Hi, this is a documentation of my use of the Crowd Counting Computer Vision project, provided by @Leeyeehoo and directed by AnalyticsVidhya's article. This documentation is useful for following reasons:
- It has already corrected @Leeyeehoo's project according to AnalyticsVidhya's article and also where AnalyticsVidhya missed.
- It is corrected based on the Shanghaitech Dataset that you can get from Kaggle, which is slightly different from that used by AnalyticsVidhya's article nor @Leeyeehoo.
- You can easily change your root directory and .json files by using the converter.py that I provided.
- I am using a laptop with RTX3050, and it took my laptop about 40 hours to train 360~epochs.
- The MAE after validation gave me about 69 MAE, better than 75 MAE given by AnalyticsVidhya.
- After about 280ish epochs, the loss barely decreases. Sharp reduction of loss was seen after 30-50 epochs.
- I think the loss was within 30 people (even with the outliers) when it reached 150 epochs. That could be good enough.
- After 360 epochs, the loss is within 1 digit, but sometimes outliers could reach about 15 people.
to run training in terminal vsc terminal cd CSRNet-pytorch python train.py part_B_train.json part_B_val.json 0 0