Bug in ResNet of ModulatedDeformConv version

[wang2wei1]

File backbones/resnet.py -> class BasicBlock(nn.Module) -> def forward(self, x) ->

offset = offset_mask[:, :18, :, :]
mask = offset_mask[:, -9:, :, :].sigmoid()

we know shape(offset_mask)[1] = deformable_groups * offset_channels with deformable_groups = 1 and offset_channels = 27. From definition, offset_channels = 27 is fixed, but deformable_groups is a parameter to be passed in with default value 1. If a value other than 1 is passed in to deformable_groups, then shape(offset_mask)[1] = deformable_groups * offset_channels > 27 and values in offset_mask[:, 18:-9, :, :] is discarded in the code segment above. Same problem in class Bottleneck.

[wang2wei1]

Another bug:
backbones/resnet_fpn.py -> def Resnet34FPN ->

bottom_up = resnet50(pretrained=resnet_pretrained)
top_down = FPNTopDown([2048, 1024, 512, 256], 256)

should be

bottom_up = resnet34(pretrained=resnet_pretrained)
top_down = FPNTopDown([512, 256, 128, 64], 256)

[wanzysky]

Thanks a lot for pointing out.
We will double check and fix it.

[wang2wei1]

data/meta_loaders/json_meta_loader.py -> class JsonMetaLoader(Configurable) -> def load_meta(self, json_path) ->

with open(self.same_dir_with(json_path), self.image_folder) as reader:

should be

with open(self.same_dir_with(json_path, self.image_folder)) as reader: