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

layer benchmarks crash on keras 3
keras-team/keras#23862 · opened on Oct 7, 2026
1
circle loss crashes with a dynamic-batch symbolic input, root cause in ops.ndim
keras-team/keras#23861 · opened on Oct 7, 2026
Rebuilding Keras op tests for scale and coverage
Feature
keras-team/keras#23855 · opened on Oct 6, 2026
sparse_sigmoid returns wrong values for integer and bool input
keras-team/keras#23852 · opened on Oct 6, 2026
1
CategoricalGeneralizedCrossEntropy and Lamb are missing from ALL_OBJECTS, so get and deserialize do not find them
keras-team/keras#23849 · opened on Oct 6, 2026
Loss.get_config does not include dtype, so a restored loss computes in floatx
keras-team/keras#23848 · opened on Oct 6, 2026
1
erf returns wrong values for integer input on the numpy and openvino backends
keras-team/keras#23846 · opened on Oct 6, 2026
1
TF backend: image.resize(pad_to_aspect_ratio=True) raises TypeError on integer images with the default fill_value
keras-team/keras#23832 · opened on Oct 3, 2026
1
from_config skips the vocabulary-file safe-mode check when no scope is active
keras-team/keras#23817 · opened on Oct 2, 2026
keras.ops.hypot returns incorrect second derivative on TensorFlow backend when one component is zero
keras-team/keras#23816 · opened on Oct 1, 2026
Security: crafted .npz can bypass RestrictedUnpickler array-state validation in npz_utils (OOB read/write)
keras-team/keras#23805 · opened on Sep 30, 2026
2
GlobalAveragePooling1D emits NaNs for fully masked sequences
keras-team/keras#23797 · opened on Sep 30, 2026
4
Clarify unknown-rank dataset errors with steps_per_execution greater than one
keras-team/keras#23793 · opened on Sep 29, 2026
5
keras.ops.cross fails for 2-D vectors on the torch backend
keras-team/keras#23789 · opened on Sep 29, 2026
1
boston_housing_test.py test fail. Expected AssertionError to pass, but received ValueError
keras-team/keras#23787 · opened on Sep 29, 2026
1
LayerNormalization returns 0 instead of beta for large constant rows
keras-team/keras#23780 · opened on Sep 29, 2026
1
Should enable fullgraph=True for torch.compile?
keras-team/keras#23775 · opened on Sep 28, 2026
7
RandomRotation: boxes are rotated about a different center than the image (half-pixel offset)
keras-team/keras#23773 · opened on Sep 26, 2026
Resizing with crop_to_aspect_ratio / pad_to_aspect_ratio: boxes use real-valued crop/pad geometry while the image uses integer-rounded geometry
keras-team/keras#23772 · opened on Sep 26, 2026
3
RandomShear: boxes are sheared in normalized coordinates while the image is sheared in pixel coordinates
keras-team/keras#23771 · opened on Sep 26, 2026
Fix several Doc issues for mobilenet, mobilenet_v2, and mobilenet_v3
keras-team/keras#23764 · opened on Sep 25, 2026
2
Layer.__delattr__ does not untrack sub-layers
keras-team/keras#23761 · opened on Sep 25, 2026
1
Backend kernels do not follow the op spec: ignored dtype arguments, rejected index dtypes, wrong output structure
keras-team/keras#23757 · opened on Sep 24, 2026
OpenVINO backend: ops return wrong values or fail (scalar operands in the wrong element type, three CPU-plugin bugs, shape bugs)
keras-team/keras#23756 · opened on Sep 24, 2026
Backend ops do not canonicalize list and tuple arguments or convert Python inputs, so Functional models fail after a .keras reload
keras-team/keras#23755 · opened on Sep 24, 2026
1

Navigation Menu

Repository navigation

Issues

Issues

layer benchmarks crash on keras 3

circle loss crashes with a dynamic-batch symbolic input, root cause in ops.ndim

Rebuilding Keras op tests for scale and coverage

sparse_sigmoid returns wrong values for integer and bool input

`CategoricalGeneralizedCrossEntropy` and `Lamb` are missing from `ALL_OBJECTS`, so `get` and `deserialize` do not find them

`Loss.get_config` does not include `dtype`, so a restored loss computes in `floatx`

erf returns wrong values for integer input on the numpy and openvino backends

TF backend: image.resize(pad_to_aspect_ratio=True) raises TypeError on integer images with the default fill_value

from_config skips the vocabulary-file safe-mode check when no scope is active

`keras.ops.hypot` returns incorrect second derivative on TensorFlow backend when one component is zero

Security: crafted .npz can bypass RestrictedUnpickler array-state validation in npz_utils (OOB read/write)

GlobalAveragePooling1D emits NaNs for fully masked sequences

Clarify unknown-rank dataset errors with steps_per_execution greater than one

keras.ops.cross fails for 2-D vectors on the torch backend

boston_housing_test.py test fail. Expected AssertionError to pass, but received ValueError

LayerNormalization returns 0 instead of beta for large constant rows

Should enable `fullgraph=True` for `torch.compile`?

RandomRotation: boxes are rotated about a different center than the image (half-pixel offset)

Resizing with crop_to_aspect_ratio / pad_to_aspect_ratio: boxes use real-valued crop/pad geometry while the image uses integer-rounded geometry

RandomShear: boxes are sheared in normalized coordinates while the image is sheared in pixel coordinates

Fix several Doc issues for `mobilenet`, `mobilenet_v2`, and `mobilenet_v3`

Layer.delattr does not untrack sub-layers

Backend kernels do not follow the op spec: ignored `dtype` arguments, rejected index dtypes, wrong output structure

OpenVINO backend: ops return wrong values or fail (scalar operands in the wrong element type, three CPU-plugin bugs, shape bugs)

Backend ops do not canonicalize list and tuple arguments or convert Python inputs, so Functional models fail after a `.keras` reload