Open Galaxy-PC opened 4 months ago
In addition, what's mean about "model.load_state_dict(torch.load("/root/autodl-tmp/weights/Modern_1652_2023-06-23-03:45:26/net_059.pth"))" , how to get net_059.pth? Thanks.
Please tell me the specific error message. If the training issue is due to the line model.load_state_dict(torch.load("/root/autodl-tmp/weights/Modern_1652_2023-06-23-03:45:26/net_059.pth")), you can simply comment out this line as it does not affect obtaining the MEBG model.
If you want to obtain MEBG weights, please refer to the backbone and MEBG correspondence table in Table 1 of the paper. Also, check the Hugging Face website at https://huggingface.co/timm/eva02_large_patch14_448.mim_m38m_ft_in22k_in1k for instructions on how to use EVA's initial weights. Modify the weights loaded by timm in lines 45 and 50 of model_.py and retrain the model.
感谢!就只是单纯的使用EVA模型,但是训练出来的模型效果并没有您在论文展示的那么好,我在University-1652k测试只得到50多scores的Recall@1,不知道是不是还有其他设置?例如超参数?
麻烦重新下载model_.py的代码后再尝试训练一下,我在前两天修复了一个bug。您也可以通过下载训练权重(Baidu Cloud Disk Link: https://pan.baidu.com/s/1k1z90EyLaL85PqeSxxlWOw?pwd=1652 提取码: 1652)复现论文中的结果。
Hello, thanks for your open source work. I would like to ask whether MBEG-L1.pth is the model file obtained directly through training train.py? Running train.py directly and setting lr=0.1 doesn't work very well.