keras-team/keras

Description Three public objects are missing from the ALL_OBJECTS registry of their module: keras.losses.CategoricalGeneralizedCrossEntropy and keras.losses.categorical_generalized_cross_entropy. keras/src/losses/__init__.py does not import them. keras.optimizers.Lamb. keras/src/optimizers/__init__.py does not import it. keras.losses.get, keras.losses.deserialize, keras.optimizers.get and keras.optimizers.deserialize find objects only through ALL_OBJECTS_DICT. Thus, for these three objects: keras.losses.get and keras.optimizers.get raise ValueError: Could not interpret ... identifier for the class name, for the function name and for a serialized config. Model.compile uses get for string identifiers. compile(optimizer="lamb") raises this ValueError. compile(loss="CategoricalGeneralizedCrossEntropy") and compile(loss="categorical_generalized_cross_entropy") raise it at the first fit or evaluate. The other losses and optimizers accept their names, for example compile(loss="categorical_

Doc Issue Same as #21562, several docs of mobilenet should be fixed: In mobilenet.py: (the first one is described properly in comments) keras/keras/src/applications/mobilenet.py Lines 119 to 124 in 58db6d6 if weights == "imagenet" and include_top and classes != 1000: raise ValueError( "If using `weights='imagenet'` with `include_top=True`, " "`classes` should be 1000. " f"Received classes={classes}" ) keras/keras/src/applications/mobilenet.py Lines 158 to 164 in 58db6d6 if weights == "imagenet": if d

When calling keras.ops.multi_hot on a symbolic KerasTensor (e.g., dtype="int32" or dtype="int64"), MultiHot.compute_output_spec returns a KerasTensor with inputs.dtype ("int32"/"int64") instead of self.dtype (backend.floatx() by default, or the user-specified dtype). In eager mode, keras.ops.multi_hot returns "float32" by default (or the explicitly requested dtype), causing a dtype mismatch between symbolic graph construction and eager execution. Root Cause In keras/src/ops/nn.py: MultiHot.__init__ initializes self.dtype = dtype or backend.floatx(). However, MultiHot.compute_output_spec returns: return KerasTensor(x_shape, dtype=inputs.dtype, sparse=self.sparse) Using dtype=inputs.dtype instead of dtype=self.dtype causes symbolic execution to always inherit the integer index dtype of inputs and ignore both the default floatx() dtype and any explicit dtype argument. Additionally, MultiHot.call omits sparse=self.sparse when calling backend.nn.multi_hot. Reproduction colab: https://col

Calling keras.ops.batch_normalization on symbolic KerasTensor inputs when offset is provided and scale=None (the default) crashes with AttributeError: 'NoneType' object has no attribute 'shape'. Root Cause In BatchNorm.compute_output_spec in keras/src/ops/nn.py, line 2430 has a typo: if offset is not None: self._check_shape("offset", tuple(offset.shape), shape) if offset is not scale: # <-- Typo for `if scale is not None:` self._check_shape("scale", tuple(scale.shape), shape) When offset is a tensor (not None) and scale=None, offset is not scale evaluates to True, and tuple(scale.shape) fails with AttributeError: 'NoneType' object has no attribute 'shape'. Conversely, if the same tensor instance is passed for both offset and scale, offset is not scale evaluates to False and skips validating scale's shape. Reproduction Colab: https://colab.research.google.com/drive/1eilbjHoAgXb-sMUehi8XwqtVN7sxeAyF?usp=sharing

When calling keras.ops.categorical_crossentropy on symbolic KerasTensor inputs with a non-default axis (e.g., axis=1 or axis=-2), CategoricalCrossentropy.compute_output_spec ignores self.axis and always strips the last dimension (output.shape[:-1]). As a result, symbolic execution returns the wrong output shape whenever axis != -1 and fails to raise a ValueError when axis is out of bounds. Root Cause In CategoricalCrossentropy.compute_output_spec in keras/src/ops/nn.py (line 2085): def compute_output_spec(self, target, output): ... return KerasTensor(output.shape[:-1], dtype=output.dtype) Unlike SparseCategoricalCrossentropy.compute_output_spec immediately below it (which canonicalizes self.axis via canonicalize_axis(self.axis, len(output.shape)) and returns output.shape[:axis] + output.shape[axis + 1 :]), CategoricalCrossentropy.compute_output_spec hardcodes output.shape[:-1]. Reproduction colab: https://colab.research.google.com/drive/1ieqLsG4DEnXpgTyLxQTA_k-4hERBdXBC?usp=sha