Using this network in 2020, I found an error when importing model weights. I will not attach a trace, because it is very large, it can be found in the corresponding discussion at this repository. It referred to the network architecture and was probably caused by package updates (as I understood, it had been several years since the network was written). To fix this, I had to add the track_running_stats=True parameter in the Bottleneck, UpBottleneck, Net classes to norm_layer, which is called directly (and not passed as an argument). This was enough to fix the code, I didn't change anything else in the code itself, the architecture was fixed in the attached file.
Using this network in 2020, I found an error when importing model weights. I will not attach a trace, because it is very large, it can be found in the corresponding discussion at this repository. It referred to the network architecture and was probably caused by package updates (as I understood, it had been several years since the network was written). To fix this, I had to add the track_running_stats=True parameter in the Bottleneck, UpBottleneck, Net classes to norm_layer, which is called directly (and not passed as an argument). This was enough to fix the code, I didn't change anything else in the code itself, the architecture was fixed in the attached file.
You can find it here as well: https://pastebin.com/HNyL4uRK fixed_architecture.zip