Bug in testing Set5 X3 with `args.chop=True`

When I set args.chop=False, everything goes well. When I set it to True, it outputs errors as:

Evaluation:################################################   2019-10-10-16:38:03
 40%|██████████████████                           | 2/5 [00:02<00:04,  1.36s/it]
Traceback (most recent call last):
  File "/anaconda3/lib/python3.7/site-packages/IPython/core/interactiveshell.py", line 3325, in run_code
    exec(code_obj, self.user_global_ns, self.user_ns)
  File "<ipython-input-2-8439669b053c>", line 1, in <module>
    runfile('/Code/EDSR-PyTorch-master/src/main.py', wdir='/Code/EDSR-PyTorch-master/src')
  File "/.pycharm_helpers/pydev/_pydev_bundle/pydev_umd.py", line 197, in runfile
    pydev_imports.execfile(filename, global_vars, local_vars)  # execute the script
  File "/.pycharm_helpers/pydev/_pydev_imps/_pydev_execfile.py", line 18, in execfile
    exec(compile(contents+"\n", file, 'exec'), glob, loc)
  File "/Code/EDSR-PyTorch-master/src/main.py", line 41, in <module>
    main()
  File "/Code/EDSR-PyTorch-master/src/main.py", line 32, in main
    while not t.terminate():
  File "/Code/EDSR-PyTorch-master/src/trainer.py", line 141, in terminate
    self.test()
  File "/Code/EDSR-PyTorch-master/src/trainer.py", line 91, in test
    sr = self.model(lr, idx_scale)
  File "/anaconda3/lib/python3.7/site-packages/torch/nn/modules/module.py", line 547, in __call__
    result = self.forward(*input, **kwargs)
  File "/Code/EDSR-PyTorch-master/src/model/__init__.py", line 62, in forward
    return forward_function(x)
  File "/Code/EDSR-PyTorch-master/src/model/__init__.py", line 175, in forward_chop
    _y[..., top, right] = y_chop[1][..., top, right_r]
RuntimeError: The expanded size of the tensor (127) must match the existing size (128) at non-singleton dimension 3.  Target sizes: [1, 3, 127, 127].  Tensor sizes: [3, 127, 128]

Note that I have no problem with training. In testing, for X2 and X4, args.chop=True works good. For X3, some images in Set5 and Set14 would have problems. By the way, I download the processed datasets directly from your link in Readme.

[HolmesShuan]

Edit EDSR-PyTorch-master/src/model/__init__.py as follow:

def forward_chop(self, *args, shave=10, min_size=160000):
        scale = 1 if self.input_large else self.scale[self.idx_scale]
        n_GPUs = min(self.n_GPUs, 4)
        # height, width
        h, w = args[0].size()[-2:]

        h_half, w_half = h//2, w//2
        h_size, w_size = h_half + shave, w_half + shave

        top = slice(0, h_size)
        bottom = slice(h - h_size, h)
        left = slice(0, w_size)
        right = slice(w - w_size, w)
        x_chops = [torch.cat([
            a[..., top, left],
            a[..., top, right],
            a[..., bottom, left],
            a[..., bottom, right]
        ]) for a in args]

        y_chops = []
        if h * w < 4 * min_size:
            for i in range(0, 4, n_GPUs):
                x = [x_chop[i:(i + n_GPUs)] for x_chop in x_chops]
                y = P.data_parallel(self.model, *x, range(n_GPUs))
                if not isinstance(y, list): y = [y]
                if not y_chops:
                    y_chops = [[c for c in _y.chunk(n_GPUs, dim=0)] for _y in y]
                else:
                    for y_chop, _y in zip(y_chops, y):
                        y_chop.extend(_y.chunk(n_GPUs, dim=0))
        else:
            for p in zip(*x_chops):
                p1 = [p[0].unsqueeze(0)]
                y = self.forward_chop(*p1, shave=shave, min_size=min_size)
                if not isinstance(y, list): y = [y]
                if not y_chops:
                    y_chops = [[_y] for _y in y]
                else:
                    for y_chop, _y in zip(y_chops, y): y_chop.append(_y)

        h, w = scale * h, scale * w
        h_half, w_half = scale * h_half, scale * w_half
        h_size, w_size = scale * h_size, scale * w_size
        shave *= scale

        # h *= scale
        # w *= scale

        # top = slice(0, h_half)
        # bottom = slice(h - h_half, h)
        # bottom_r = slice(h//2 - h, None)
        # left = slice(0, w_half)
        # right = slice(w - w_half, w)
        # right_r = slice(w//2 - w, None)

        # batch size, number of color channels
        b, c = y_chops[0][0].size()[:-2]
        y = [y_chop[0].new(b, c, h, w) for y_chop in y_chops]
        for y_chop, _y in zip(y_chops, y):
            _y[..., 0:h_half, 0:w_half] = y_chop[0][..., 0:h_half, 0:w_half]
            _y[..., 0:h_half, w_half:w] = y_chop[1][..., 0:h_half, (w_size - w + w_half):w_size]
            _y[..., h_half:h, 0:w_half] = y_chop[2][..., (h_size - h + h_half):h_size, 0:w_half]
            _y[..., h_half:h, w_half:w] = y_chop[3][..., (h_size - h + h_half):h_size, (w_size - w + w_half):w_size]

        if len(y) == 1: y = y[0]

        return y

[Senwang98]

@HolmesShuan, Thanks for your reply! I have solved this issue by replacing forward_chop in EDSR with forward_chop in RCAN(PyTorch0.4.0). My model shows this change doesn’t affect my accuracy. But, I still thank you very much!