Repository navigation

Issues

New issue

🗓️ Keras Community Meetings: Schedule, Invites, and Notes
#22410 · laxmareddyp opened on Mar 12, 2026
🚀 Contributing to Keras: Open Projects, Ideas and Roadmap 🚀
#18442 · fchollet opened on Jul 14, 2023
56
[RFC] New Pull Request Policy 🕵️
#23601 · divyashreepathihalli opened on Sep 10, 2026

Add tests for op edge cases that hid bugs in the MLX backend
keras-team/keras#24103 · opened on Oct 30, 2026
Distiller runs the student two times in each training step with named feature layers
keras-team/keras#24101 · opened on Oct 30, 2026
2
R2Score loses precision on targets far from zero and can return a score above 1
keras-team/keras#24099 · opened on Oct 29, 2026
1
Unpin torch in requirements-torch-cuda.txt
keras-team/keras#24096 · opened on Oct 29, 2026
LossScaleOptimizer misses overflow on non first replicas under MirroredStrategy and corrupts weights
keras-team/keras#24090 · opened on Oct 28, 2026
4
Torch graph recompiles at every _clear_losses if add_loss is ever used.
keras-team/keras#24084 · opened on Oct 27, 2026
3
Unnecessary graph break in batchnorm implementation for torch dynamo
keras-team/keras#24082 · opened on Oct 27, 2026
3
keras.ops.roll rejects tensor shift/axis on the TensorFlow backend
keras-team/keras#24080 · opened on Oct 27, 2026
3
keras.ops.unique with a tensor size only works for int32 on the TensorFlow backend
keras-team/keras#24079 · opened on Oct 27, 2026
4
pad_sequences raises TypeError inside tf.function with dynamic-shape sequence tensors
keras-team/keras#24078 · opened on Oct 27, 2026
3
Distiller.fit() executes student model twice per step, causing duplicate BatchNorm updates and mismatched Dropout masks
keras-team/keras#24074 · opened on Oct 26, 2026
6
Distiller evaluation mutates student BatchNormalization state by running feature extraction with training=True
keras-team/keras#24068 · opened on Oct 25, 2026
7
RandomShear with float factors cannot be loaded from its own config
keras-team/keras#24067 · opened on Oct 25, 2026
1
Torch backend: ops.take and ops.take_along_axis with axis=None mishandle negative indices for inputs with more than one dimension
keras-team/keras#24066 · opened on Oct 25, 2026
1
Hashing with output_mode="one_hot" gives the wrong output shape when the last axis isn't 1
keras-team/keras#24064 · opened on Oct 24, 2026
ops.in_top_k and TopKCategoricalAccuracy crash on torch/numpy when k > number of classes
keras-team/keras#24051 · opened on Oct 22, 2026
2
[Feature Request] Add native DoRA (Weight-Decomposed Low-Rank Adaptation) support to core Keras layers
keras-team/keras#24049 · opened on Oct 22, 2026
keras.ops.sinc returns NaN gradient at zero on the TensorFlow backend
keras-team/keras#24043 · opened on Oct 22, 2026
3
Add an x64_mode setting for 64-bit dtype support (off / explicit_x64 / enable_x64)
Feature
keras-team/keras#24040 · opened on Oct 21, 2026
1
[Feature / Bug]: Missing @keras_export for functional cosine_similarity metric and docstring discrepancy
keras-team/keras#24019 · opened on Oct 20, 2026
keras.ops.image.resize no longer rejects a non-positive size given as NumPy integers
keras-team/keras#24007 · opened on Oct 20, 2026
1
ops.one_hot raises for indices >= num_classes on the torch backend but returns zeros on TensorFlow
keras-team/keras#24006 · opened on Oct 20, 2026
1
ops.arange raises on an empty range on TensorFlow and torch
keras-team/keras#24005 · opened on Oct 20, 2026
3
ops.split fails with negative or out-of-range indices_or_sections on TensorFlow and torch
keras-team/keras#24004 · opened on Oct 20, 2026
3
ops.diagonal, ops.diag, ops.trace and ops.eye raise for out-of-range offsets on the TensorFlow backend
keras-team/keras#24003 · opened on Oct 20, 2026
1

Navigation Menu

Repository navigation

Issues

Issues

Add tests for op edge cases that hid bugs in the MLX backend

`Distiller` runs the student two times in each training step with named feature layers

R2Score loses precision on targets far from zero and can return a score above 1

Unpin torch in requirements-torch-cuda.txt

LossScaleOptimizer misses overflow on non first replicas under MirroredStrategy and corrupts weights

Torch graph recompiles at every _clear_losses if add_loss is ever used.

Unnecessary graph break in batchnorm implementation for torch dynamo

`keras.ops.roll` rejects tensor `shift`/`axis` on the TensorFlow backend

`keras.ops.unique` with a tensor `size` only works for int32 on the TensorFlow backend

pad_sequences raises TypeError inside tf.function with dynamic-shape sequence tensors

`Distiller.fit()` executes student model twice per step, causing duplicate BatchNorm updates and mismatched Dropout masks

`Distiller` evaluation mutates student BatchNormalization state by running feature extraction with `training=True`

RandomShear with float factors cannot be loaded from its own config

Torch backend: `ops.take` and `ops.take_along_axis` with `axis=None` mishandle negative indices for inputs with more than one dimension

Hashing with output_mode="one_hot" gives the wrong output shape when the last axis isn't 1

ops.in_top_k and TopKCategoricalAccuracy crash on torch/numpy when k > number of classes

[Feature Request] Add native DoRA (Weight-Decomposed Low-Rank Adaptation) support to core Keras layers

keras.ops.sinc returns NaN gradient at zero on the TensorFlow backend

Add an `x64_mode` setting for 64-bit dtype support (`off` / `explicit_x64` / `enable_x64`)

[Feature / Bug]: Missing @keras_export for functional cosine_similarity metric and docstring discrepancy

`keras.ops.image.resize` no longer rejects a non-positive `size` given as NumPy integers

ops.one_hot raises for indices >= num_classes on the torch backend but returns zeros on TensorFlow

ops.arange raises on an empty range on TensorFlow and torch

ops.split fails with negative or out-of-range indices_or_sections on TensorFlow and torch

ops.diagonal, ops.diag, ops.trace and ops.eye raise for out-of-range offsets on the TensorFlow backend