mismatch with torchvision resnets

[CarloLucibello]

I'm using the script
https://github.com/FluxML/Metalhead.jl/blob/master/scripts/port_torchvision.jl
to load torchvision's models and copy their weights into Metalhead's ones.

With vggX model all is fine.

With resnets instead I get the following mismatchs:

• ResNet18:

  flux_key = "model.layers[1].layers[3].layers[1].layers[1].layers[1].conv_weight"
  size(flux_param) = (1, 1, 64, 128)
  pytorch_key = "layer2.0.conv1.weight"
  size(pytorch_param) = (3, 3, 64, 128)
  

• ResNet34

  flux_key = "model.layers[1].layers[3].layers[1].layers[1].layers[1].conv_weight"
  size(flux_param) = (1, 1, 64, 128)
  pytorch_key = "layer2.0.conv1.weight"
  size(pytorch_param) = (3, 3, 64, 128)
  

• ResNet50

  flux_key = "model.layers[1].layers[2].layers[1].layers[1].layers[1].conv_weight"
  size(flux_param) = (1, 1, 64, 256)
  pytorch_key = "layer1.0.conv1.weight"
  size(pytorch_param) = (1, 1, 64, 64)
  

• ResNet101

  flux_key = "model.layers[1].layers[2].layers[1].layers[1].layers[1].conv_weight"
  size(flux_param) = (1, 1, 64, 256)
  pytorch_key = "layer1.0.conv1.weight"
  size(pytorch_param) = (1, 1, 64, 64)
  

• ResNet152

  flux_key = "model.layers[1].layers[2].layers[1].layers[1].layers[1].conv_weight"
  size(flux_param) = (1, 1, 64, 256)
  pytorch_key = "layer1.0.conv1.weight"
  size(pytorch_param) = (1, 1, 64, 64)
  

[theabhirath]

It's because Parallel gives the layers in a different order than the one in torchvision. Enumerating over reverse(node.layers) at

Metalhead.jl/scripts/pytorch2flux.jl

Line 79 in e3a82ff

for (i, n) in enumerate(node.layers)

should fix this