oneAPI RNN State and Activation Quantizers

[laurilaatu]

Description

Adding support for QKeras state_quantizer for RNNs

Type of change

Extract state_quantizer bitwidth and set the accum_t to correspond to it. Change hidden state to be of accum_t instead of res_T.

Extract quantized activations from RNNs, with the exception of hard activations.

• Bug fix (non-breaking change that fixes an issue)

Tests

model = Sequential()

model.add(qkeras.QLSTM(15,
                            kernel_quantizer=qkeras.quantizers.quantized_bits(8,1,0,alpha=1),
                            recurrent_quantizer=qkeras.quantizers.quantized_bits(9,1,0,alpha=1),
                            state_quantizer=qkeras.quantizers.quantized_bits(10,1,0,alpha=1),
                            bias_quantizer=qkeras.quantizers.quantized_bits(11,1,0,alpha=1),
                            activation=qkeras.utils.quantized_tanh(use_real_tanh=True),
                            recurrent_activation=qkeras.utils.quantized_sigmoid(use_real_sigmoid=True),
                            input_shape=(10,4)
                           ))

model.add(qkeras.QDense(1, activation=qkeras.utils.quantized_tanh(use_real_tanh=True)))

model.compile(optimizer='adam', loss='mse')



config = hls4ml.utils.config_from_keras_model(model, granularity='name', backend='oneAPI')

hls_model = hls4ml.converters.convert_from_keras_model(
    model, hls_config=config, backend='oneAPI', output_dir='model_1/hls4ml_prj', part='Agilex7'
)

hls_model.compile()

Checklist

• I have read the guidelines for contributing.
• I have commented my code, particularly in hard-to-understand areas.
• [] I have made corresponding changes to the documentation.
• My changes generate no new warnings.
• I have installed and run pre-commit on the files I edited or added.
• I have added tests that prove my fix is effective or that my feature works.

[vloncar]

Looks good overall, minor cosmetic changes would be nice before finalizing