General

• Implementation crowd-counting to traffic domain and counting vehicle on road
• paper reference: CSRNet: Dilated Convolutional Neural Networks for Understanding the Highly Congested Scenes

Requirements

• torch
• torchvision
• pillow
• opencv-python
• numpy
• GPU-CUDA (optional)

Training

python3 train.py --train <path to folder training set> -cfg model.json --use_pretrain --density <path to dataset density> --cuda -i 200 -lr 0.000001 -wk 3 -bs 4

--cuda: flag to use GPU NVIDIA to training phase. If you don't want use GPUs just skip it.

-wk:(worker) multi-processor to train, it make training phase more faster( if you got strong cpu like i5, ryzen5, xeon, ...)

-bs: batch-size

-lr: learning rate, make it small enough to training, can you monitor with tensorboard and watch total gradient graph it must be fluctuate

Load from checkpoint and continue to train:

python3 train.py --train <path to folder training set> -cfg model.json --density <path to dataset density> --cuda -i 200 -lr 0.000001 -wk 3 -bs 4 --checkpoint

Validation

python3 eval.py --test <path to folder training set> -cfg model.json --density <path to eval density> --cuda -wk 3 -bs 4

Test with video

python3 video_test.py --input <path to video> -cfg model.json --cuda -l ./logs