tf-keras-nightly 2.17.0.dev2024050509__py3-none-any.whl → 2.19.0.dev2024101709__py3-none-any.whl

tf_keras/src/layers/rnn/legacy_cells.py

@@ -246,11 +246,6 @@ class RNNCell(base_layer.Layer):
         """Integer or TensorShape: size of outputs produced by this cell."""
         raise NotImplementedError("Abstract method")
 
-    def build(self, _):
-        # This tells the parent Layer object that it's OK to call
-        # self.add_weight() inside the call() method.
-        pass
-
     def get_initial_state(self, inputs=None, batch_size=None, dtype=None):
         if inputs is not None:
             # Validate the given batch_size and dtype against inputs if
@@ -445,15 +440,15 @@ class BasicRNNCell(LayerRNNCell):
         return self._num_units
 
     @tf_utils.shape_type_conversion
-    def build(self, inputs_shape):
-        if inputs_shape[-1] is None:
+    def build(self, input_shape):
+        if input_shape[-1] is None:
             raise ValueError(
                 "Expected inputs.shape[-1] to be known, "
-                f"received shape: {inputs_shape}"
+                f"received shape: {input_shape}"
             )
         _check_supported_dtypes(self.dtype)
 
-        input_depth = inputs_shape[-1]
+        input_depth = input_shape[-1]
         self._kernel = self.add_weight(
             _WEIGHTS_VARIABLE_NAME,
             shape=[input_depth + self._num_units, self._num_units],
@@ -464,7 +459,7 @@ class BasicRNNCell(LayerRNNCell):
             initializer=tf.compat.v1.zeros_initializer(dtype=self.dtype),
         )
 
-        self.built = True
+        super().build(input_shape)
 
     def call(self, inputs, state):
         """Most basic RNN: output = new_state = act(W * input + U * state +
@@ -563,14 +558,14 @@ class GRUCell(LayerRNNCell):
         return self._num_units
 
     @tf_utils.shape_type_conversion
-    def build(self, inputs_shape):
-        if inputs_shape[-1] is None:
+    def build(self, input_shape):
+        if input_shape[-1] is None:
             raise ValueError(
                 "Expected inputs.shape[-1] to be known, "
-                f"received shape: {inputs_shape}"
+                f"received shape: {input_shape}"
             )
         _check_supported_dtypes(self.dtype)
-        input_depth = inputs_shape[-1]
+        input_depth = input_shape[-1]
         self._gate_kernel = self.add_weight(
             f"gates/{_WEIGHTS_VARIABLE_NAME}",
             shape=[input_depth + self._num_units, 2 * self._num_units],
@@ -600,7 +595,7 @@ class GRUCell(LayerRNNCell):
             ),
         )
 
-        self.built = True
+        super().build(input_shape)
 
     def call(self, inputs, state):
         """Gated recurrent unit (GRU) with nunits cells."""
@@ -774,14 +769,14 @@ class BasicLSTMCell(LayerRNNCell):
         return self._num_units
 
     @tf_utils.shape_type_conversion
-    def build(self, inputs_shape):
-        if inputs_shape[-1] is None:
+    def build(self, input_shape):
+        if input_shape[-1] is None:
             raise ValueError(
                 "Expected inputs.shape[-1] to be known, "
-                f"received shape: {inputs_shape}"
+                f"received shape: {input_shape}"
             )
         _check_supported_dtypes(self.dtype)
-        input_depth = inputs_shape[-1]
+        input_depth = input_shape[-1]
         h_depth = self._num_units
         self._kernel = self.add_weight(
             _WEIGHTS_VARIABLE_NAME,
@@ -793,7 +788,7 @@ class BasicLSTMCell(LayerRNNCell):
             initializer=tf.compat.v1.zeros_initializer(dtype=self.dtype),
         )
 
-        self.built = True
+        super().build(input_shape)
 
     def call(self, inputs, state):
         """Long short-term memory cell (LSTM).
@@ -1017,14 +1012,14 @@ class LSTMCell(LayerRNNCell):
         return self._output_size
 
     @tf_utils.shape_type_conversion
-    def build(self, inputs_shape):
-        if inputs_shape[-1] is None:
+    def build(self, input_shape):
+        if input_shape[-1] is None:
             raise ValueError(
                 "Expected inputs.shape[-1] to be known, "
-                f"received shape: {inputs_shape}"
+                f"received shape: {input_shape}"
             )
         _check_supported_dtypes(self.dtype)
-        input_depth = inputs_shape[-1]
+        input_depth = input_shape[-1]
         h_depth = self._num_units if self._num_proj is None else self._num_proj
         maybe_partitioner = (
             tf.compat.v1.fixed_size_partitioner(self._num_unit_shards)
@@ -1076,7 +1071,7 @@ class LSTMCell(LayerRNNCell):
                 partitioner=maybe_proj_partitioner,
             )
 
-        self.built = True
+        super().build(input_shape)
 
     def call(self, inputs, state):
         """Run one step of LSTM.

tf_keras/src/layers/rnn/lstm.py

@@ -236,7 +236,6 @@ class LSTMCell(DropoutRNNCellMixin, base_layer.BaseRandomLayer):
             )
         else:
             self.bias = None
-        self.built = True
 
     def _compute_carry_and_output(self, x, h_tm1, c_tm1):
         """Computes carry and output using split kernels."""
@@ -1063,11 +1062,13 @@ def gpu_lstm(
             mask, time_major
         )
 
-    if not time_major and sequence_lengths is None:
-        inputs = tf.transpose(inputs, perm=(1, 0, 2))
-        seq_axis, batch_axis = (0, 1)
-    else:
-        seq_axis, batch_axis = (0, 1) if time_major else (1, 0)
+    seq_axis, batch_axis = (0, 1) if time_major else (1, 0)
+
+    if sequence_lengths is None:
+        max_sequence_length = tf.shape(inputs)[seq_axis]
+        batch_size = tf.shape(inputs)[batch_axis]
+        sequence_lengths = tf.fill([batch_size], max_sequence_length)
+
     # For init_h and init_c, cuDNN expects one more dim of num_layers before or
     # after batch dim for time major or batch major inputs respectively
     init_h = tf.expand_dims(init_h, axis=seq_axis)
@@ -1099,52 +1100,36 @@ def gpu_lstm(
         transpose_weights=True,
     )
 
-    if sequence_lengths is not None:
-        if go_backwards:
-            # Three reversals are required. E.g.,
-            # normal input = [1, 2, 3, 0, 0]  # where 0 need to be masked
-            # reversed_input_to_cudnn = [3, 2, 1, 0, 0]
-            # output_from_cudnn = [6, 5, 4, 0, 0]
-            # expected_output = [0, 0, 6, 5 ,4]
-            inputs = tf.reverse_sequence(
-                inputs,
-                sequence_lengths,
-                seq_axis=seq_axis,
-                batch_axis=batch_axis,
-            )
-        outputs, h, c, _, _ = tf.raw_ops.CudnnRNNV3(
-            input=inputs,
-            input_h=init_h,
-            input_c=init_c,
-            params=params,
-            is_training=True,
-            rnn_mode="lstm",
-            sequence_lengths=sequence_lengths,
-            time_major=time_major,
+    if go_backwards:
+        # Three reversals are required. E.g.,
+        # normal input = [1, 2, 3, 0, 0]  # where 0 need to be masked
+        # reversed_input_to_cudnn = [3, 2, 1, 0, 0]
+        # output_from_cudnn = [6, 5, 4, 0, 0]
+        # expected_output = [0, 0, 6, 5 ,4]
+        inputs = tf.reverse_sequence(
+            inputs,
+            sequence_lengths,
+            seq_axis=seq_axis,
+            batch_axis=batch_axis,
         )
-        if go_backwards:
-            outputs = tf.reverse_sequence(
-                outputs,
-                sequence_lengths,
-                seq_axis=seq_axis,
-                batch_axis=batch_axis,
-            )
-            outputs = tf.reverse(outputs, axis=[seq_axis])
-    else:
-        # # Fill the array with shape [batch] with value of max timesteps.
-        # sequence_length = array_ops.fill([array_ops.shape(inputs)[1]],
-        #                                  array_ops.shape(inputs)[0])
-        if go_backwards:
-            # Reverse axis 0 since the input is already convert to time major.
-            inputs = tf.reverse(inputs, axis=[0])
-        outputs, h, c, _ = tf.raw_ops.CudnnRNN(
-            input=inputs,
-            input_h=init_h,
-            input_c=init_c,
-            params=params,
-            is_training=True,
-            rnn_mode="lstm",
+    outputs, h, c, _, _ = tf.raw_ops.CudnnRNNV3(
+        input=inputs,
+        input_h=init_h,
+        input_c=init_c,
+        params=params,
+        is_training=True,
+        rnn_mode="lstm",
+        sequence_lengths=sequence_lengths,
+        time_major=time_major,
+    )
+    if go_backwards:
+        outputs = tf.reverse_sequence(
+            outputs,
+            sequence_lengths,
+            seq_axis=seq_axis,
+            batch_axis=batch_axis,
         )
+        outputs = tf.reverse(outputs, axis=[seq_axis])
 
     last_output = outputs[-1]
     if not time_major and sequence_lengths is None and return_sequences:

tf_keras/src/layers/rnn/simple_rnn.py

@@ -189,7 +189,6 @@ class SimpleRNNCell(DropoutRNNCellMixin, base_layer.BaseRandomLayer):
             )
         else:
             self.bias = None
-        self.built = True
 
     def call(self, inputs, states, training=None):
         prev_output = states[0] if tf.nest.is_nested(states) else states

tf_keras/src/layers/rnn/stacked_rnn_cells.py

@@ -166,6 +166,7 @@ class StackedRNNCells(base_layer.Layer):
 
     @tf_utils.shape_type_conversion
     def build(self, input_shape):
+        super().build(input_shape)
         if isinstance(input_shape, list):
             input_shape = input_shape[0]
 
@@ -195,7 +196,6 @@ class StackedRNNCells(base_layer.Layer):
                 input_shape = tuple(
                     [batch_size] + tf.TensorShape(output_dim).as_list()
                 )
-        self.built = True
 
     def get_config(self):
         cells = []

tf_keras/src/layers/rnn/time_distributed.py

@@ -135,7 +135,6 @@ class TimeDistributed(Wrapper):
         )
         child_input_shape = tf_utils.convert_shapes(child_input_shape)
         super().build(tuple(child_input_shape))
-        self.built = True
 
     def compute_output_shape(self, input_shape):
         input_shape = tf_utils.convert_shapes(input_shape, to_tuples=False)

tf_keras/src/mixed_precision/autocast_variable.py

@@ -124,20 +124,21 @@ class AutoCastVariable(tf.Variable, tf.__internal__.types.Tensor):
     def _should_cast(self):
         """Returns True if this variable should be casted when accessed."""
         autocast_dtype = getattr(_autocast_dtype, "dtype", None)
-        return autocast_dtype is not None and self.dtype != autocast_dtype
+        return autocast_dtype is not None and self.true_dtype != autocast_dtype
 
     @property
     def dtype(self):
-        """The dtype of the underlying variable, before any casts are done."""
-        return self._variable.dtype
+        """The dtype when the value is accessed, that is after casting."""
+        return self._cast_dtype
 
     @property
     def true_dtype(self):
-        """Deprecated alias of `dtype`."""
+        """The dtype of the underlying variable, before any casts are done."""
         return self._variable.dtype
 
     @property
     def _cast_dtype(self):
+        """The dtype after casting."""
         dtype = getattr(_autocast_dtype, "dtype", None)
         return dtype or self._variable.dtype
 
@@ -202,7 +203,8 @@ class AutoCastVariable(tf.Variable, tf.__internal__.types.Tensor):
         if tf.executing_eagerly() and not self._in_graph_mode:
             repr_str = (
                 "<AutoCastVariable '{v.name}' shape={v.shape} "
-                "dtype={v.dtype.name} dtype_to_cast_to={v._cast_dtype.name}, "
+                "dtype={v.true_dtype.name} "
+                "dtype_to_cast_to={v._cast_dtype.name}, "
                 "numpy={np_repr}>"
             )
             return repr_str.format(
@@ -211,7 +213,8 @@ class AutoCastVariable(tf.Variable, tf.__internal__.types.Tensor):
         else:
             repr_str = (
                 "<AutoCastVariable '{v.name}' shape={v.shape} "
-                "dtype={v.dtype.name} dtype_to_cast_to={v._cast_dtype.name}>"
+                "dtype={v.true_dtype.name} "
+                "dtype_to_cast_to={v._cast_dtype.name}>"
             )
             return repr_str.format(v=self)
 
@@ -261,6 +264,9 @@ class AutoCastVariable(tf.Variable, tf.__internal__.types.Tensor):
     def _apply_assign_update(
         self, update_fn, value, use_locking=None, name=None, read_value=True
     ):
+        # In auto cast scope, we cast back to the actual variable dtype.
+        if self._should_cast():
+            value = tf.cast(value, self.true_dtype)
         # TODO(b/146181571): This logic can be simplified once
         # DistributedVariable.assign returns a DistributedVariable. Currently
         # for MirroredStrategy, it returns a Mirrored value.

tf_keras/src/mixed_precision/test_util.py

@@ -171,14 +171,14 @@ class MultiplyLayer(AssertTypeLayer):
             activity_regularizer=self._activity_regularizer, **kwargs
         )
 
-    def build(self, _):
+    def build(self, input_shape):
         self.v = self.add_weight(
             self._var_name,
             (),
             initializer="ones",
             regularizer=self._regularizer,
         )
-        self.built = True
+        super().build(input_shape)
 
     def call(self, inputs):
         self.assert_input_types(inputs)
@@ -205,7 +205,7 @@ class MultiplyLayer(AssertTypeLayer):
 class MultiplyLayerWithoutAutoCast(MultiplyLayer):
     """Same as MultiplyLayer, but does not use AutoCastVariables."""
 
-    def build(self, _):
+    def build(self, input_shape):
         dtype = self.dtype
         if dtype in ("float16", "bfloat16"):
             dtype = "float32"
@@ -214,10 +214,11 @@ class MultiplyLayerWithoutAutoCast(MultiplyLayer):
             (),
             initializer="ones",
             dtype=dtype,
-            experimental_autocast=False,
+            autocast=False,
             regularizer=self._regularizer,
         )
-        self.built = True
+        # Call Layer.build() to skip MultiplyLayer.build() which we override.
+        base_layer.Layer.build(self, input_shape)
 
     def call(self, inputs):
         self.assert_input_types(inputs)

tf_keras/src/optimizers/legacy/optimizer_v2.py

@@ -1033,6 +1033,13 @@ class OptimizerV2(tf.__internal__.tracking.Trackable):
         slot_dict = self._slots.setdefault(var_key, {})
         weight = slot_dict.get(slot_name, None)
         if weight is None:
+            # Under a mixed precision policy, variables report their "cast"
+            # dtype. However, we want to use the original dtype for slots.
+            if hasattr(var, "true_dtype"):
+                dtype = var.true_dtype
+            else:
+                dtype = var.dtype
+
             if isinstance(initializer, str) or callable(initializer):
                 initializer = initializers.get(initializer)
                 if isinstance(
@@ -1043,7 +1050,7 @@ class OptimizerV2(tf.__internal__.tracking.Trackable):
                 else:
                     slot_shape = var.shape
                 initial_value = functools.partial(
-                    initializer, shape=slot_shape, dtype=var.dtype
+                    initializer, shape=slot_shape, dtype=dtype
                 )
             else:
                 initial_value = initializer
@@ -1064,7 +1071,7 @@ class OptimizerV2(tf.__internal__.tracking.Trackable):
                 with strategy.extended.colocate_vars_with(var):
                     weight = tf.Variable(
                         name=f"{var._shared_name}/{slot_name}",
-                        dtype=var.dtype,
+                        dtype=dtype,
                         trainable=False,
                         initial_value=initial_value,
                     )

tf_keras/src/optimizers/optimizer.py

@@ -498,26 +498,28 @@ class _BaseOptimizer(tf.__internal__.tracking.AutoTrackable):
         Returns:
           An optimizer variable.
         """
+        # Under a mixed precision policy, variables report their "cast"
+        # dtype. However, we want to use the original dtype for slots.
+        if hasattr(model_variable, "true_dtype"):
+            dtype = model_variable.true_dtype
+        else:
+            dtype = model_variable.dtype
         if initial_value is None:
             if shape is None:
                 if model_variable.shape.rank is None:
                     # When the rank is None, we cannot get a concrete
                     # `model_variable.shape`, we use dynamic shape.
-                    initial_value = tf.zeros_like(
-                        model_variable, dtype=model_variable.dtype
-                    )
+                    initial_value = tf.zeros_like(model_variable, dtype=dtype)
                 else:
                     # We cannot always use `zeros_like`, because some cases
                     # the shape exists while values don't.
-                    initial_value = tf.zeros(
-                        model_variable.shape, dtype=model_variable.dtype
-                    )
+                    initial_value = tf.zeros(model_variable.shape, dtype=dtype)
             else:
-                initial_value = tf.zeros(shape, dtype=model_variable.dtype)
+                initial_value = tf.zeros(shape, dtype=dtype)
         variable = tf.Variable(
             initial_value=initial_value,
             name=f"{variable_name}/{model_variable._shared_name}",
-            dtype=model_variable.dtype,
+            dtype=dtype,
             trainable=False,
         )
         # If model_variable is a shard of a ShardedVariable, we should add a
@@ -1188,10 +1190,17 @@ class Optimizer(_BaseOptimizer):
                         self._mesh, rank=initial_value.shape.rank
                     ),
                 )
+            # Under a mixed precision policy, variables report their "cast"
+            # dtype. However, we want to use the original dtype for optimizer
+            # variables.
+            if hasattr(model_variable, "true_dtype"):
+                dtype = model_variable.true_dtype
+            else:
+                dtype = model_variable.dtype
             variable = tf.experimental.dtensor.DVariable(
                 initial_value=initial_value,
                 name=f"{variable_name}/{model_variable._shared_name}",
-                dtype=model_variable.dtype,
+                dtype=dtype,
                 trainable=False,
             )
             self._variables.append(variable)

tf_keras/src/premade_models/linear.py

@@ -156,7 +156,8 @@ class LinearModel(training.Model):
             )
         else:
             self.bias = None
-        self.built = True
+        # Call Layer.build() to skip Model.build() which we override here.
+        base_layer.Layer.build(self, input_shape)
 
     def call(self, inputs):
         result = None

tf_keras/src/utils/data_utils.py

@@ -1108,7 +1108,7 @@ def pad_sequences(
         maxlen = np.max(lengths)
 
     is_dtype_str = np.issubdtype(dtype, np.str_) or np.issubdtype(
-        dtype, np.unicode_
+        dtype, np.str_
     )
     if isinstance(value, str) and dtype != object and not is_dtype_str:
         raise ValueError(

tf_keras/src/utils/steps_per_execution_tuning.py

@@ -229,7 +229,7 @@ class StepsPerExecutionTuner:
 
         if current_spe >= spe_limit:
             new_spe = current_spe
-        elif current_spe == 0:
+        elif current_spe <= 0:
             new_spe = self.init_spe
 
         self._steps_per_execution.assign(np.round(new_spe))

tf_keras/src/utils/timeseries_dataset.py

@@ -110,16 +110,24 @@ def timeseries_dataset_from_array(
     timesteps to predict the next timestep, you would use:
 
     ```python
-    input_data = data[:-10]
-    targets = data[10:]
+    data = tf.range(15)
+    sequence_length = 10
+    input_data = data[:]
+    targets = data[sequence_length:]
     dataset = tf.keras.utils.timeseries_dataset_from_array(
-        input_data, targets, sequence_length=10)
+        input_data, targets, sequence_length=sequence_length
+    )
     for batch in dataset:
       inputs, targets = batch
-      assert np.array_equal(inputs[0], data[:10])  # First sequence: steps [0-9]
+      # First sequence: steps [0-9]
+      assert np.array_equal(inputs[0], data[:sequence_length])
       # Corresponding target: step 10
-      assert np.array_equal(targets[0], data[10])
+      assert np.array_equal(targets[0], data[sequence_length])
       break
+    # To view the generated dataset
+    for batch in dataset.as_numpy_iterator():
+      input, label = batch
+      print(f"Input:{input}, target:{label}")
     ```
 
     Example 3: Temporal regression for many-to-many architectures.

{tf_keras_nightly-2.17.0.dev2024050509.dist-info → tf_keras_nightly-2.19.0.dev2024101709.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: tf_keras-nightly
-Version: 2.17.0.dev2024050509
+Version: 2.19.0.dev2024101709
 Summary: Deep learning for humans.
 Home-page: https://keras.io/
 Download-URL: https://github.com/keras-team/tf-keras/tags
@@ -26,7 +26,7 @@ Classifier: Topic :: Software Development
 Classifier: Topic :: Software Development :: Libraries
 Classifier: Topic :: Software Development :: Libraries :: Python Modules
 Requires-Python: >=3.9
-Requires-Dist: tf-nightly ~=2.17.0.dev
+Requires-Dist: tf-nightly ~=2.19.0.dev
 
 TF-Keras is a deep learning API written in Python,
 running on top of the machine learning platform TensorFlow.