RuntimeError: inconsistent tensor size

1

Hi, when I train CNN, I got error message below:

$ python train.py

0 0.39379167556762695
1 0.3715594410896301
2 0.21380575001239777
3 0.18399034440517426
4 0.14919331669807434
5 0.3483908176422119
6 0.261790931224823
7 0.3059869706630707
8 0.3544842600822449
9 0.3003101348876953
10 0.28798148036003113
11 0.2801303565502167
12 0.17284557223320007
13 0.15419822931289673
14 0.18459680676460266
15 0.1608618199825287
Traceback (most recent call last):
  ...
RuntimeError: inconsistent tensor size, expected tensor [3 x 384 x 256]
 and src [1 x 384 x 256] to have the same number of elements, but got 
294912 and 98304 elements respectively at /opt/conda/conda-bld/pytor
ch-cpu_1515613813020/work/torch/lib/TH/generic/THTensorCopy.c:86

I’m sorry but I can’t show you the running code.

I tried several time but this problem only occurs on 15th epoch.
So I guess this is related to illegal format of input image which is sampled on 15th epoch. Is this right?
and When the code encounters the illegal format image, how can I ignore or skip it?

Thanks.

2

The error message is very clear. You even hid the trace so I can’t know which call errors, but it seems to require two input tensors to have same numel. You can check that and skip if you want.

3

As you said this is simple problem and solved.Thanks.