I found you did not use the self.cls_nonlocal() and self.reg_nonlocal in roi_box_feature_extractors.py.The number of blocks "self.cls_num_stack" and "self.reg_num_stack" are zero

[yangmin666]

codes are here:https://github.com/wuyuebupt/doubleheadsrcnn/blob/master/maskrcnn_benchmark/modeling/roi_heads/box_head/roi_box_feature_extractors.py.
And another question:I found no difference between the reg and cls operations. All the network structure parameters are the same except some details.

[wuyuebupt]

@yangmin666 You are right, cls_nonloca and reg_nonlocal is not used. Instead, self.shared_nonlocal is used.

The code is at:
https://github.com/wuyuebupt/doubleheadsrcnn/blob/master/cmd_train.sh#L35

doubleheadsrcnn/maskrcnn_benchmark/modeling/roi_heads/box_head/roi_box_feature_extractors.py

Line 186 in a744b41

for i in range(self.shared_num_stack):

There are four losses in total: conv reg, conv cls, fc reg and fc cls. Conv cls and conv cls will share one head, e.g.

doubleheadsrcnn/maskrcnn_benchmark/modeling/roi_heads/box_head/roi_box_feature_extractors.py

Line 227 in a744b41

x_conv = self.shared_nonlocal[i](x_conv)

Fc reg and fc cls will shared the other head, e.g.

doubleheadsrcnn/maskrcnn_benchmark/modeling/roi_heads/box_head/roi_box_feature_extractors.py

Line 245 in a744b41

identity = F.relu(self.fc6(identity))

The corresponding predictor is at
https://github.com/wuyuebupt/doubleheadsrcnn/blob/master/maskrcnn_benchmark/modeling/roi_heads/box_head/roi_box_predictors.py#L122

The reg and cls losses are also different, e.g. reg uses L1 loss, cls uses cross-entropy loss. I am not sure if those answered your question. But let me know if you need more info.