How to properly train SPTNet

[Iranb]

Thank you for your cool work on GCD.

I ran the training script using the script in the Readme and tried to train a model based on DINO pretraining on the CUB dataset, but it seems that there are issues with the results.

CUDA_VISIBLE_DEVICES=0 python train_spt.py \
    --dataset_name 'CUB' \
    --batch_size 128 \
    --grad_from_block 11 \
    --epochs 1000 \
    --num_workers 8 \
    --use_ssb_splits \
    --sup_weight 0.35 \
    --weight_decay 5e-4 \
    --transform 'imagenet' \
    --lr 1 \
    --lr2 0.05 \
    --prompt_size 1 \
    --freq_rep_learn 20 \
    --pretrained_model_path ./pretrained/dino_vitbase16_pretrain.pth \
    --prompt_type 'all' \
    --eval_funcs 'v2' \
    --warmup_teacher_temp 0.07 \
    --teacher_temp 0.04 \
    --warmup_teacher_temp_epochs 10 \
    --memax_weight 1 \
    --model_path ./model_save

Here is the results.txt , which records the accuracy changes of each epoch during the training of 1000 epochs.

result.txt

What parameters do I need to modify to reproduce the results in the paper?

I look forward to your response and would like to thank you once again for your great work !

[whj363636]

Hi @Iranb

Thank you for your interest! I have reviewed the results you provided. I kindly request you to take note of the reminder mentioned below the training scripts. Our model is designed to enhance the compatibility with GCD by adjusting both the model parameters and prompt parameters. If you wish to reproduce the results mentioned in the paper, it is necessary to obtain the pretrained model and replace 'pretrained_model_path' with the SimGCD pretrained model, as SimGCD was the model we utilized in the paper. However, please note that our method can be applied to any other pretrained model, and SimGCD is just one of the optional choices. You can further replace the pretrained model with any other model, but need to carefully select the hyperparameters and training scheme.

[Iranb]

Hi @Iranb

Thank you for your interest! I have reviewed the results you provided. I kindly request you to take note of the reminder mentioned below the training scripts. Our model is designed to enhance the compatibility with GCD by adjusting both the model parameters and prompt parameters. If you wish to reproduce the results mentioned in the paper, it is necessary to obtain the pretrained model and replace 'pretrained_model_path' with the SimGCD pretrained model, as SimGCD was the model we utilized in the paper. However, please note that our method can be applied to any other pretrained model, and SimGCD is just one of the optional choices. You can further replace the pretrained model with any other model, but need to carefully select the hyperparameters and training scheme.

Thank you for your response. Through the replication experiments of SimGCD before, I did find that the current method's loss design is sensitive to hyperparameters. I will retry the SPTNet based on the pretrained SimGCD weights. 😊

[whj363636]

Sure. The default setting of the SimGCD paper should be sufficient to obtain a good enough pretrained model (for CUB, be aware of the difference of hyperparameters between their latest versions and previous ones), and proceed with building our SPTNet on top of it. Please feel free to reopen this issue if you encounter any difficulties.