear,
+        groups=groups,
     )
 
 
@@ -81,21 +145,36 @@ def convert_weights(backbone, loader, timm_config):
     stackwise_depth = backbone.stackwise_depth
     stage_type = backbone.stage_type
     block_type = backbone.block_type
+    strides = backbone.stackwise_strides
 
     for idx, block in enumerate(stackwise_depth):
-        port_conv2d(
-            f"stages.{idx}.conv_down.conv",
-            f"stage_{idx}_{stage_type}_conv_down_1",
-        )
-        port_batch_normalization(
-            f"stages.{idx}.conv_down.bn", f"stage_{idx}_{stage_type}_bn_1"
-        )
-        port_conv2d(
-            f"stages.{idx}.conv_exp.conv", f"stage_{idx}_{stage_type}_conv_exp"
-        )
-        port_batch_normalization(
-            f"stages.{idx}.conv_exp.bn", f"stage_{idx}_{stage_type}_bn_2"
-        )
+        if strides[idx] != 1 or stage_type == "dark":
+            if strides[idx] == 2 and backbone.avg_down:
+                port_conv2d(
+                    f"stages.{idx}.conv_down.1.conv",
+                    f"stage_{idx}_{stage_type}_conv_down_1",
+                )
+                port_batch_normalization(
+                    f"stages.{idx}.conv_down.1.bn",
+                    f"stage_{idx}_{stage_type}_bn_1",
+                )
+            else:
+                port_conv2d(
+                    f"stages.{idx}.conv_down.conv",
+                    f"stage_{idx}_{stage_type}_conv_down_1",
+                )
+                port_batch_normalization(
+                    f"stages.{idx}.conv_down.bn",
+                    f"stage_{idx}_{stage_type}_bn_1",
+                )
+        if stage_type != "dark":
+            port_conv2d(
+                f"stages.{idx}.conv_exp.conv",
+                f"stage_{idx}_{stage_type}_conv_exp",
+            )
+            port_batch_normalization(
+                f"stages.{idx}.conv_exp.bn", f"stage_{idx}_{stage_type}_bn_2"
+            )
 
         for i in range(block):
             port_conv2d(
@@ -133,16 +212,8 @@ def convert_weights(backbone, loader, timm_config):
                 f"stages.{idx}.conv_transition_b.bn",
                 f"stage_{idx}_{stage_type}_transition_b_bn",
             )
-            port_conv2d(
-                f"stages.{idx}.conv_transition.conv",
-                f"stage_{idx}_{stage_type}_conv_transition",
-            )
-            port_batch_normalization(
-                f"stages.{idx}.conv_transition.bn",
-                f"stage_{idx}_{stage_type}_transition_bn",
-            )
 
-        else:
+        if stage_type != "dark":
             port_conv2d(
                 f"stages.{idx}.conv_transition.conv",
                 f"stage_{idx}_{stage_type}_conv_transition",

keras_hub/src/utils/timm/preset_loader.py

@@ -16,17 +16,17 @@ class TimmPresetLoader(PresetLoader):
     def __init__(self, preset, config):
         super().__init__(preset, config)
         architecture = self.config["architecture"]
-        if "resnet" in architecture:
+        if architecture.startswith("resnet"):
             self.converter = convert_resnet
-        elif "csp" in architecture:
+        elif architecture.startswith(("csp", "dark")):
             self.converter = convert_cspnet
-        elif "densenet" in architecture:
+        elif architecture.startswith("densenet"):
             self.converter = convert_densenet
-        elif "mobilenet" in architecture:
+        elif architecture.startswith("mobilenet"):
             self.converter = convert_mobilenet
-        elif "vgg" in architecture:
+        elif architecture.startswith("vgg"):
             self.converter = convert_vgg
-        elif "efficientnet" in architecture:
+        elif architecture.startswith("efficientnet"):
             self.converter = convert_efficientnet
         else:
             raise ValueError(