Access scale and zero_point from the weight within QuantizedLinearReLU

quantization
1

Hello,

I try to access scale and zero_point from the weight within QuantizedLinearReLU by using methods scale and q_scale. The results are failed as shown in below:

model.fc1
Out[7]: QuantizedLinearReLU(in_features=4, out_features=4, scale=0.04960649833083153, zero_point=0, qscheme=torch.per_channel_affine)

model.fc1.weight().scale
Out[8]: AttributeError: 'Tensor' object has no attribute 'scale'

model.fc1.weight().q_scale()
Out[9]: RuntimeError: Expected quantizer->qscheme() == kPerTensorAffine to be true, but got false.  (Could this error message be improved?  If so, please report an enhancement request to PyTorch.)

The scale and zero_point that I want to access are zero_point=tensor([0, 0, 0, 0]), axis=0 and scale=tensor([0.0145, 0.0016, 0.0132, 0.0124], dtype=torch.float64) as below:

model.fc1.weight()
Out[9]: 
tensor([[-0.1880, -1.3018,  0.8100,  1.8369],
        [-0.0033,  0.1172, -0.0065, -0.2083],
        [-0.9236,  0.3299, -1.6889, -1.3195],
        [-0.2718,  0.4078,  1.0997,  1.5693]], size=(4, 4), dtype=torch.qint8,
       quantization_scheme=torch.per_channel_affine,
       scale=tensor([0.0145, 0.0016, 0.0132, 0.0124], dtype=torch.float64),
       zero_point=tensor([0, 0, 0, 0]), axis=0)

Is there any way to access these parameters without manually copy from the prompt? Thank you.

1 Like
2

For per_channel_affine quant scheme, please use q_per_channel_scales() and q_per_channel_zero_points()

1 Like
3

It works perfectly, thank you @dskhudia.