Merge master into mlx (#19557)

* Introduce float8 training (#19488)

* Add float8 training support

* Add tests for fp8 training

* Add `quantize_and_dequantize` test

* Fix bugs and add float8 correctness tests

* Cleanup

* Address comments and cleanup

* Add docstrings and some minor refactoring

* Add `QuantizedFloat8DTypePolicy`

* Add dtype policy setter

* Fix torch dynamo issue by using `self._dtype_policy`

* Improve test coverage

* Add LoRA to ConvND layers (#19516)

* Add LoRA to `BaseConv`

* Add tests

* Fix typo

* Fix tests

* Fix tests

* Add path to run keras on dm-tree when optree is not available.

* feat(losses): add Tversky loss implementation (#19511)

* feat(losses): add Tversky loss implementation

* adjusted documentation

* Update KLD docs

* Models and layers now return owned metrics recursively. (#19522)

- added `Layer.metrics` to return all metrics owned by the layer and its sub-layers recursively.
- `Layer.metrics_variables` now returns variables from all metrics recursively, not just the layer and its direct sub-layers.
- `Model.metrics` now returns all metrics recursively, not just the model level metrics.
- `Model.metrics_variables` now returns variables from all metrics recursively, not just the model level metrics.
- added test coverage to test metrics and variables 2 levels deep.

This is consistent with the Keras 2 behavior and how `Model/Layer.variables` and `Model/Layer.weights` work.

* Update IoU ignore_class handling

* Fix `RandomBrightness`, Enhance `IndexLookup` Initialization and Expand Test Coverage for `Preprocessing Layers` (#19513)

* Add tests for CategoryEncoding class in category_encoding_test.py

* fix

* Fix IndexLookup class initialization and add test cases

* Add test case for IndexLookupLayerTest without vocabulary

* Fix IndexLookup class initialization

* Add normalization test cases

* Add test cases for Hashing class

* Fix value range validation error in RandomBrightness class

* Refactor IndexLookup class initialization and add test cases

* Reffix ndexLookup class initialization and afix est cases

* Add test for spectral norm

* Add missing test decorator

* Fix torch test

* Fix code format

* Generate API (#19530)

* API Generator for Keras

* API Generator for Keras

* Generates API Gen via api_gen.sh

* Remove recursive import of _tf_keras

* Generate API Files via api_gen.sh

* Update APIs

* Added metrics from custom `train_step`/`test_step` are now returned. (#19529)

This works the same way as in Keras 2, whereby the metrics are returned directly from the logs if the set of keys doesn't match the model metrics.

* Use temp dir and abs path in `api_gen.py` (#19533)

* Use temp dir and abs path

* Use temp dir and abs path

* Update Readme

* Update API

* Fix gradient accumulation when using `overwrite_with_gradient` during float8 training (#19534)

* Fix gradient accumulation with `overwrite_with_gradient` in float8 training

* Add comments

* Fix annotation

* Update code path in ignore path (#19537)

* Add operations per run (#19538)

* Include input shapes in model visualization.

* Add pad_to_aspect_ratio feature in ops.image.resize

* Add pad_to_aspect_ratio feature in Resizing layer.

* Fix incorrect usage of `quantize` (#19541)

* Add logic to prevent double quantization

* Add detailed info for double quantization error

* Update error msg

* Add eigh op.

* Add keepdim in argmax/argmin.

* Fix small bug in model.save_weights (#19545)

* Update public APIs.

* eigh should work on JAX GPU

* Copy init to keras/__init__.py (#19551)

* Revert "Copy init to keras/__init__.py (#19551)" (#19552)

This reverts commit da9af61.

* fixes for master

---------

Co-authored-by: james77777778 <[email protected]>
Co-authored-by: Francois Chollet <[email protected]>
Co-authored-by: Luca Pizzini <[email protected]>
Co-authored-by: hertschuh <[email protected]>
Co-authored-by: Faisal Alsrheed <[email protected]>
Co-authored-by: Ramesh Sampath <[email protected]>
Co-authored-by: Sachin Prasad <[email protected]>
Co-authored-by: Uwe Schmidt <[email protected]>

Author: 9 people

.github/workflows/actions.yml

@@ -24,13 +24,13 @@ jobs:
       KERAS_HOME: .github/workflows/config/${{ matrix.backend }}
     steps:
       - uses: actions/checkout@v4
-      - name: Check for changes in keras/applications
+      - name: Check for changes in keras/src/applications
         uses: dorny/paths-filter@v3
         id: filter
         with:
           filters: |
             applications:
-              - 'keras/applications/**'
+              - 'keras/src/applications/**'
       - name: Set up Python
         uses: actions/setup-python@v5
         with:
@@ -49,13 +49,13 @@ jobs:
         run: |
           pip install -r requirements.txt --progress-bar off --upgrade
           pip uninstall -y keras keras-nightly
-          pip install tf_keras==2.16.0rc0 --progress-bar off --upgrade
+          pip install tf_keras==2.16.0 --progress-bar off --upgrade
           pip install -e "." --progress-bar off --upgrade
       - name: Test applications with pytest
         if: ${{ steps.filter.outputs.applications == 'true' }}
         run: |
-          pytest keras/applications --cov=keras/applications
-          coverage xml --include='keras/applications/*' -o apps-coverage.xml
+          pytest keras/src/applications --cov=keras/src/applications
+          coverage xml --include='keras/src/applications/*' -o apps-coverage.xml
       - name: Codecov keras.applications
         if: ${{ steps.filter.outputs.applications == 'true' }}
         uses: codecov/codecov-action@v4
@@ -80,8 +80,8 @@ jobs:
           pytest integration_tests/torch_workflow_test.py
       - name: Test with pytest
         run: |
-          pytest keras --ignore keras/applications --cov=keras
-          coverage xml --omit='keras/applications/*' -o core-coverage.xml
+          pytest keras --ignore keras/src/applications --cov=keras
+          coverage xml --omit='keras/src/applications/*,keras/api' -o core-coverage.xml
       - name: Codecov keras
         uses: codecov/codecov-action@v4
         with:
@@ -115,5 +115,14 @@ jobs:
           pip install -r requirements.txt --progress-bar off --upgrade
           pip uninstall -y keras keras-nightly
           pip install -e "." --progress-bar off --upgrade
+      - name: Check for API changes
+        run: |
+          bash shell/api_gen.sh
+          git status
+          clean=$(git status | grep "nothing to commit")
+          if [ -z "$clean" ]; then
+            echo "Please run shell/api_gen.sh to generate API."
+            exit 1
+          fi
       - name: Lint
         run: bash shell/lint.sh

.github/workflows/nightly.yml

@@ -55,7 +55,7 @@ jobs:
           pytest integration_tests/torch_workflow_test.py
       - name: Test with pytest
         run: |
-          pytest keras --ignore keras/applications --cov=keras
+          pytest keras --ignore keras/src/applications --cov=keras
 
   format:
     name: Check the code format
@@ -81,6 +81,15 @@ jobs:
           pip install -r requirements.txt --progress-bar off --upgrade
           pip uninstall -y keras keras-nightly
           pip install -e "." --progress-bar off --upgrade
+      - name: Check for API changes
+        run: |
+          bash shell/api_gen.sh
+          git status
+          clean=$(git status | grep "nothing to commit")
+          if [ -z "$clean" ]; then
+            echo "Please run shell/api_gen.sh to generate API."
+            exit 1
+          fi
       - name: Lint
         run: bash shell/lint.sh
 
@@ -108,4 +117,4 @@ jobs:
         with:
           password: ${{ secrets.PYPI_NIGHTLY_API_TOKEN }}
           packages-dir: dist/
-          verbose: true
+          verbose: true

.github/workflows/stale-issue-pr.yaml

@@ -4,6 +4,8 @@ on:
     - cron: "30 1 * * *"
 jobs:
   close-issues:
+    # Don't do this in forks
+    if: github.repository == 'keras-team/keras'
     runs-on: ubuntu-latest
     permissions:
       issues: write
@@ -12,6 +14,7 @@ jobs:
       - name: Awaiting response issues
         uses: actions/stale@v9
         with:
+          operations-per-run: 500
           days-before-issue-stale: 14
           days-before-issue-close: 14
           stale-issue-label: "stale"
@@ -34,6 +37,7 @@ jobs:
       - name: Contribution issues
         uses: actions/stale@v9
         with:
+          operations-per-run: 500
           days-before-issue-stale: 180
           days-before-issue-close: 365
           stale-issue-label: "stale"

.kokoro/github/ubuntu/gpu/build.sh

@@ -34,8 +34,8 @@ then
    python3 -c 'import tensorflow as tf;assert len(tf.config.list_physical_devices("GPU")) > 0'
 
    # TODO: keras/layers/merging/merging_test.py::MergingLayersTest::test_sparse_dot_2d Fatal Python error: Aborted
-   pytest keras --ignore keras/applications \
-               --ignore keras/layers/merging/merging_test.py \
+   pytest keras --ignore keras/src/applications \
+               --ignore keras/src/layers/merging/merging_test.py \
                --cov=keras
 fi
 
@@ -51,11 +51,11 @@ then
    # TODO: keras/layers/merging/merging_test.py::MergingLayersTest::test_sparse_dot_2d Fatal Python error: Aborted
    # TODO: keras/trainers/data_adapters/py_dataset_adapter_test.py::PyDatasetAdapterTest::test_basic_flow0 Fatal Python error: Aborted
    # keras/backend/jax/distribution_lib_test.py is configured for CPU test for now.
-   pytest keras --ignore keras/applications \
-               --ignore keras/layers/merging/merging_test.py \
-               --ignore keras/trainers/data_adapters/py_dataset_adapter_test.py \
-               --ignore keras/backend/jax/distribution_lib_test.py \
-               --ignore keras/distribution/distribution_lib_test.py \
+   pytest keras --ignore keras/src/applications \
+               --ignore keras/src/layers/merging/merging_test.py \
+               --ignore keras/src/trainers/data_adapters/py_dataset_adapter_test.py \
+               --ignore keras/src/backend/jax/distribution_lib_test.py \
+               --ignore keras/src/distribution/distribution_lib_test.py \
                --cov=keras
 fi
 
@@ -68,6 +68,6 @@ then
    # Raise error if GPU is not detected.
    python3 -c 'import torch;assert torch.cuda.is_available()'
 
-   pytest keras --ignore keras/applications \
+   pytest keras --ignore keras/src/applications \
                --cov=keras
 fi

README.md

@@ -50,6 +50,12 @@ pip install -r requirements.txt
 python pip_build.py --install
 ```
 
+3. Run API generation script when creating PRs that update `keras_export` public APIs:
+
+```
+./shell/api_gen.sh
+```
+
 #### Adding GPU support
 
 The `requirements.txt` file will install a CPU-only version of TensorFlow, JAX, and PyTorch. For GPU support, we also

api_gen.py

@@ -0,0 +1,201 @@
+"""Script to generate keras public API in `keras/api` directory.
+
+Usage:
+
+Run via `./shell/api_gen.sh`.
+It generates API and formats user and generated APIs.
+"""
+
+import os
+import shutil
+
+import namex
+
+package = "keras"
+
+
+def ignore_files(_, filenames):
+    return [f for f in filenames if f.endswith("_test.py")]
+
+
+def copy_source_to_build_directory(root_path):
+    # Copy sources (`keras/` directory and setup files) to build dir
+    build_dir = os.path.join(root_path, "tmp_build_dir")
+    if os.path.exists(build_dir):
+        shutil.rmtree(build_dir)
+    os.mkdir(build_dir)
+    shutil.copytree(
+        package, os.path.join(build_dir, package), ignore=ignore_files
+    )
+    return build_dir
+
+
+def create_legacy_directory(package_dir):
+    src_dir = os.path.join(package_dir, "src")
+    api_dir = os.path.join(package_dir, "api")
+    # Make keras/_tf_keras/ by copying keras/
+    tf_keras_dirpath_parent = os.path.join(api_dir, "_tf_keras")
+    tf_keras_dirpath = os.path.join(tf_keras_dirpath_parent, "keras")
+    os.makedirs(tf_keras_dirpath, exist_ok=True)
+    with open(os.path.join(tf_keras_dirpath_parent, "__init__.py"), "w") as f:
+        f.write("from keras.api._tf_keras import keras\n")
+    with open(os.path.join(api_dir, "__init__.py")) as f:
+        init_file = f.read()
+        init_file = init_file.replace(
+            "from keras.api import _legacy",
+            "from keras.api import _tf_keras",
+        )
+    with open(os.path.join(api_dir, "__init__.py"), "w") as f:
+        f.write(init_file)
+    # Remove the import of `_tf_keras` in `keras/_tf_keras/keras/__init__.py`
+    init_file = init_file.replace("from keras.api import _tf_keras\n", "\n")
+    with open(os.path.join(tf_keras_dirpath, "__init__.py"), "w") as f:
+        f.write(init_file)
+    for dirname in os.listdir(api_dir):
+        dirpath = os.path.join(api_dir, dirname)
+        if os.path.isdir(dirpath) and dirname not in (
+            "_legacy",
+            "_tf_keras",
+            "src",
+        ):
+            destpath = os.path.join(tf_keras_dirpath, dirname)
+            if os.path.exists(destpath):
+                shutil.rmtree(destpath)
+            shutil.copytree(
+                dirpath,
+                destpath,
+                ignore=ignore_files,
+            )
+
+    # Copy keras/_legacy/ file contents to keras/_tf_keras/keras
+    legacy_submodules = [
+        path[:-3]
+        for path in os.listdir(os.path.join(src_dir, "legacy"))
+        if path.endswith(".py")
+    ]
+    legacy_submodules += [
+        path
+        for path in os.listdir(os.path.join(src_dir, "legacy"))
+        if os.path.isdir(os.path.join(src_dir, "legacy", path))
+    ]
+
+    for root, _, fnames in os.walk(os.path.join(package_dir, "_legacy")):
+        for fname in fnames:
+            if fname.endswith(".py"):
+                legacy_fpath = os.path.join(root, fname)
+                tf_keras_root = root.replace("/_legacy", "/_tf_keras/keras")
+                core_api_fpath = os.path.join(
+                    root.replace("/_legacy", ""), fname
+                )
+                if not os.path.exists(tf_keras_root):
+                    os.makedirs(tf_keras_root)
+                tf_keras_fpath = os.path.join(tf_keras_root, fname)
+                with open(legacy_fpath) as f:
+                    legacy_contents = f.read()
+                    legacy_contents = legacy_contents.replace(
+                        "keras.api._legacy", "keras.api._tf_keras.keras"
+                    )
+                if os.path.exists(core_api_fpath):
+                    with open(core_api_fpath) as f:
+                        core_api_contents = f.read()
+                    core_api_contents = core_api_contents.replace(
+                        "from keras.api import _tf_keras\n", ""
+                    )
+                    for legacy_submodule in legacy_submodules:
+                        core_api_contents = core_api_contents.replace(
+                            f"from keras.api import {legacy_submodule}\n",
+                            "",
+                        )
+                        core_api_contents = core_api_contents.replace(
+                            f"keras.api.{legacy_submodule}",
+                            f"keras.api._tf_keras.keras.{legacy_submodule}",
+                        )
+                    legacy_contents = core_api_contents + "\n" + legacy_contents
+                with open(tf_keras_fpath, "w") as f:
+                    f.write(legacy_contents)
+
+    # Delete keras/api/_legacy/
+    shutil.rmtree(os.path.join(api_dir, "_legacy"))
+
+
+def export_version_string(api_init_fname):
+    with open(api_init_fname) as f:
+        contents = f.read()
+    with open(api_init_fname, "w") as f:
+        contents += "from keras.src.version import __version__\n"
+        f.write(contents)
+
+
+def update_package_init(init_fname):
+    contents = """
+# Import everything from /api/ into keras.
+from keras.api import *  # noqa: F403
+from keras.api import __version__  # Import * ignores names start with "_".
+
+import os
+
+# Add everything in /api/ to the module search path.
+__path__.append(os.path.join(os.path.dirname(__file__), "api"))  # noqa: F405
+
+# Don't pollute namespace.
+del os
+
+# Never autocomplete `.src` or `.api` on an imported keras object.
+def __dir__():
+    keys = dict.fromkeys((globals().keys()))
+    keys.pop("src")
+    keys.pop("api")
+    return list(keys)
+
+
+# Don't import `.src` or `.api` during `from keras import *`.
+__all__ = [
+    name
+    for name in globals().keys()
+    if not (name.startswith("_") or name in ("src", "api"))
+]"""
+    with open(init_fname) as f:
+        init_contents = f.read()
+    with open(init_fname, "w") as f:
+        f.write(init_contents.replace("\nfrom keras import api", contents))
+
+
+def build():
+    # Backup the `keras/__init__.py` and restore it on error in api gen.
+    root_path = os.path.dirname(os.path.abspath(__file__))
+    code_api_dir = os.path.join(root_path, package, "api")
+    code_init_fname = os.path.join(root_path, package, "__init__.py")
+    # Create temp build dir
+    build_dir = copy_source_to_build_directory(root_path)
+    build_api_dir = os.path.join(build_dir, package, "api")
+    build_init_fname = os.path.join(build_dir, package, "__init__.py")
+    build_api_init_fname = os.path.join(build_api_dir, "__init__.py")
+    try:
+        os.chdir(build_dir)
+        # Generates `keras/api` directory.
+        if os.path.exists(build_api_dir):
+            shutil.rmtree(build_api_dir)
+        if os.path.exists(build_init_fname):
+            os.remove(build_init_fname)
+        os.makedirs(build_api_dir)
+        namex.generate_api_files(
+            "keras", code_directory="src", target_directory="api"
+        )
+        # Creates `keras/__init__.py` importing from `keras/api`
+        update_package_init(build_init_fname)
+        # Add __version__ to keras package
+        export_version_string(build_api_init_fname)
+        # Creates `_tf_keras` with full keras API
+        create_legacy_directory(package_dir=os.path.join(build_dir, package))
+        # Copy back the keras/api and keras/__init__.py from build directory
+        if os.path.exists(code_api_dir):
+            shutil.rmtree(code_api_dir)
+        shutil.copytree(build_api_dir, code_api_dir)
+        shutil.copy(build_init_fname, code_init_fname)
+    finally:
+        # Clean up: remove the build directory (no longer needed)
+        shutil.rmtree(build_dir)
+
+
+if __name__ == "__main__":
+    build()

conftest.py

@@ -14,7 +14,7 @@
 
 import pytest  # noqa: E402
 
-from keras.backend import backend  # noqa: E402
+from keras.src.backend import backend  # noqa: E402
 
 
 def pytest_configure(config):

integration_tests/basic_full_flow.py

@@ -2,11 +2,11 @@
 import pytest
 
 import keras
-from keras import layers
-from keras import losses
-from keras import metrics
-from keras import optimizers
-from keras import testing
+from keras.src import layers
+from keras.src import losses
+from keras.src import metrics
+from keras.src import optimizers
+from keras.src import testing
 
 
 class MyModel(keras.Model):