What else is done after forward?

Ardeal · September 20, 2023, 4:31am

Hi,

	out  = self.qconv5_qbn5_qrelu5(out4) 
	out5 = self.qpool5(out)

	y = self.qinterpolate1(out5)

For the upper code, out5 is correct.

in self.qinterpolate1(out5) forward function, the return value x is also correct.

class QInterpolate(QModule):
    def __init__(self, qi=True, qo=True, num_bits=8, scale_factor=2, mode='bilinear', align_corners=True):
        super(QInterpolate, self).__init__(qi=qi, qo=qo, num_bits=num_bits)
        self.num_bits = num_bits
        self.scale_factor=scale_factor
        self.mode=mode
        self.align_corners = align_corners

    def forward(self, x):
        if hasattr(self, 'qi'):
            self.qi.update(x)
            x = FakeQuantize.apply(x, self.qi)
        x = F.interpolate(x, scale_factor=self.scale_factor, mode=self.mode, align_corners=self.align_corners)
        if hasattr(self, 'qo'):
            self.qo.update(x)
            x = FakeQuantize.apply(x, self.qo)
        return x

however,
y = self.qinterpolate1(out5) in the first piece of code is not correct.

More details:
x is:

while y is:
2dfd904270dbacdd18d28ab4a562d55

I cannot understand why round function is applied to the returned tensor y?

yiftach · September 20, 2023, 8:53am

Thanks for the detailed question, can you paste the data as code and not as an image though? Will make it easier to debug.

Ardeal · September 20, 2023, 11:58am

@yiftach
Thank you for your reply!

I am debugging the code on Windows 10 with vscode. The data you mentioned can’t be copied. I am trying to save the data to txt file and then paste to the forum.

however, I suddenly found a very weird issue:
please see the following picture, the data in y is different from yy, but I didn’t change anything in y. I just sliced y[0,0,:,:] to make yy

yiftach · September 20, 2023, 12:17pm

Interesting, but it is still hard to debug your issue without a complete code example. You could just print the tensor instead of saving to a file (you will not need to alter it). Try to see if this still happens for a small tensor, and if so - print the input too, so we can have a code snippet we can run to reproduce and study the issue.

Ardeal · September 20, 2023, 12:28pm

@yiftach

Please check the data I printed.
It is so weird!!!

I saw that the y in vscode is round-off, however, the printed y is correct.


print x
tensor([[[[0.0000, 0.3015, 0.6108,  ..., 1.4767, 0.9974, 0.5180],
          [0.0000, 0.1933, 0.3943,  ..., 1.7086, 1.2602, 0.8118],
          [0.0000, 0.0850, 0.1701,  ..., 1.9406, 1.5231, 1.1056],
          ...,
          [0.0000, 0.2087, 0.4175,  ..., 1.7473, 1.3066, 0.8659],
          [0.0000, 0.2629, 0.5180,  ..., 1.4226, 0.9278, 0.4330],
          [0.0000, 0.3093, 0.6185,  ..., 1.0979, 0.5489, 0.0000]],

         [[0.0000, 0.3015, 0.6108,  ..., 1.4767, 0.9974, 0.5180],
          [0.0000, 0.1933, 0.3943,  ..., 1.7086, 1.2602, 0.8118],
          [0.0000, 0.0850, 0.1701,  ..., 1.9406, 1.5231, 1.1056],
          ...,
          [0.0000, 0.2087, 0.4175,  ..., 1.7473, 1.3066, 0.8659],
          [0.0000, 0.2629, 0.5180,  ..., 1.4226, 0.9278, 0.4330],
          [0.0000, 0.3093, 0.6185,  ..., 1.0979, 0.5489, 0.0000]],

         [[0.0000, 0.3015, 0.6108,  ..., 1.4767, 0.9974, 0.5180],
          [0.0000, 0.1933, 0.3943,  ..., 1.7086, 1.2602, 0.8118],
          [0.0000, 0.0850, 0.1701,  ..., 1.9406, 1.5231, 1.1056],
          ...,
          [0.0000, 0.2087, 0.4175,  ..., 1.7473, 1.3066, 0.8659],
          [0.0000, 0.2629, 0.5180,  ..., 1.4226, 0.9278, 0.4330],
          [0.0000, 0.3093, 0.6185,  ..., 1.0979, 0.5489, 0.0000]],

         ...,

         [[0.0000, 0.3015, 0.6108,  ..., 1.4767, 0.9974, 0.5180],
          [0.0000, 0.1933, 0.3943,  ..., 1.7086, 1.2602, 0.8118],
          [0.0000, 0.0850, 0.1701,  ..., 1.9406, 1.5231, 1.1056],
          ...,
          [0.0000, 0.2087, 0.4175,  ..., 1.7473, 1.3066, 0.8659],
          [0.0000, 0.2629, 0.5180,  ..., 1.4226, 0.9278, 0.4330],
          [0.0000, 0.3093, 0.6185,  ..., 1.0979, 0.5489, 0.0000]],

         [[0.0000, 0.3015, 0.6108,  ..., 1.4767, 0.9974, 0.5180],
          [0.0000, 0.1933, 0.3943,  ..., 1.7086, 1.2602, 0.8118],
          [0.0000, 0.0850, 0.1701,  ..., 1.9406, 1.5231, 1.1056],
          ...,
          [0.0000, 0.2087, 0.4175,  ..., 1.7473, 1.3066, 0.8659],
          [0.0000, 0.2629, 0.5180,  ..., 1.4226, 0.9278, 0.4330],
          [0.0000, 0.3093, 0.6185,  ..., 1.0979, 0.5489, 0.0000]],

         [[0.0000, 0.3015, 0.6108,  ..., 1.4767, 0.9974, 0.5180],
          [0.0000, 0.1933, 0.3943,  ..., 1.7086, 1.2602, 0.8118],
          [0.0000, 0.0850, 0.1701,  ..., 1.9406, 1.5231, 1.1056],
          ...,
          [0.0000, 0.2087, 0.4175,  ..., 1.7473, 1.3066, 0.8659],
          [0.0000, 0.2629, 0.5180,  ..., 1.4226, 0.9278, 0.4330],
          [0.0000, 0.3093, 0.6185,  ..., 1.0979, 0.5489, 0.0000]]]],
       device='cuda:0', grad_fn=<FakeQuantizeBackward>)
print y
tensor([[[[0.0000, 0.3015, 0.6108,  ..., 1.4767, 0.9974, 0.5180], 
          [0.0000, 0.1933, 0.3943,  ..., 1.7086, 1.2602, 0.8118], 
          [0.0000, 0.0850, 0.1701,  ..., 1.9406, 1.5231, 1.1056], 
          ...,
          [0.0000, 0.2087, 0.4175,  ..., 1.7473, 1.3066, 0.8659], 
          [0.0000, 0.2629, 0.5180,  ..., 1.4226, 0.9278, 0.4330], 
          [0.0000, 0.3093, 0.6185,  ..., 1.0979, 0.5489, 0.0000]],

         [[0.0000, 0.3015, 0.6108,  ..., 1.4767, 0.9974, 0.5180], 
          [0.0000, 0.1933, 0.3943,  ..., 1.7086, 1.2602, 0.8118], 
          [0.0000, 0.0850, 0.1701,  ..., 1.9406, 1.5231, 1.1056], 
          ...,
          [0.0000, 0.2087, 0.4175,  ..., 1.7473, 1.3066, 0.8659], 
          [0.0000, 0.2629, 0.5180,  ..., 1.4226, 0.9278, 0.4330], 
          [0.0000, 0.3093, 0.6185,  ..., 1.0979, 0.5489, 0.0000]],

         [[0.0000, 0.3015, 0.6108,  ..., 1.4767, 0.9974, 0.5180], 
          [0.0000, 0.1933, 0.3943,  ..., 1.7086, 1.2602, 0.8118], 
          [0.0000, 0.0850, 0.1701,  ..., 1.9406, 1.5231, 1.1056], 
          ...,
          [0.0000, 0.2087, 0.4175,  ..., 1.7473, 1.3066, 0.8659], 
          [0.0000, 0.2629, 0.5180,  ..., 1.4226, 0.9278, 0.4330], 
          [0.0000, 0.3093, 0.6185,  ..., 1.0979, 0.5489, 0.0000]],

         ...,

         [[0.0000, 0.3015, 0.6108,  ..., 1.4767, 0.9974, 0.5180], 
          [0.0000, 0.1933, 0.3943,  ..., 1.7086, 1.2602, 0.8118], 
          [0.0000, 0.0850, 0.1701,  ..., 1.9406, 1.5231, 1.1056], 
          ...,
          [0.0000, 0.2087, 0.4175,  ..., 1.7473, 1.3066, 0.8659], 
          [0.0000, 0.2629, 0.5180,  ..., 1.4226, 0.9278, 0.4330], 
          [0.0000, 0.3093, 0.6185,  ..., 1.0979, 0.5489, 0.0000]],

         [[0.0000, 0.3015, 0.6108,  ..., 1.4767, 0.9974, 0.5180], 
          [0.0000, 0.1933, 0.3943,  ..., 1.7086, 1.2602, 0.8118], 
          [0.0000, 0.0850, 0.1701,  ..., 1.9406, 1.5231, 1.1056], 
          ...,
          [0.0000, 0.2087, 0.4175,  ..., 1.7473, 1.3066, 0.8659],
          [0.0000, 0.2629, 0.5180,  ..., 1.4226, 0.9278, 0.4330],
          [0.0000, 0.3093, 0.6185,  ..., 1.0979, 0.5489, 0.0000]],

         [[0.0000, 0.3015, 0.6108,  ..., 1.4767, 0.9974, 0.5180],
          [0.0000, 0.1933, 0.3943,  ..., 1.7086, 1.2602, 0.8118],
          [0.0000, 0.0850, 0.1701,  ..., 1.9406, 1.5231, 1.1056],
          ...,
          [0.0000, 0.2087, 0.4175,  ..., 1.7473, 1.3066, 0.8659],
          [0.0000, 0.2629, 0.5180,  ..., 1.4226, 0.9278, 0.4330],
          [0.0000, 0.3093, 0.6185,  ..., 1.0979, 0.5489, 0.0000]]]],
       device='cuda:0', grad_fn=<FakeQuantizeBackward>)

Ardeal · September 21, 2023, 1:56am

@yiftach

Thank you for you reply!

This issue is fixed.
The reason is not related with Pytorch. It is the reason of VScode.
For any unknown reason, I cleaned all vscode cache files, and then, the issue is fixed.