-
Notifications
You must be signed in to change notification settings - Fork 85
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Pytorch with Resnet50 #37
Comments
Hi raymongL, I've also tried to adopted the PyTorch code to CUB200 dataset with finetuned ResNet50, but I can't make the loss converge. I've tried different optimizers like SGD, Adam, RMSprop, and different class_num values like 1.0 and 200.0, different lr values from 1e-5 to 1e-3. Here is a set of parameters which I tried:
But the training loss is always around 0.69, and mAP is extremely as low as 0.04. Could you please give me some hints or share your train script with me? Thanks. |
我用的RMSprop,lr=1e-5,Resnet最后一层fc后面接一个tanh输出网络特征。除此之外都跟源码一致。Class_num是200,为什么要改成1?没太懂。。 |
I think the original code has already tanh after fc, Do you use the original code?
I'm quite confused about class_num because in coco dataset it's 1.0 even there are 80 classes in coco. I've also perform an experiment with parameters that you mentioned above.
Thel loss is shocking during the interval from 0.3 to 0.5, here is a glance of loss:
MAP: 0.043476254169203095 But the final mAP is still very low. Would you mind to share your code on your GitHub? |
OK, can i have your email? |
Thanks |
Can anybody tell me the reason for low mAP? I am getting around 0.03. Any suggestion would be great. Thanks. |
I change the network to resnet50 in the pytorch version HashNet, and i can not produce acceptable map on CUB dataset. I have tried some fine tuning, and the best map i can get is arount 40 with 64 bits. SGD and Adam won't work, the result is given by RMSprop.
Any idea what's the problem and how can i fix it?
The text was updated successfully, but these errors were encountered: