The impact of the result of the first stage("Train teacher model with available labeled data only").

[CStar-777]

Hi,
Excellent work. But I have a question that I hope you can answer.
Training is split into two consecutive steps in this project, And then, what is the impact of different training results in the first stage on the training in the second stage?
Such as I run the code twice:
python3 train_net.py --config-file ./configs/coco/instance-segmentation/maskformer2_R50_bs16_50ep.yaml --num-gpus 4 --num-machines 1 SSL.PERCENTAGE 5 SSL.TRAIN_SSL False OUTPUT_DIR out/TEACHER/COCO/ResNet50/5
The AP are 26.06(50k) and 25.56(40k) respectively. Will it have any impact on the final result if I use the “model_best.pt” of these two results separately to train the second stage("Train semi-supervised model using pretrained checkpoint").
Looking forward to your answer.

[CStar-777]

I think I should train the first stage until it stops on its own. I have another request. Can you provide the AP that the Teacher model can achieve after completing the first stage of training?

[TariqBerrada]

Hello,
The second stage seems relatively robust to the score achieved during the first stage, but can vary with the number of burnin iterations.
For the first stage of training, the AP results are reported in the paper.
During training, you can check the results of evaluation if you want to do early stopping.

[CStar-777]

Hello, The second stage seems relatively robust to the score achieved during the first stage, but can vary with the number of burnin iterations. For the first stage of training, the AP results are reported in the paper. During training, you can check the results of evaluation if you want to do early stopping.

Sorry, I couldn't find any records about the first stage AP results in the paper, except for "No Buin in" in Table 4. Could you please tell me in detail which page or table it is on? I apologize for any inconvenience caused by my carelessness.