different layer names with official pytorch ResNet

[Keson96]

Although the _ConvBatchNormReLU abstraction is very handy for building the model, it requires extra conversion when loading a pretrained ResNet model as the names of parameters differ. Could you use the official pytorch ResNet code? Thanks.

[kazuto1011]

I would like to keep readability and minimality of the model definition with the abstraction, for caffe-to-pytorch porting and comparison in versions.

To leverage the torchvision ResNet, I think it's more simple to dynamically modify it to DeepLab.

# PyTorch ResNet-101 -> DeepLab v2
# A bit different in the stride position
import torch
import torch.nn as nn
from torchvision import models
from collections import OrderedDict

from libs.models.deeplabv2 import _ASPPModule

model = models.resnet101(pretrained=True)

# Layer 3 (OS=16 -> OS=8)
model.layer3[0].conv2.stride = (1, 1)
model.layer3[0].downsample[0].stride = (1, 1)
for m in model.layer3[1:]:
    m.conv2.padding = (2, 2)
    m.conv2.dilation = (2, 2)

# Layer 4 (OS=32 -> OS=8)
model.layer4[0].conv2.stride = (1, 1)
model.layer4[0].downsample[0].stride = (1, 1)
for m in model.layer4[1:]:
    m.conv2.padding = (4, 4)
    m.conv2.dilation = (4, 4)

# Remove "avgpool" and "fc", and add ASPP
model = list(model.named_children())[:-2]
model += [("aspp", _ASPPModule(2048, 21, [6, 12, 18, 24]))]
model = nn.Sequential(OrderedDict(model))

# The model output is sub-sampled by 8 instead of 32
image = torch.randn(1, 3, 224, 224)
logit = model(image)

print(logit.shape)

[Keson96]

I haven't considered dynamic modification. Thanks.