swuxyj

Same problem.

> I personally solved it by deleting the line `conda-forge` inside the `.condarc` config file and by not using conda-forge while installing packages/new envs. Hope it helps someone. it works...

I get mAP 0.844 when use ` optimizer = torch.optim.RMSprop(params_list, lr = 1e-5, weight_decay=10 ** -5) `

> > May I know your specific parameters in NUS_ Wide？I use your library's parameters to run out the result is only 0.8188 using resnet model and 64 bit hash...

> > > > May I know your specific parameters in NUS_ Wide？I use your library's parameters to run out the result is only 0.8188 using resnet model and 64...

> @wujiyang Hi could you please share us the .list file along with your CASIA-WebFace dataset ? Maybe you can try try: ``` import os with open('webface_align_train.list', 'w') as f:...

> Hi, swuxyj. Nice work for this community. It is noted that the training loss of DTSH contains a for loop which is somewhat time-consuming. Is there any change to...

可能是我这边实现的问题，可以参考下原作者的代码

You can refer to these two codes to divide the dataset https://github.com/swuxyj/DeepHash-pytorch/blob/master/data/voc2012/code.py https://github.com/swuxyj/DeepHash-pytorch/blob/master/data/mirflickr/code.py

也是用的torch.nn.CrossEntropyLoss()，这两个数据集用相同的流程也能跑出一个结果出来