0 DSPs

[vkvasan]

How do you force a neural network to take '0' DSPs in HLS4ML ? Also is there a way that we could implement BNN or TNN with HLS4ML APIs ?

[vloncar]

DSPs won't be used if the precision of the weights and activations is less than 10 bits. You can train BNNs and TNNs with QKeras

[vkvasan]

I am having the precision less than 10 bits. The network is given below-

inputs = Input(shape=[98, ], name='qinput_features')
x = QDense(20, kernel_quantizer=quantized_bits(6,0,alpha=1), bias_quantizer=quantized_bits(6,0,alpha=1), name='fc1')(inputs)
x = QActivation(activation=quantized_relu(6), name='relu1')(x)
x = QDense(1, kernel_quantizer=quantized_bits(6,0,alpha=1), bias_quantizer=quantized_bits(6,0,alpha=1), name='output')(x)
x = Activation(activation="sigmoid", name='output_act')(x)
model = Model(inputs, x)

If I change the input shape from 98 to 50 , the number of DSP slices increases from 0 to 33. Can you please explain why this happens ?

[vloncar]

Look into the synthesis reports to see what exactly is using DSPs.