Dobiasd/frugally-deep v0.20.0

on GitHub

Minor release adding broad new Keras 3 layer support (#465).

Recurrent layers (restored after the Keras 3 weight-format change broke the previous implementation, plus newly added):

• LSTM, GRU, SimpleRNN, Bidirectional (all merge_mode values)
• ConvLSTM1D, ConvLSTM2D, ConvLSTM3D
• RNN wrapping LSTMCell / GRUCell / SimpleRNNCell / StackedRNNCells

Attention / dense:

• EinsumDense (generic einsum interpreter; covers transformer-style projections)
• GroupedQueryAttention (with optional use_gate=True)

Pooling / convolution:

• AdaptiveAveragePooling1D/2D/3D, AdaptiveMaxPooling1D/2D/3D
• DepthwiseConv1D

Normalization:

• RMSNormalization
• GroupNormalization

Preprocessing:

• Discretization, IntegerLookup
• Masking (passthrough)
• Many training-only image augmentation layers as identity passthroughs (RandomBrightness, RandomFlip, RandomCrop, RandomContrast, RandomRotation, RandomTranslation, RandomZoom, RandomHue, RandomSaturation, RandomSharpness, AugMix, CutMix, MixUp, RandAugment, Equalization, Solarization, Pipeline, MaxNumBoundingBoxes, etc.)

Numerical correctness verified end-to-end against Keras 3.14 across multiple distinct configurations of every new layer.