Some questions about the model structure and the reappearance of the results. about dg-net

[layumi]

Hi @Liu-1994

1. It is due to that I added the non-local layer to the model before the paper submission. But we did not use it. When we decided to release our code, we simplified the code and removed the non-local layers which takes extra GPU memory. You may try to load the model by 'strict=False'.

2. Did you use fp16 or anything else? fp16 will lead to performance drop 1 percent.

[Liu-1994]

@layumi Thank you for your reply.
I hava understood the structure of the provided model.
About the mAP, I think I do not use fp16 as the config is apex: false. and I have not installed the NVIDIA/apex.
I have a small problem. As the provided trained model have weights about non_local layers, would the structure of the model affect the mAP?
By the way, what version of pytorch did you use? According to Pytorch's official documentation, PyTorch 1.1.0 and later versions adjust the order of lr_scheduler.step() and optimizer.step(), which may affect the reappearance of the results.

[layumi]

@Liu-1994
No. It will not affect the performance, since I did not include the non-local layer in the forward function.

I might utilize the pytorch 1.0.0 before the paper submission.
By the way, how many GPU do you use? It may affect the performance.

[Liu-1994]

@layumi Thanks for your reply.
I should only use a GPU, as the gpu-ids is only a number 0.