Why does vgg19_bn (pretrained=True) give size mismatch error with images with large sizes?

1

I’m using the following code :
vgg_model = models.vgg19_bn(pretrained=True)

If I crop my images to 240*240, the network works fine, but for the following sizes, it throws a size mismatch error
(3L, 480L, 640L)
DAMN --------------------
size mismatch, m1: [1 x 153600], m2: [25088 x 4096] at /pytorch/torch/lib/TH/generic/THTensorMath.c:1293
(3L, 426L, 640L)
DAMN --------------------
size mismatch, m1: [1 x 133120], m2: [25088 x 4096] at /pytorch/torch/lib/TH/generic/THTensorMath.c:1293
(3L, 428L, 640L)
DAMN --------------------
size mismatch, m1: [1 x 133120], m2: [25088 x 4096] at /pytorch/torch/lib/TH/generic/THTensorMath.c:1293
(3L, 425L, 640L)
DAMN --------------------
size mismatch, m1: [1 x 133120], m2: [25088 x 4096] at /pytorch/torch/lib/TH/generic/THTensorMath.c:1293
(3L, 640L, 481L)
DAMN --------------------
size mismatch, m1: [1 x 153600], m2: [25088 x 4096] at /pytorch/torch/lib/TH/generic/THTensorMath.c:1293

Any ideas as to how to correct this?

2

The first Linear layer in vgg is defined as nn.Linear(512 * 7 * 7, 4096): if your input image is 240x240, after “feature process”, the output can be view as 512*7*7. But if your input is 480x640，the output of “feature process” is not 512*7*7，so there is the error of mismatch.

You can see more detail in the vgg

1 Like