Using pytorch >=1.5.0, I have compiled maskrcnn-benchmark. But when running siam-mot, it can't import DFConv2d

[lihanlin99]

I have solved the problem, thank you!

I am using python3.6, pytorch 1.7.0 and cuda 11.1.

According to the INSTALL.md, I have run the commands below to change the two files, "/deform_conv_cuda.cu" and "/deform_pool_cuda.cu".

cuda_dir="maskrcnn_benchmark/csrc/cuda"
perl -i -pe 's/AT_CHECK/TORCH_CHECK/' $cuda_dir/deform_pool_cuda.cu $cuda_dir/deform_conv_cuda.cu

Then, the maskrcnn_benchmark can be compiled without any error. But when running siam-mot, it reports the error following:
Traceback (most recent call last):
File "tools/test_net.py", line 13, in
from siammot.modelling.rcnn import build_siammot
File "/root/siam-mot/siammot/modelling/rcnn.py", line 12, in
from .backbone.backbone_ext import build_backbone
File "/root/siam-mot/siammot/modelling/backbone/backbone_ext.py", line 8, in
from .dla import dla
File "/root/siam-mot/siammot/modelling/backbone/dla.py", line 8, in
from maskrcnn_benchmark.layers import DFConv2d
ImportError: cannot import name 'DFConv2d'

I have checked the code of the maskrcnn-benchmark package that has been installed in /opt/conda/lib/python3.6/site-pakages. These is no DFConv2d.