MAE-pytorch
MAE-pytorch copied to clipboard
pengzhiliang
knn eval of MAE
I eval vit_base of 500/1600 pretraining on imagenet1000 using knn metric. By loading all the pretained parameter with vit GAP method (not need cls token), the knn 20-NN result is 33.4 in imagenet100 dataset, which is very low and not match the accuracy of linear prob.
I just tried your latest updata of end-to-end finetuning, it seems good. But I think linear prob still is a metric cannot avoided.
Thanks for you suggestions, we actually ignore the linear prob metric. In fact, I am not very familiar with Linear Prob. Can you help me try to implement it? Thank you very much!
https://github.com/facebookresearch/dino/blob/main/eval_linear.py dino contains the code of knn and linear eval code. I am not sure how to treat the cls token, as the linear prob only finetune the last head, but for MAE , the cls token is not pre-trained.
Hello, have you finished the end-to-end fine-tuning of vit-base/1600e? Can you tell me the result? Thank you!
Hi, I finished the epoch 1600 training, but I only got fine-tuning result of 83.15 for epoch 1400 and 82.97 for epoch 1600. which is lower than your reported epoch 400 and the paper results.
From your pretrained log of vit_base, I found your max learning rate is 0.0024, is you run with 128X32 batch size? according to the code: args.lr = args.lr * total_batch_size / 256, which should be 0.0006 for batchsize of 128X8.
Ok, that is very strange. I run vit-base with 512 x 8 = 4096, where the lr: 1.5e-4 * 512 * 8 / 256 = 0.0024.
ok, I will try your setting to reimplement your results for epoch 400. But the results of epoch 1600 is on batchsize 4096, still not good enough. the ft accuracy incrase slowly with epoch: 82.71/200, 82.82/400,82.87/600, 83/800,82.78/1000,82.96/1200,83.15/1400,82.97/1600.
OK, thank you for your so much experiments! Maybe there is still some problems, I will check it carefully.
@Dongshengjiang Have you tried the LinearProbe evaluation with cls token?
The paper said: As ViT has a class token [16], to adapt to this design, in our MAE pre-training we append an auxiliary dummy token to the encoder input. This token will be treated as the class token for training the classifier in linear probing and fine-tuning.
It seems that the author just adds a dummy token when pre-training, and directly uses it as the feature for linear probing.
