SuperResolution: How to troubleshoot tensor size mismatch error?

Part 1 (2019) Advanced (Part 1 v3)
1

Would appreciate any help on how to troubleshoot tensor size mismatch error. I’m using the superesolution notebook as a template, but I keep getting size mismatch error.

Initially, I thought it was because my image is a rectangle (one of the dimension errors is the dimensions of my image), so I resized it to a square of 135 x 135, but no luck.

It fails on the .fit command, and here’s the end of the error output:

~/anaconda3/envs/dl_2019_06/lib/python3.6/site-packages/fastai/basic_train.py in loss_batch(model, xb, yb, loss_func, opt, cb_handler)
     24     if not is_listy(xb): xb = [xb]
     25     if not is_listy(yb): yb = [yb]
---> 26     out = model(*xb)
     27     out = cb_handler.on_loss_begin(out)
     28 

~/anaconda3/envs/dl_2019_06/lib/python3.6/site-packages/torch/nn/modules/module.py in __call__(self, *input, **kwargs)
    491             result = self._slow_forward(*input, **kwargs)
    492         else:
--> 493             result = self.forward(*input, **kwargs)
    494         for hook in self._forward_hooks.values():
    495             hook_result = hook(self, input, result)

~/anaconda3/envs/dl_2019_06/lib/python3.6/site-packages/fastai/layers.py in forward(self, x)
    153         for l in self.layers:
    154             res.orig = x
--> 155             nres = l(res)
    156             # We have to remove res.orig to avoid hanging refs and therefore memory leaks
    157             res.orig = None

~/anaconda3/envs/dl_2019_06/lib/python3.6/site-packages/torch/nn/modules/module.py in __call__(self, *input, **kwargs)
    491             result = self._slow_forward(*input, **kwargs)
    492         else:
--> 493             result = self.forward(*input, **kwargs)
    494         for hook in self._forward_hooks.values():
    495             hook_result = hook(self, input, result)

~/anaconda3/envs/dl_2019_06/lib/python3.6/site-packages/fastai/layers.py in forward(self, x)
    170         self.dense=dense
    171 
--> 172     def forward(self, x): return torch.cat([x,x.orig], dim=1) if self.dense else (x+x.orig)
    173 
    174 def res_block(nf, dense:bool=False, norm_type:Optional[NormType]=NormType.Batch, bottle:bool=False, **conv_kwargs):

RuntimeError: invalid argument 0: Sizes of tensors must match except in dimension 1. Got 135 and 136 in dimension 2 at /opt/conda/conda-bld/pytorch_1556653099582/work/aten/src/THC/generic/THCTensorMath.cu:71

I should also point out that I’ve checked that all my input data are of dimensions (135,135,3), and I have no idea where the 136 dimension in the error message is coming from!!!

Here’s my datagen. I’m using a pandas df that has corresponding filenames for the crappy and non-crappy images (filename_y, filename_x):

image_gen = get_transforms(do_flip=False, max_rotate=0.0, max_zoom=1.2, max_lighting=0.0, max_warp=0.0)

trn_src = (ImageImageList.from_df(trnval_df, path="/", cols='filename_y')
               .split_by_rand_pct(valid_pct=0.1, seed=7)
               .label_from_df(cols='filename_x'))
    
trn_data = (trn_src.transform(image_gen, size=(135,135), tfm_y=True)
            .databunch(bs=batch_size)
            .normalize(imagenet_stats, do_y=True))
2

You need to make your dims divisible by 2, these errors arise from padding issues. With fastai unets I usually ensure div by 32 so this error doesn’t arise at an interim layer. So try 128. Please post back if this helps.

See Gland segmentation using Lesson 3 fastai v1

3

Thanks, it worked on 128x128. Then I tried 136x136 and 136x240 (closest to my original image resolution), and those worked. However 136 is not divisible by 32…, and it’s also going to be difficult to get both dimensions divisible by 32 if I want to keep the aspect ratio.

4

I’ve padded input images when needed, with a solid colour like yellow that doesn’t show up in images, in order to keep dims sane. Also if input images have varying dims. You can do this with preprocessing or your own open_image function.