Rework Model.export and keras.export.ExportArchive to support exporting in TFLite and ONNX formats in the future (#20631)

* Rework `Model.export` and `keras.export.ExportArchive`

* Try fixing PyDatasetAdapterTest CI issues

Author: james77777778

keras/src/backend/jax/export.py

@@ -0,0 +1,194 @@
+import copy
+import inspect
+import itertools
+import string
+import warnings
+
+from keras.src import layers
+from keras.src import tree
+from keras.src.backend.common.stateless_scope import StatelessScope
+from keras.src.utils.module_utils import tensorflow as tf
+
+
+class JaxExportArchive:
+    def __init__(self):
+        self._backend_variables = []
+        self._backend_trainable_variables = []
+        self._backend_non_trainable_variables = []
+
+    def track(self, resource):
+        if not isinstance(resource, layers.Layer):
+            raise ValueError(
+                "Invalid resource type. Expected an instance of a "
+                "JAX-based Keras `Layer` or `Model`. "
+                f"Received instead an object of type '{type(resource)}'. "
+                f"Object received: {resource}"
+            )
+
+        if isinstance(resource, layers.Layer):
+            # Variables in the lists below are actually part of the trackables
+            # that get saved, because the lists are created in __init__.
+            trainable_variables = resource.trainable_variables
+            non_trainable_variables = resource.non_trainable_variables
+
+            self._tf_trackable.trainable_variables += tree.map_structure(
+                self._convert_to_tf_variable, trainable_variables
+            )
+            self._tf_trackable.non_trainable_variables += tree.map_structure(
+                self._convert_to_tf_variable, non_trainable_variables
+            )
+            self._tf_trackable.variables = (
+                self._tf_trackable.trainable_variables
+                + self._tf_trackable.non_trainable_variables
+            )
+
+            self._backend_trainable_variables += trainable_variables
+            self._backend_non_trainable_variables += non_trainable_variables
+            self._backend_variables = (
+                self._backend_trainable_variables
+                + self._backend_non_trainable_variables
+            )
+
+    def add_endpoint(self, name, fn, input_signature=None, **kwargs):
+        jax2tf_kwargs = kwargs.pop("jax2tf_kwargs", None)
+        # Use `copy.copy()` to avoid modification issues.
+        jax2tf_kwargs = copy.copy(jax2tf_kwargs) or {}
+        is_static = bool(kwargs.pop("is_static", False))
+
+        # Configure `jax2tf_kwargs`
+        if "native_serialization" not in jax2tf_kwargs:
+            jax2tf_kwargs["native_serialization"] = (
+                self._check_device_compatible()
+            )
+        if "polymorphic_shapes" not in jax2tf_kwargs:
+            jax2tf_kwargs["polymorphic_shapes"] = self._to_polymorphic_shape(
+                input_signature
+            )
+
+        # Note: we truncate the number of parameters to what is specified by
+        # `input_signature`.
+        fn_signature = inspect.signature(fn)
+        fn_parameters = list(fn_signature.parameters.values())
+
+        if is_static:
+            from jax.experimental import jax2tf
+
+            jax_fn = jax2tf.convert(fn, **jax2tf_kwargs)
+            jax_fn.__signature__ = inspect.Signature(
+                parameters=fn_parameters[0 : len(input_signature)],
+                return_annotation=fn_signature.return_annotation,
+            )
+
+            decorated_fn = tf.function(
+                jax_fn,
+                input_signature=input_signature,
+                autograph=False,
+            )
+        else:
+            # 1. Create a stateless wrapper for `fn`
+            # 2. jax2tf the stateless wrapper
+            # 3. Create a stateful function that binds the variables with
+            #    the jax2tf converted stateless wrapper
+            # 4. Make the signature of the stateful function the same as the
+            #    original function
+            # 5. Wrap in a `tf.function`
+            def stateless_fn(variables, *args, **kwargs):
+                state_mapping = zip(self._backend_variables, variables)
+                with StatelessScope(state_mapping=state_mapping) as scope:
+                    output = fn(*args, **kwargs)
+
+                # Gather updated non-trainable variables
+                non_trainable_variables = []
+                for var in self._backend_non_trainable_variables:
+                    new_value = scope.get_current_value(var)
+                    non_trainable_variables.append(new_value)
+                return output, non_trainable_variables
+
+            jax2tf_stateless_fn = self._convert_jax2tf_function(
+                stateless_fn, input_signature, jax2tf_kwargs=jax2tf_kwargs
+            )
+
+            def stateful_fn(*args, **kwargs):
+                output, non_trainable_variables = jax2tf_stateless_fn(
+                    # Change the trackable `ListWrapper` to a plain `list`
+                    list(self._tf_trackable.variables),
+                    *args,
+                    **kwargs,
+                )
+                for var, new_value in zip(
+                    self._tf_trackable.non_trainable_variables,
+                    non_trainable_variables,
+                ):
+                    var.assign(new_value)
+                return output
+
+            stateful_fn.__signature__ = inspect.Signature(
+                parameters=fn_parameters[0 : len(input_signature)],
+                return_annotation=fn_signature.return_annotation,
+            )
+
+            decorated_fn = tf.function(
+                stateful_fn,
+                input_signature=input_signature,
+                autograph=False,
+            )
+        return decorated_fn
+
+    def _convert_jax2tf_function(self, fn, input_signature, jax2tf_kwargs=None):
+        from jax.experimental import jax2tf
+
+        variables_shapes = self._to_polymorphic_shape(
+            self._backend_variables, allow_none=False
+        )
+        input_shapes = list(jax2tf_kwargs["polymorphic_shapes"])
+        jax2tf_kwargs["polymorphic_shapes"] = [variables_shapes] + input_shapes
+        return jax2tf.convert(fn, **jax2tf_kwargs)
+
+    def _to_polymorphic_shape(self, struct, allow_none=True):
+        if allow_none:
+            # Generates unique names: a, b, ... z, aa, ab, ... az, ba, ... zz
+            # for unknown non-batch dims. Defined here to be scope per endpoint.
+            dim_names = itertools.chain(
+                string.ascii_lowercase,
+                itertools.starmap(
+                    lambda a, b: a + b,
+                    itertools.product(string.ascii_lowercase, repeat=2),
+                ),
+            )
+
+        def convert_shape(x):
+            poly_shape = []
+            for index, dim in enumerate(list(x.shape)):
+                if dim is not None:
+                    poly_shape.append(str(dim))
+                elif not allow_none:
+                    raise ValueError(
+                        f"Illegal None dimension in {x} with shape {x.shape}"
+                    )
+                elif index == 0:
+                    poly_shape.append("batch")
+                else:
+                    poly_shape.append(next(dim_names))
+            return "(" + ", ".join(poly_shape) + ")"
+
+        return tree.map_structure(convert_shape, struct)
+
+    def _check_device_compatible(self):
+        from jax import default_backend as jax_device
+
+        if (
+            jax_device() == "gpu"
+            and len(tf.config.list_physical_devices("GPU")) == 0
+        ):
+            warnings.warn(
+                "JAX backend is using GPU for export, but installed "
+                "TF package cannot access GPU, so reloading the model with "
+                "the TF runtime in the same environment will not work. "
+                "To use JAX-native serialization for high-performance export "
+                "and serving, please install `tensorflow-gpu` and ensure "
+                "CUDA version compatibility between your JAX and TF "
+                "installations."
+            )
+            return False
+        else:
+            return True

keras/src/backend/numpy/export.py

@@ -0,0 +1,10 @@
+class NumpyExportArchive:
+    def track(self, resource):
+        raise NotImplementedError(
+            "`track` is not implemented in the numpy backend."
+        )
+
+    def add_endpoint(self, name, fn, input_signature=None, **kwargs):
+        raise NotImplementedError(
+            "`add_endpoint` is not implemented in the numpy backend."
+        )

keras/src/backend/tensorflow/export.py

@@ -0,0 +1,32 @@
+import tensorflow as tf
+
+from keras.src import layers
+
+
+class TFExportArchive:
+    def track(self, resource):
+        if not isinstance(resource, tf.__internal__.tracking.Trackable):
+            raise ValueError(
+                "Invalid resource type. Expected an instance of a "
+                "TensorFlow `Trackable` (such as a Keras `Layer` or `Model`). "
+                f"Received instead an object of type '{type(resource)}'. "
+                f"Object received: {resource}"
+            )
+
+        if isinstance(resource, layers.Layer):
+            # Variables in the lists below are actually part of the trackables
+            # that get saved, because the lists are created in __init__.
+            variables = resource.variables
+            trainable_variables = resource.trainable_variables
+            non_trainable_variables = resource.non_trainable_variables
+            self._tf_trackable.variables += variables
+            self._tf_trackable.trainable_variables += trainable_variables
+            self._tf_trackable.non_trainable_variables += (
+                non_trainable_variables
+            )
+
+    def add_endpoint(self, name, fn, input_signature=None, **kwargs):
+        decorated_fn = tf.function(
+            fn, input_signature=input_signature, autograph=False
+        )
+        return decorated_fn

keras/src/backend/torch/export.py

@@ -0,0 +1,35 @@
+from keras.src import layers
+from keras.src import tree
+
+
+class TorchExportArchive:
+    def track(self, resource):
+        if not isinstance(resource, layers.Layer):
+            raise ValueError(
+                "Invalid resource type. Expected an instance of a "
+                "JAX-based Keras `Layer` or `Model`. "
+                f"Received instead an object of type '{type(resource)}'. "
+                f"Object received: {resource}"
+            )
+
+        if isinstance(resource, layers.Layer):
+            # Variables in the lists below are actually part of the trackables
+            # that get saved, because the lists are created in __init__.
+            variables = resource.variables
+            trainable_variables = resource.trainable_variables
+            non_trainable_variables = resource.non_trainable_variables
+            self._tf_trackable.variables += tree.map_structure(
+                self._convert_to_tf_variable, variables
+            )
+            self._tf_trackable.trainable_variables += tree.map_structure(
+                self._convert_to_tf_variable, trainable_variables
+            )
+            self._tf_trackable.non_trainable_variables += tree.map_structure(
+                self._convert_to_tf_variable, non_trainable_variables
+            )
+
+    def add_endpoint(self, name, fn, input_signature=None, **kwargs):
+        # TODO: torch-xla?
+        raise NotImplementedError(
+            "`add_endpoint` is not implemented in the torch backend."
+        )