tensorflow_model_optimization/python/core/quantization/keras/default_8bit/default_8bit_quantize_registry.py [72:180]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - quantize_registry.QuantizeRegistry, _RNNHelper): """QuantizationRegistry for built-in Keras classes for default 8-bit scheme.""" # TODO(tfmot): expand layers test in quantize_functional_test.py # to add more layers to allowlist. _LAYER_QUANTIZE_INFO = [ # Activation Layers _QuantizeInfo(layers.ReLU, [], [], True), _QuantizeInfo(layers.Softmax, [], []), # Enable once verified. # layers.ELU, _QuantizeInfo(layers.LeakyReLU, [], [], True), # layers.PReLU, # layers.ThresholdedReLU, # Convolution Layers # _QuantizeInfo(layers.Conv1D, ['kernel'], ['activation']), # layers.Conv2D is supported and handled in code below. # layers.DepthwiseConv2D is supported and handled in code below. # _QuantizeInfo(layers.Conv3D, ['kernel'], ['activation']), # _QuantizeInfo(layers.Conv3DTranspose, ['kernel'], ['activation']), _QuantizeInfo(layers.Concatenate, [], [], True), _no_quantize(layers.Cropping1D), _no_quantize(layers.Cropping2D), _no_quantize(layers.Cropping3D), # _no_quantize(layers.UpSampling1D), # TODO(tfmot): Reduce the quantization errors for bilinear interpolation # type for UpSampling2D op. UpSampling2D supports two interpolation types, # nearest and bilinear. we convert the op to ResizeBilnear integer op on # TFLite. This ResizeBilinear TFLite op only for input and output has the # same quantization parameters. (scale and zero_point) To do that, The # TFLite converter inserts quantization cast op right after the input to # match quantization params for the output. Current QAT doesn’t consider # this behavior yet, so now we have larger quantization errors than we # expected. We have to add support for it on QAT or change the TFLite # kernel op to support different quantization params for input and output. # (Note that the nearest case just copies the number so there’s no more # errors even if the quantization order is different.) _QuantizeInfo(layers.UpSampling2D, [], [], True), # _no_quantize(layers.UpSampling3D), _no_quantize(layers.ZeroPadding1D), _no_quantize(layers.ZeroPadding2D), # _no_quantize(layers.ZeroPadding3D), # Supported via modifications in Transforms. # layers.SeparableConv1D, layers.SeparableConv2D, # Core Layers _no_quantize(layers.ActivityRegularization), _QuantizeInfo(layers.Dense, ['kernel'], ['activation']), _no_quantize(layers.Dropout), _no_quantize(layers.Flatten), # _no_quantize(layers.Masking), _no_quantize(layers.Permute), # _no_quantize(layers.RepeatVector), _no_quantize(layers.Reshape), _no_quantize(layers.SpatialDropout1D), _no_quantize(layers.SpatialDropout2D), _no_quantize(layers.SpatialDropout3D), # layers.Lambda needs custom handling by the user. # Pooling Layers _QuantizeInfo(layers.AveragePooling1D, [], [], True), _QuantizeInfo(layers.AveragePooling2D, [], [], True), # _QuantizeInfo(layers.AveragePooling3D, [], [], True), _QuantizeInfo(layers.GlobalAveragePooling1D, [], [], True), _QuantizeInfo(layers.GlobalAveragePooling2D, [], [], True), _QuantizeInfo(layers.GlobalAveragePooling3D, [], [], True), _no_quantize(layers.GlobalMaxPooling1D), _no_quantize(layers.GlobalMaxPooling2D), _no_quantize(layers.GlobalMaxPooling3D), # _no_quantize(layers.MaxPooling1D), _no_quantize(layers.MaxPooling2D), # _no_quantize(layers.MaxPooling3D), # _QuantizeInfo(layers.LocallyConnected1D, ['kernel'], ['activation']), # _QuantizeInfo(layers.LocallyConnected2D, ['kernel'], ['activation']), _QuantizeInfo(layers.Add, [], [], True), # Enable once verified with TFLite behavior. # layers.Embedding: ['embeddings'], # BatchNormalization is handled elsewhere, in the cases # where it's preceded by convolutional layers. # layers.BatchNormalization: [], # Merge layers to be added. # RNN Cells # TODO(pulkitb): Verify RNN layers behavior. # TODO(tfmot): check if we still need to allowlist via compat.v1 and # compat.v2 to support legacy TensorFlow 2.X # behavior where the v2 RNN uses the v1 RNNCell instead of the v2 RNNCell. # See b/145939875 for details. # _QuantizeInfo(tf.keras.layers.GRUCell, ['kernel', 'recurrent_kernel'], # ['activation', 'recurrent_activation']), # _QuantizeInfo(tf.keras.layers.LSTMCell, ['kernel', 'recurrent_kernel'], # ['activation', 'recurrent_activation']), # _QuantizeInfo(tf.keras.experimental.PeepholeLSTMCell, # ['kernel', 'recurrent_kernel'], # ['activation', 'recurrent_activation']), # _QuantizeInfo(tf.keras.layers.SimpleRNNCell, # ['kernel', 'recurrent_kernel'], # ['activation', 'recurrent_activation']), ] - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - tensorflow_model_optimization/python/core/quantization/keras/experimental/default_n_bit/default_n_bit_quantize_registry.py [81:189]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - quantize_registry.QuantizeRegistry, _RNNHelper): """QuantizationRegistry for built-in Keras classes for default 8-bit scheme.""" # TODO(tfmot): expand layers test in quantize_functional_test.py # to add more layers to allowlist. _LAYER_QUANTIZE_INFO = [ # Activation Layers _QuantizeInfo(layers.ReLU, [], [], True), _QuantizeInfo(layers.Softmax, [], []), # Enable once verified. # layers.ELU, _QuantizeInfo(layers.LeakyReLU, [], [], True), # layers.PReLU, # layers.ThresholdedReLU, # Convolution Layers # _QuantizeInfo(layers.Conv1D, ['kernel'], ['activation']), # layers.Conv2D is supported and handled in code below. # layers.DepthwiseConv2D is supported and handled in code below. # _QuantizeInfo(layers.Conv3D, ['kernel'], ['activation']), # _QuantizeInfo(layers.Conv3DTranspose, ['kernel'], ['activation']), _QuantizeInfo(layers.Concatenate, [], [], True), _no_quantize(layers.Cropping1D), _no_quantize(layers.Cropping2D), _no_quantize(layers.Cropping3D), # _no_quantize(layers.UpSampling1D), # TODO(tfmot): Reduce the quantization errors for bilinear interpolation # type for UpSampling2D op. UpSampling2D supports two interpolation types, # nearest and bilinear. we convert the op to ResizeBilnear integer op on # TFLite. This ResizeBilinear TFLite op only for input and output has the # same quantization parameters. (scale and zero_point) To do that, The # TFLite converter inserts quantization cast op right after the input to # match quantization params for the output. Current QAT doesn’t consider # this behavior yet, so now we have larger quantization errors than we # expected. We have to add support for it on QAT or change the TFLite # kernel op to support different quantization params for input and output. # (Note that the nearest case just copies the number so there’s no more # errors even if the quantization order is different.) _QuantizeInfo(layers.UpSampling2D, [], [], True), # _no_quantize(layers.UpSampling3D), _no_quantize(layers.ZeroPadding1D), _no_quantize(layers.ZeroPadding2D), # _no_quantize(layers.ZeroPadding3D), # Supported via modifications in Transforms. # layers.SeparableConv1D, layers.SeparableConv2D, # Core Layers _no_quantize(layers.ActivityRegularization), _QuantizeInfo(layers.Dense, ['kernel'], ['activation']), _no_quantize(layers.Dropout), _no_quantize(layers.Flatten), # _no_quantize(layers.Masking), _no_quantize(layers.Permute), # _no_quantize(layers.RepeatVector), _no_quantize(layers.Reshape), _no_quantize(layers.SpatialDropout1D), _no_quantize(layers.SpatialDropout2D), _no_quantize(layers.SpatialDropout3D), # layers.Lambda needs custom handling by the user. # Pooling Layers _QuantizeInfo(layers.AveragePooling1D, [], [], True), _QuantizeInfo(layers.AveragePooling2D, [], [], True), # _QuantizeInfo(layers.AveragePooling3D, [], [], True), _QuantizeInfo(layers.GlobalAveragePooling1D, [], [], True), _QuantizeInfo(layers.GlobalAveragePooling2D, [], [], True), _QuantizeInfo(layers.GlobalAveragePooling3D, [], [], True), _no_quantize(layers.GlobalMaxPooling1D), _no_quantize(layers.GlobalMaxPooling2D), _no_quantize(layers.GlobalMaxPooling3D), # _no_quantize(layers.MaxPooling1D), _no_quantize(layers.MaxPooling2D), # _no_quantize(layers.MaxPooling3D), # _QuantizeInfo(layers.LocallyConnected1D, ['kernel'], ['activation']), # _QuantizeInfo(layers.LocallyConnected2D, ['kernel'], ['activation']), _QuantizeInfo(layers.Add, [], [], True), # Enable once verified with TFLite behavior. # layers.Embedding: ['embeddings'], # BatchNormalization is handled elsewhere, in the cases # where it's preceded by convolutional layers. # layers.BatchNormalization: [], # Merge layers to be added. # RNN Cells # TODO(pulkitb): Verify RNN layers behavior. # TODO(tfmot): check if we still need to allowlist via compat.v1 and # compat.v2 to support legacy TensorFlow 2.X # behavior where the v2 RNN uses the v1 RNNCell instead of the v2 RNNCell. # See b/145939875 for details. # _QuantizeInfo(tf.keras.layers.GRUCell, ['kernel', 'recurrent_kernel'], # ['activation', 'recurrent_activation']), # _QuantizeInfo(tf.keras.layers.LSTMCell, ['kernel', 'recurrent_kernel'], # ['activation', 'recurrent_activation']), # _QuantizeInfo(tf.keras.experimental.PeepholeLSTMCell, # ['kernel', 'recurrent_kernel'], # ['activation', 'recurrent_activation']), # _QuantizeInfo(tf.keras.layers.SimpleRNNCell, # ['kernel', 'recurrent_kernel'], # ['activation', 'recurrent_activation']), ] - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -