About the accuracy of fine-tuning on the target fine-grained datasetes
eafn opened this issue · comments
Hi, I have read your paper which is a great work. But I have a question, how is the performance of fine-tuning the entire ResNet50 (both encoder and classifier) on these target fine-grained datasetes with supervised learning? Do you have the result?
Hi and thanks for your good question!
Yes, in Table 7b, we summarized end-to-end supervised fine-tuning results with SSL pretrained ResNet50 on four fine-grained datasets.
But, we only reported not fully-labeled scenarios as in previous SSL works.
Here, we additionally summarize e2e fine-tuning results on fully-labeled (100%) target datasets:
Pretrain | Aircraft | Cars | Pet | Birds |
---|---|---|---|---|
X | 79.89 | 88.63 | 78.24 | 67.46 |
OS | 81.28 | 87.43 | 85.01 | 70.12 |
SimCore | 83.28 | 89.54 | 85.95 | 71.24 |
cf) Note that we fine-tuned models for 100 epochs with a momentum SGD optimizer and weight decay of 1e-4.
We searched the optimal learning rate among three logarithmically spaced values from 1e-1 to 1e-2 (i.e., {1e-1, 3e-2, 1e-2}), and they are decayed after 60 and 80 epochs by the ratio of 0.1.