Keras v3.15.0

v3.15.0
Added 9
  • Added `export_torch` function to export Keras models to native PyTorch `nn.Module` format
  • Added `sliding_window` parameter to `MultiHeadAttention` and `GroupedQueryAttention` for efficient long-context attention
  • Added `unique`, `fabs`, `fmax`, `fmin`, `dsplit`, `erfc`, `percentile`, and `nanpercentile` operations to `keras.ops.numpy`
  • Added `pinv` (pseudo-inverse) and `matrix_rank` operations to `keras.ops.linalg`
  • Added `sobel_edges` and `ssim` (structural similarity) operations to `keras.ops.image`
  • Added `MultiOptimizer` to support training sub-networks with different optimizers
Changed 5
  • Causal-only attention in `MultiHeadAttention` and `GroupedQueryAttention` now automatically dispatches to Flash Attention (cuDNN SDPA)
  • `keras.ops.transpose` now supports negative axis values
  • JAX backend now fuses Bidirectional LSTM into a single cuDNN call
  • Added fused bidirectional GRU for Torch backend
  • Added distribution information (`num_processes`, `num_model_replicas`, `data_shard_id`) to `ModelParallel`
Fixed 4
  • Fixed manual attention path to correctly apply causal masking
  • Fixed LiteRT export for Keras 3 with TensorFlow 2.20 and Python 3.13
  • Fixed float16 numerical instability in `GroupNormalization` with small epsilon
  • Fixed GPU OOM with rank-3 input in Dense layer due to `BatchMatMulV2` gradient materialization
Security 2
  • Hardened model reloading against HDF5 exploits by rejecting `ExternalLink`, `SoftLink` groups, virtual datasets, and shape-bomb datasets
  • Hardened archive loading by rejecting tar members and links escaping extraction directory and ZIP/NPZ members declaring excessive data
Highlights
  • Keras-to-Torch Export: New export_torch enables exporting Keras models to native PyTorch nn.Module format, along with LiteRT (TFLite) export support for the PyTorch backend.
  • Sliding Window Attention: Added sliding_window parameter to MultiHeadAttention and GroupedQueryAttention for efficient long-context attention.
  • Flash / Fused SDPA: Causal-only MHA/GQA now automatically dispatches to Flash Attention (cuDNN SDPA), and the manual attention path correctly applies causal masking.
  • Multi-Optimizer Training: New MultiOptimizer supports assigning different optimizers to sub-networks.
  • New Math Operations: Added unique, pinv, matrix_rank, fabs, fmax, fmin, erfc, dsplit, percentile, nanpercentile, sobel_edges, and ssim (structural similarity) to keras.ops.
  • Security Hardening: Comprehensive hardening of model reloading against HDF5 exploits, tar/zip traversal attacks, insecure deserialization.
New Features and Operations
Multi-Backend Operations
  • New NumPy Operations: Added unique, fabs, fmax, fmin, dsplit, erfc, percentile, nanpercentile in keras.ops.numpy.
  • New Linear Algebra Operations: Added pinv (pseudo-inverse) and matrix_rank in keras.ops.linalg.
  • New Image Operations: Added sobel_edges for edge detection and ssim (structural similarity) in keras.ops.image.
  • Negative Axes in Transpose: keras.ops.transpose now supports negative axis values.
Layers and Attention
  • Sliding Window Attention: MultiHeadAttention and GroupedQueryAttention layers support the sliding_window parameter for efficient long-sequence processing.
  • Flash Attention Engagement: Causal-only attention in MHA/GQA now uses Flash SDPA for significant speedups.
  • Fused Bidirectional LSTM/GRU: JAX backend now fuses Bidirectional LSTM into a single cuDNN call; fused bidirectional GRU added for Torch backend.
  • CTC Beam Search Decoder: Added CTC beam search decoding for the Torch backend.
Training and Optimizers
  • MultiOptimizer: Supports training sub-networks with different optimizers.
  • SKLearn Classifier: Added predict_proba method to SKLearnClassifier.
Export and Deployment
  • Keras-to-Torch Export: Export Keras models to native PyTorch nn.Module via model.export(..., format="torch").
  • LiteRT (TFLite) Export for PyTorch: Added LiteRT export support for models using the PyTorch backend.
  • LiteRT Compatibility Fix: Fixed LiteRT export for Keras 3 + TF 2.20 + Python 3.13.
  • ONNX Export: Support for dict/list inputs in Torch ONNX export; documented static input signature requirement for LiteRT PyTorch export.
Distribution and Parallelism
  • ModelParallel Improvements: Defined contiguous replica-group data shard ID convention; added distribution information (num_processes, num_model_replicas, data_shard_id).
  • Initializer Distribution Layout: Initializers can now handle the distribution layout directly with JAX.
  • TF Dataset Distribution: Refactored TF dataset distribution with centralized sharding routing; fixed data distribution for model training in JAX.
OpenVINO Backend Support

The OpenVINO backend received continued improvements:

  • New Operations: Implemented glu, sparsemax, gaussian_blur, logdet, cholesky, lu_factor, erfc, segment_min, segment_prod, percentile, nanmedian, nanpercentile, unique, flash_attn, greedy ctc_decode, solve_triangular, compute_homography_matrix, and image transforms (affine, perspective, elastic).
  • Opset Upgrades: Upgraded to opset16 for select operations and full upgrade.
  • Fixes: Dynamic/symbolic shape handling, mask propagation, random seed determinism, dropout during predict, Lanczos interpolation in resize, dynamic batch shape propagation, and improved efficiency using native ops.
Security
  • HDF5 Hardening: Reject ExternalLink/SoftLink groups, virtual datasets, and shape-bomb datasets in model loading.
  • Archive Hardening: Reject tar members and links escaping extraction directory, ZIP/NPZ members declaring excessive data. Validate asset paths from Orbax checkpoints.
  • Deserialization Safety: Disable pickle in np.load, fix insecure deserialization in dataset utilities, and make Lambda/TorchModuleWrapper from_config fail closed when safe_mode is unset.
  • CI/Workflow: Fix prompt injection in issue triage workflow.
Bug Fixes and Improvements
Backend Specific Improvements
  • PyTorch: Fixed convert_to_tensor for Python scalars, divide_no_nan() NaN gradients, BiLSTM dispatch, lstsq with rcond, SymInt/SymFloat handling in convert_to_tensor and slice, and median for even-length inputs.
  • JAX: Fused Bidirectional LSTM into cuDNN call.
  • TensorFlow: Fixed depthwise/separable conv with stride and dilation. Optimized tf.tensordot by removing redundant float casts.
Layers and Ops
  • Mixed Precision Fix: Fixed float16 numerical instability in GroupNormalization with small epsilon; disabled autocast for mixed precision stability.
  • Dense Layer OOM: Fixed GPU OOM with rank-3 input due to BatchMatMulV2 gradient materialization.
  • GroupQueryAttention: Fixed symbolic output shape with return_attention_scores.
  • Conv Transpose: Save output_padding in Conv1D/2D/3DTranspose get_config.
  • Attention Layer: Fixed stale return_attention_scores flag in compute_output_spec; save seed in get_config.
  • Ops Validation: Added comprehensive axis validation in softmax, normalize, swapaxes, moveaxis, sort, argsort, cumsum, cumprod, take, stack, concatenate, split, diff, transpose, and more.
  • EinsumDense: Fixed compute_output_shape to work before build.
  • Discretization: Fixed bin boundaries calculation.
Model Saving and Loading
  • Nested Sublayers: Fixed save/load for custom models/layers with sublayers in nested lists.
  • Orbax: Fixed bug from Orbax's recent rename from "pytree" to "state".
  • Sequential: Improved error handling for missing keys during deserialization.
  • Pipeline: Validated from_config layers and avoid mutating input config.
Other Improvements
  • Callbacks: Fixed EarlyStopping/ReduceLROnPlateau resetting self.best between fit calls; fixed TensorBoard callback step counter never updating.
  • Progress Bar: Removed double averaging of metrics.
  • LoRA Weights: Use float32 to avoid underflow/overflow risk.
  • Tree Utilities: Optimized tree.flatten and tree.map_structure for common cases.
  • Depthwise/Separable Conv: Removed backend-specific strides + dilation_rate restriction; validated output shapes in build; transposed channels_first to NHWC on CPU.
  • Regularizers: Allow plain callables as regularizers; fixed L1L2 regularizer.
  • Added AI Contribution Policy.
  • Added CITATION.cff for repository citation.
New Contributors

We would like to thank our new contributors for making their first contribution to the Keras project:

Full Changelog: https://github.com/keras-team/keras/compare/v3.14.0...v3.15.0

View original

Upgraded? How did it go?

Discussion