merge master (#20741)

* Specify window_length dtype requirement in tf.keras.ops.istft in math.py (#20728) The `window_length` parameter in `tf.keras.ops.istft` requires `tf.int32` dtype, but this isn't documented. This can cause unexpected `ValueError` when using `tf.int64` and `tf.int16` Here is the Example case: ``` import tensorflow as tf input_dict = { 'stfts': tf.constant([[-0.87817144+1.14583987j, -0.32066484+0.25565411j]], dtype=tf.complex128), 'frame_length': tf.constant(256, dtype=tf.int16), 'frame_step': tf.constant(5120,dtype=tf.int64) } result = tf.signal.inverse_stft(**input_dict) print(result) ``` The code throws the following error: ``` ValueError: window_length: Tensor conversion requested dtype int32 for Tensor with dtype int64 ``` * Add rand_augment processing layer (#20716) * Add rand_augment init * Update rand_augment init * Add rand_augment * Add NotImplementedError * Add some test cases * Fix failed test case * Update rand_augment * Update rand_augment test * Fix random_rotation bug * Add build method to supress warning. * Add implementation for transform_bboxes * Fixing batch_dim_name attribute (#20674) * fixing wrong trainer assumption that batch dim is always the first one in the mesh * need functools partial * lint * fix test failure when distribution=None * lint2 * fix for test failure * added data sharding for 3D+ meshes * lint3 * added @Property for batch_dim_name + refactoring * fix typo * Add support for `dtype` / `DTypePolicy` to `JaxLayer` and `FlaxLayer`. (#20732) The `dtype` / `DTypePolicy` is applied to all float variables. * Allow dynamic shape in `STFTSpectrogram` layer. (#20736) by simply using `ops.shape(x)` instead of `x.shape`. * Remove duplicate export tests in `model_test`. (#20735) The same tests exist at: - https://github.com/keras-team/keras/blob/master/keras/src/export/saved_model_test.py#L66 - https://github.com/keras-team/keras/blob/master/keras/src/export/onnx_test.py#L62 The goal is to isolate the use of `onnxruntime` to a single file, `onnx_test.py`. * Add OpenVINO into README.md (#20739) * Add OpenVINO into README.md Signed-off-by: Kazantsev, Roman <[email protected]> * Update README.md --------- Signed-off-by: Kazantsev, Roman <[email protected]> * Multiple Example Title has removed in metrics.MeanIoU method (#20738) Multiple Example Title has removed in metrics.MeanIoU method --------- Signed-off-by: Kazantsev, Roman <[email protected]> Co-authored-by: LakshmiKalaKadali <[email protected]> Co-authored-by: Ugeun Park <[email protected]> Co-authored-by: Martin Görner <[email protected]> Co-authored-by: hertschuh <[email protected]> Co-authored-by: Roman Kazantsev <[email protected]> Co-authored-by: LavanyaKV1234 <[email protected]>
keras-team · Jan 9, 2025 · e66ab50 · e66ab50
1 parent 1adaaec
commit e66ab50
Show file tree

Hide file tree

Showing 22 changed files with 489 additions and 121 deletions.
diff --git a/README.md b/README.md
@@ -1,6 +1,6 @@
 # Keras 3: Deep Learning for Humans
 
-Keras 3 is a multi-backend deep learning framework, with support for JAX, TensorFlow, and PyTorch.
+Keras 3 is a multi-backend deep learning framework, with support for JAX, TensorFlow, PyTorch, and OpenVINO (for inference-only).
 Effortlessly build and train models for computer vision, natural language processing, audio processing,
 timeseries forecasting, recommender systems, etc.
 
@@ -73,7 +73,7 @@ python pip_build.py --install
 ## Configuring your backend
 
 You can export the environment variable `KERAS_BACKEND` or you can edit your local config file at `~/.keras/keras.json`
-to configure your backend. Available backend options are: `"tensorflow"`, `"jax"`, `"torch"`. Example:
+to configure your backend. Available backend options are: `"tensorflow"`, `"jax"`, `"torch"`, `"openvino"`. Example:
 
 ```
 export KERAS_BACKEND="jax"
@@ -91,6 +91,10 @@ import keras
 **Note:** The backend must be configured before importing `keras`, and the backend cannot be changed after 
 the package has been imported.
 
+**Note:** The OpenVINO backend is an inference-only backend, meaning it is designed only for running model
+predictions using `model.predict()` method.
+To use `openvino` backend, install the required dependencies from the `requirements-openvino.txt` file.
+
 ## Backwards compatibility
 
 Keras 3 is intended to work as a drop-in replacement for `tf.keras` (when using the TensorFlow backend). Just take your

diff --git a/keras/api/_tf_keras/keras/layers/__init__.py b/keras/api/_tf_keras/keras/layers/__init__.py
@@ -152,6 +152,9 @@
     MaxNumBoundingBoxes,
 )
 from keras.src.layers.preprocessing.image_preprocessing.mix_up import MixUp
+from keras.src.layers.preprocessing.image_preprocessing.rand_augment import (
+    RandAugment,
+)
 from keras.src.layers.preprocessing.image_preprocessing.random_brightness import (
     RandomBrightness,
 )

diff --git a/keras/api/layers/__init__.py b/keras/api/layers/__init__.py
@@ -152,6 +152,9 @@
     MaxNumBoundingBoxes,
 )
 from keras.src.layers.preprocessing.image_preprocessing.mix_up import MixUp
+from keras.src.layers.preprocessing.image_preprocessing.rand_augment import (
+    RandAugment,
+)
 from keras.src.layers.preprocessing.image_preprocessing.random_brightness import (
     RandomBrightness,
 )

diff --git a/keras/src/backend/jax/distribution_lib.py b/keras/src/backend/jax/distribution_lib.py
@@ -100,7 +100,7 @@ def distribute_tensor(tensor, layout):
         return global_value
 
 
-def distribute_data_input(per_process_batch, layout):
+def distribute_data_input(per_process_batch, layout, batch_dim_name):
     """Distribute the input data with the corresponding layout.
 
     Note that the inputs here is a local worker batch. Within the local worker,
@@ -117,9 +117,13 @@ def distribute_data_input(per_process_batch, layout):
     if not isinstance(layout, jax.sharding.Sharding):
         layout = _to_jax_layout(layout)
 
-    mesh_shape = list(layout.mesh.shape.values())
-    num_model_replicas_total = mesh_shape[0]  # batch dimension of the mesh
-    mesh_model_dim_size = mesh_shape[1] if len(mesh_shape) > 1 else 1
+    num_model_replicas_total = layout.mesh.shape[batch_dim_name]
+
+    mesh_model_dim_size = 1
+    for name, dim_size in layout.mesh.shape.items():
+        if not name == batch_dim_name:
+            mesh_model_dim_size *= dim_size
+
     num_model_replicas_per_process = num_model_replicas_total / num_processes()
     per_process_batch_size = per_process_batch.shape[0]
 

diff --git a/keras/src/backend/jax/distribution_lib_test.py b/keras/src/backend/jax/distribution_lib_test.py
@@ -337,7 +337,9 @@ def test_distribute_data_input(self):
             mesh, jax.sharding.PartitionSpec("batch", None)
         )
 
-        result = backend_dlib.distribute_data_input(per_process_batch, layout)
+        result = backend_dlib.distribute_data_input(
+            per_process_batch, layout, "batch"
+        )
 
         # Check the shape of the global batch array
         self.assertEqual(

diff --git a/keras/src/backend/jax/export.py b/keras/src/backend/jax/export.py
@@ -119,7 +119,7 @@ def stateful_fn(*args, **kwargs):
                     self._tf_trackable.non_trainable_variables,
                     non_trainable_variables,
                 ):
-                    var.assign(new_value)
+                    var.assign(tf.cast(new_value, var.dtype))
                 return output
 
             stateful_fn.__signature__ = inspect.Signature(

diff --git a/keras/src/backend/jax/trainer.py b/keras/src/backend/jax/trainer.py
@@ -1,5 +1,6 @@
 import collections
 import itertools
+from functools import partial
 
 import jax
 import numpy as np
@@ -988,15 +989,18 @@ def _get_jax_state(
 
 def _distribute_data(data, layouts=None):
     distribution = distribution_lib.distribution()
+
     if distribution is not None:
         if layouts is None:
             layouts = tree.map_structure(
                 lambda d: distribution.get_data_layout(d.shape),
                 data,
             )
-        return tree.map_structure(
-            jax_distribution_lib.distribute_data_input, data, layouts
+        jax_dist_data_input = partial(
+            jax_distribution_lib.distribute_data_input,
+            batch_dim_name=distribution.batch_dim_name,
         )
+        return tree.map_structure(jax_dist_data_input, data, layouts)
 
     return tree.map_structure(jax.device_put, data)
 

diff --git a/keras/src/distribution/distribution_lib.py b/keras/src/distribution/distribution_lib.py
@@ -287,8 +287,9 @@ class Distribution:
         device_mesh: A `DeviceMesh` instance.
     """
 
-    def __init__(self, device_mesh):
+    def __init__(self, device_mesh, batch_dim_name=None):
         self._device_mesh = device_mesh
+        self._batch_dim_name = batch_dim_name
 
     def get_data_layout(self, data_shape):
         """Retrieve the `TensorLayout` for the input data.
@@ -341,6 +342,10 @@ def scope(self):
     def device_mesh(self):
         return self._device_mesh
 
+    @property
+    def batch_dim_name(self):
+        return self._batch_dim_name
+
     def distribute_dataset(self, dataset):
         """Create a distributed dataset instance from the original user dataset.
 
@@ -395,7 +400,6 @@ def __init__(self, device_mesh=None, devices=None, auto_shard_dataset=True):
         else:
             self._initialize_mesh_from_list_devices()
 
-        self._batch_dim_name = self.device_mesh.axis_names[0]
         # Those following attributes might get convert to public methods.
         self._num_process = distribution_lib.num_processes()
         self._process_id = distribution_lib.process_id()
@@ -408,7 +412,7 @@ def _initialize_with_device_mesh(self, device_mesh):
                 "Expect `mesh` to be an instance of `DeviceMesh`. "
                 f"Received: mesh={device_mesh} (of type {type(device_mesh)})"
             )
-        super().__init__(device_mesh)
+        super().__init__(device_mesh, device_mesh.axis_names[0])
         if self.device_mesh.devices.ndim != 1:
             warnings.warn(
                 "Expect the input mesh to be 1D, but received "
@@ -424,7 +428,7 @@ def _initialize_mesh_from_devices(self, devices):
             axis_names=[DEFAULT_BATCH_DIM_NAME],
             devices=devices,
         )
-        super().__init__(device_mesh)
+        super().__init__(device_mesh, DEFAULT_BATCH_DIM_NAME)
 
     def _initialize_mesh_from_list_devices(self):
         devices = np.array(list_devices())
@@ -433,11 +437,11 @@ def _initialize_mesh_from_list_devices(self):
             axis_names=[DEFAULT_BATCH_DIM_NAME],
             devices=devices,
         )
-        super().__init__(device_mesh)
+        super().__init__(device_mesh, DEFAULT_BATCH_DIM_NAME)
 
     def get_data_layout(self, data_shape):
         data_shard_spec = [None] * len(data_shape)
-        data_shard_spec[0] = self._batch_dim_name  # Shard on the first dim
+        data_shard_spec[0] = self.batch_dim_name  # Shard on the first dim
         return TensorLayout(data_shard_spec, self.device_mesh)
 
     def get_variable_layout(self, variable):
@@ -590,7 +594,7 @@ def __init__(self, *, layout_map=None, batch_dim_name=None, **kwargs):
 
     def get_data_layout(self, data_shape):
         data_shard_spec = [None] * len(data_shape)
-        data_shard_spec[0] = self._batch_dim_name  # Shard on the first dim
+        data_shard_spec[0] = self.batch_dim_name  # Shard on the first dim
         return TensorLayout(data_shard_spec, self.device_mesh)
 
     def get_variable_layout(self, variable):
@@ -631,7 +635,7 @@ def distribute_dataset(self, dataset):
         # Note that this might be smaller than one if model replicas are sharded
         # across multiple processes.
         mesh_batch_dim_index = self.device_mesh.axis_names.index(
-            self._batch_dim_name
+            self.batch_dim_name
         )
         num_model_replicas = self.device_mesh.shape[mesh_batch_dim_index]
         if num_model_replicas == 1:

diff --git a/keras/src/distribution/distribution_lib_test.py b/keras/src/distribution/distribution_lib_test.py
@@ -186,7 +186,7 @@ def test_create_with_device_mesh(self):
         device_mesh = distribution.device_mesh
         self.assertEqual(len(device_mesh.devices), 8)
         self.assertEqual(device_mesh.axis_names, ["data"])
-        self.assertEqual(distribution._batch_dim_name, "data")
+        self.assertEqual(distribution.batch_dim_name, "data")
 
         self.assertFalse(distribution._is_multi_process)
         self.assertEqual(distribution._process_id, 0)
@@ -197,7 +197,7 @@ def test_create_with_devices(self):
         device_mesh = distribution.device_mesh
         self.assertEqual(len(device_mesh.devices), 8)
         self.assertEqual(device_mesh.axis_names, ["batch"])
-        self.assertEqual(distribution._batch_dim_name, "batch")
+        self.assertEqual(distribution.batch_dim_name, "batch")
 
     @mock.patch.object(
         distribution_lib,
@@ -211,7 +211,7 @@ def test_create_with_list_devices(self, mock_list_devices):
         device_mesh = distribution.device_mesh
         self.assertEqual(len(device_mesh.devices), 8)
         self.assertEqual(device_mesh.axis_names, ["batch"])
-        self.assertEqual(distribution._batch_dim_name, "batch")
+        self.assertEqual(distribution.batch_dim_name, "batch")
 
     def test_get_data_layout(self):
         distribution = distribution_lib.DataParallel(

diff --git a/keras/src/layers/__init__.py b/keras/src/layers/__init__.py
@@ -96,6 +96,9 @@
     MaxNumBoundingBoxes,
 )
 from keras.src.layers.preprocessing.image_preprocessing.mix_up import MixUp
+from keras.src.layers.preprocessing.image_preprocessing.rand_augment import (
+    RandAugment,
+)
 from keras.src.layers.preprocessing.image_preprocessing.random_brightness import (
     RandomBrightness,
 )

diff --git a/keras/src/layers/layer.py b/keras/src/layers/layer.py
@@ -85,12 +85,10 @@ class Layer(BackendLayer, Operation, KerasSaveable):
         trainable: Boolean, whether the layer's variables should be trainable.
         name: String name of the layer.
         dtype: The dtype of the layer's computations and weights. Can also be a
-            `keras.DTypePolicy`,
-            which allows the computation and
-            weight dtype to differ. Defaults to `None`. `None` means to use
-            `keras.config.dtype_policy()`,
-            which is a `float32` policy unless set to different value
-            (via `keras.config.set_dtype_policy()`).
+            `keras.DTypePolicy`, which allows the computation and weight dtype
+            to differ. Defaults to `None`. `None` means to use
+            `keras.config.dtype_policy()`, which is a `float32` policy unless
+            set to different value (via `keras.config.set_dtype_policy()`).
 
     Attributes:
         name: The name of the layer (string).