search

wangchen1801 / FPD

Official code of the paper "Fine-Grained Prototypes Distillation for Few-Shot Object Detection (AAAI 2024)"
https://arxiv.org/pdf/2401.07629.pdf
27 stars 4 forks source link

code reproduction process #6

Closed wml666666 closed 4 months ago

wml666666 commented 4 months ago

I want to continue working on your work, but I encountered some problems during the code reproduction process! In fpd_r101_c4_2xb4_voc-split_10shot fine-tuning, experiments were conducted according to the code you provided. The log file is shown in 10shot. log, and the optimal accuracy NOVEL-CLASSESPLIT1 mAP: 0.672 was achieved in 2400 iterations of testing results. However, upon observing the log file you provided, only 2000 iterations were performed, achieving an accuracy of 0.684. What is the main reason for the difference in experimental results? 10shot.log

wangchen1801 commented 4 months ago

Hello, it's glad to hear that you are interest in continuing this work. In the fine-tuning stage, we train more layers (NLF, FFA) compared with other methods (only the last layer). It's a unstable process due to the limited training samples. Please try to evaluate more frequently in case the model is overfitting. Besides, base training can also affect the fine-tuning performance, you can use the pretrained weights or retrain the base model. It may take a few more experiments to get the reported performance.

wml666666 commented 4 months ago

Okay, thank you for your prompt reply!