How to use "squeezenet" as backend to train model?? about cascadepsp

[hkchengrex]

We modified the first conv in the network model to take 3 extra channels and initialized them with zeros. You need to change it in a similar way. For speed you can reduce it to a 2-level or even 1-level cascade -- the performance will drop but it will run much faster.

[wangjing60755]

can you put these codes? I can't modify

[hkchengrex]

You can find our modifications in models/psp/extractors.py.

[wangjing60755]

I find：
if v1.shape != tar_v.shape:
# Init the new segmentation channel with zeros
c, _, w, h = v1.shape
tar_v = torch.cat([
tar_v,
torch.zeros((c,3,w,h)),
], 1)

but no use to me.
when I run train.py use "model = PSPNet(sizes=(1, 2, 3, 6), psp_size=2048, deep_features_size=1024, backend='squeezenet1_1')"
or " model = PSPNet(sizes=(1, 2, 3, 6), psp_size=2048, deep_features_size=1024, backend='resnet18') ", I meet error .I can,t know how to change the code to adapt！

[hkchengrex]

Only ResNet-50 is supported in this repo. You would have to include those extensions by yourself.