13 Pull requests merged by 7 people

• Rollback tf distribute change

  #20017 merged Jul 19, 2024

• Simplify and improve CompileLoss

  #20015 merged Jul 19, 2024

• Revert to the original default behavior of dtype in Loss and Metric

  #20014 merged Jul 19, 2024

• Remove duplicated code from rnn.py

  #20013 merged Jul 19, 2024

• Avoid exceptions in save_model and load_model

  #19989 merged Jul 19, 2024

• Fix conv + channel_first + tensorflow + CPU

  #20008 merged Jul 19, 2024

• Fix gradient accumulation

  #19998 merged Jul 18, 2024

• Correct test case for ImdbLoadDataTest

  #20009 merged Jul 18, 2024

• Use Numpy arrays in data adapters for JAX backend.

  #20007 merged Jul 18, 2024

• Extend support to Sequential._maybe_build for functional/sequential models

  #20002 merged Jul 17, 2024

• Change dtypes for test_dtypes test case.

  #19987 merged Jul 17, 2024

• Actually sum gradient for tf distribute

  #19969 merged Jul 15, 2024

• add lamb optimizer

  #19981 merged Jul 14, 2024

5 Pull requests opened by 4 people

• Add test branch for layers with sublayers

  #20010 opened Jul 19, 2024

• Fixing issue #19754: len is not well defined for a symbolic Tensor

  #20016 opened Jul 19, 2024

• Faster host to device transfer on Jax backend

  #20018 opened Jul 20, 2024

• Fixing issue #20021: Unexpected result on keras.ops.nn.softmax when running with tuple axis

  #20022 opened Jul 20, 2024

• Bring back loss information for multiple outputs

  #20023 opened Jul 21, 2024

14 Issues closed by 10 people

• having error in executing object detection module

  #19888 closed Jul 20, 2024

• loss/losses are omitted from get_config/serialization

  #19884 closed Jul 19, 2024

• model.save fail after updating to keras-nightly

  #19985 closed Jul 19, 2024

• predict_on_batch fails when the model was compiled with steps_per_execution > 1

  #19996 closed Jul 19, 2024

• Regression bug when using 3D convolution with channels_first on GPU

  #19991 closed Jul 19, 2024

• PyDataset ignores __len__ during iteration, Tensorflow backend

  #20001 closed Jul 18, 2024

• (feat) Implement the deprecated `ctc_decode` function in Keras v3

  #19227 closed Jul 18, 2024

• `padding_mode` in convolutions

  #19988 closed Jul 18, 2024

• Gradient accumulation done wrong

  #19990 closed Jul 18, 2024

• Keras 3.4.2 release

  #19976 closed Jul 18, 2024

• SparseCategoricalCrossentropy consumes heavy GPU RAM on over 10K categories.

  #20005 closed Jul 18, 2024

• "ValueError: Undefined shapes are not supported." when calling model.call()

  #19982 closed Jul 17, 2024

• Irregular Tensors as output from Generator class with batch size = 1

  #19925 closed Jul 17, 2024

• LAMB optimizer for keras3

  #19965 closed Jul 15, 2024

14 Issues opened by 11 people

• Unexpected result on keras.ops.nn.softmax when running with tuple axis

  #20021 opened Jul 20, 2024

• Unexpected saving/loading `.h5` weights behavior on Keras 3+.

  #20019 opened Jul 20, 2024

• Sum Metrics reset_state only supports float types

  #20012 opened Jul 19, 2024

• Zero encountered in variance in NormalizationLayer leads to backend.epsilon() being used which is a questionable choice.

  #20011 opened Jul 19, 2024

• InvalidArgumentError: Graph execution error:

  #20006 opened Jul 17, 2024

• Custom layers with default layers composition hide their weights

  #20004 opened Jul 17, 2024

• Any plans to restore fused layer normalization

  #20003 opened Jul 17, 2024

• Intended behavior of `TorchModuleWrapper`?

  #20000 opened Jul 16, 2024

• Saving and restoring the model is changing type of model inputs and outputs

  #19999 opened Jul 16, 2024

• Assign doesn't return a value

  #19997 opened Jul 16, 2024

• Incorrect handling of `nan` in `in_top_k` of Jax backend (regression from keras 3.3.3)

  #19995 opened Jul 15, 2024

• keras.utils.PyDataset / tf.keras.utils.Sequence ignoring __len__ different behavior Keras2/Keras3 (tensorflow 2.16)

  #19994 opened Jul 15, 2024

• Accuracy Drop Across TensorFlow Versions When Using Keras 3 Instead of Keras 2

  #19993 opened Jul 15, 2024

• Optimizers in Keras V3 are slower than in V2 when working with sparse gradients

  #19992 opened Jul 15, 2024

13 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• Simplify the implementation of quantization-related methods

  #19954 commented on Jul 20, 2024 • 5 new comments

• first version of flash_attention for jax

  #19743 commented on Jul 20, 2024 • 2 new comments

• [BUG][all_reduce] INVALID_ARGUMENT: You must feed a value for placeholder tensor

  #19246 commented on Jul 16, 2024 • 0 new comments

• Unexpected depreciation warning in tensorflow 2.16.1

  #19811 commented on Jul 16, 2024 • 0 new comments

• Backward compatibility issues: arbitrary inputs and overreliance on custom layers

  #19314 commented on Jul 17, 2024 • 0 new comments

• AttributeError: 'types.SimpleNamespace' object has no attribute 'get_registered_object' from return keras.saving.get_registered_object(config["registered_name"]) in keras_nlp\src\utils\preset_utils.py

  #19456 commented on Jul 17, 2024 • 0 new comments

• Think about Keras3 preference, should I develop my distributed training CTR model based on Jax or TensorFlow?

  #19866 commented on Jul 18, 2024 • 0 new comments

• Load a model saved using keras 2.3.1 in keras 3 (containing bidirectionnal LSTM)

  #19898 commented on Jul 19, 2024 • 0 new comments

• [Feature Request] Add cuDNN-accelerated LSTM and GRU to PyTorch

  #19177 commented on Jul 19, 2024 • 0 new comments

• Sequentials `_maybe_rebuild` does not make sense

  #19916 commented on Jul 19, 2024 • 0 new comments

• Support for channels first data format with zeropadding1D

  #19984 commented on Jul 19, 2024 • 0 new comments

• Memory leak when using custom DataGenerator

  #19907 commented on Jul 19, 2024 • 0 new comments

• utility for video reader and manipulation

  #19055 commented on Jul 21, 2024 • 0 new comments