stigma0617 / vovnet.pytorch Goto Github PK

I think the resnet50 baseline from torchvision (23.85% top-1 error) is trained for 100 epochs instead of 90.

Hi,

I'm trying to test out the pre-trained model and am having issues downloading the state dictionary with the "main.py" script. Here is the error I'm getting:

line 461, in load_state_dict_from_url
hash_prefix = HASH_REGEX.search(filename).group(1)
AttributeError: 'NoneType' object has no attribute 'group'

I've tried updating PyTorch and Torchvision to the latest releases, but got the same error. I then downloaded the state dictionary manually from the provided URL and using torch.load to load it into the variable "state_dict", but ran into more issues when trying to load that with the model. I've been using the following command to run the script:

python main.py [train and validation folders] --arch vovnet39 --pretrained --savedir VoVNet39

Thanks,
Bryan

Thank you for your great works!

I have a question about the variable identity_feat.

Why do you add the identity_feat to last feature maps in the OSA stage?

I couldn't find contents like this implementation in the VoVNet paper.

Is this just your method to help training the model like RESNET or official implementation?

I use the vovnet39 without pretrained weights , the result show vovnet is fast and powerful.so I want pretrained model for transfer learning, but the Link that you provide is not useful,can you upload the pretrained model.thanks.

Because vovnet.py does not import load_state_dict_from_url,
"the state_dict = load_state_dict_from_url(model_urls[arch], progress=progress)" is not valid in Line 185-186.
I guess vovnet.py should import a module such as utils.

I want to use this code. What is the this repository license?

Your pretraining model is invalid. Where else can I get it?
Thank you very much

Hi

How do you time the models, and what batch size and input image size do you use ?
When I tried to time various models with 224x224 input on different GPUs, the timings I see differ vastly from those reported in the repo here.
TitanXp: Res50(15.5ms B1, 24.4ms B16), VoVNet39(13ms B1, 24.5ms B16)
TitanX: Res50(15.2ms B1, 42.2ms B16), VoVNet39(12.6ms B1, 43.9ms B16)