show the CAM in some of trained VGG models by pytorch using CIFAR10
VGG models: VGG9_conv, VGG9_avgpool, ......
VGG9_conv means the last classifier layer is conv2d instead of fc
VGG9_avgpool means before the fc layer, use avgpool first
When I use two GPUs to train the model, althought the acc is high enough, the results of CAM isn't good as well.
| test2.jpg | test3.jpg | test4.jpg |
|---|---|---|
 |
 |
 |
| VGG11_avgpool | VGG11_conv | VGG9_avgpool | VGG9_conv |
|---|---|---|---|
 |
 |
 |
 |
| VGG11_avgpool | VGG11_conv | VGG9_avgpool | VGG9_conv |
|---|---|---|---|
 |
 |
 |
 |
| VGG11_avgpool | VGG11_conv | VGG9_avgpool | VGG9_conv |
|---|---|---|---|
 |
 |
 |
 |
The output classes of test4.jpg are "bird", but it can easy to see that the model didn't learn the features well. I think one of the reason is that the VGGnet is too deep for the CIFAR10 dataset. After convolutional calculation, 3x32x32 images transfered to 512x1x1 features vector.