search

dvlab-research / JigsawClustering

This is the code for CVPR 2021 oral paper: Jigsaw Clustering for Unsupervised Visual Representation Learning
76 stars 6 forks source link

It seems that the model has not learned anything，What should I do？ #11

Open zbw0329 opened 2 years ago

zbw0329 commented 2 years ago

Thanks for your excellent work! I change the dataloader to use JigClu in CIFAR-10,and train the model on it by 1000epoch. But the prediction of my model is all the same. It seem that model always cluster into the same cluster

zbw0329 commented 2 years ago

When I use linear_evaluation to evaluate my model,should I train on main_lincls.py first?

zbw0329 commented 2 years ago

1637381240(1)

akuxcw commented 2 years ago

I think you only need to change the dataloader and augmentation to CIFAR-10 and change the num_classes=10. And check if you can load the pretrained weights correctly.

zbw0329 commented 2 years ago

What is the right order of training? Use main.py first to get checkpoint? And than use the main_lincls.py to evaluate the checkpoint of main.py? Or use the checkpoint as resume and train on main_lincls.py to get linear_checkpoint?

zbw0329 commented 2 years ago

I get 37 ACC1 on CIFAR-10 dataset after 300 epochs,which is far from the result of your paper. I use the learning rate of 10,should I make it smaller? 1637721732(1)

zbw0329 commented 2 years ago

I'm confused about the order in which main.py and main_lincls.py are used

akuxcw commented 2 years ago

You should, 1. use main.py to get the checkpoint; 2. use main_lincls.py to load the checkpoint as pretrained weights (not resume training).

The results of CIFAR-10 in the paper are produced using ImageNet pretrained weights. I didn't try directly pretraining on CIFAR-10.

Actually, 37% precision of linear evaluation shows the model's weight is not random. It seems the model learns some features, but not that good. The reason may lie in inappropriate hyper-parameters. Or maybe the CIFAR-10 dataset is too easy to learn, which makes the model's outputs became unchanged during the training process.

zbw0329 commented 2 years ago

Oh,I see. I use the main_lincls.py to get the checkpoint and than use it to evaluate. I will retry with the order you offer,thanks for your help.

zbw0329 commented 2 years ago

In the Table 5 of your paper,what is the different between 'finetune' and 'linear'? 1637805959(1) Is there any difference in their experimental process？ Are their assessment methods different?

akuxcw commented 2 years ago

In "linear", we load the pretrained weights and fix the backbone, then only train a classifier. In "fine-tune" we load the pretrained weights as initialization and train the whole model normally. They are two different ways to measure the quality of pretrained weights.