CameraTraps

Inference:

Install and activate environment, then run megadetector, follow the instrutions on CameraTraps github repo. This is the instruction to run megadetector

python detection/run_detector.py "path/to/md_v5a.0.0.pt" --image_file "path/to/img.jpg" --threshold 0.2 

Training:

Data Preprocessing

Split data into train and test set: 1st answer in the post

Data format conversion with this code: Coco to yolo

Launch training

Single GPU: yolo tutorial Chapter Single GPU:

python train.py --data your_new_training_data.yaml --weights path/to/md_v5a.0.0.pt

Multi-GPU single node: yolo tutorial Chapter Multi-GPU DistributedDataParallel Mode: pass

python -m torch.distributed.launch --nproc_per_node

followed by the usual arguments.

Multi-node: Use mpiexec.hydra in slurm to launch a bash script. Launch torch.distributed.launch in bash script:

mpiexec.hydra -np $NNODES -ppn 1 path/to/yolov5/train.sh \
--ngpus $GPU_PER_NODE --nnodes $NNODES --master $MAIN_RANK (in slurm)

python -m torch.distributed.launch --nnodes=$NNODES \
--nproc_per_node=$NGPUS --node_rank=$LOCAL_RANK \
--master_addr=$MASTER path/to/train.py –arguments (in train.sh)

Scripts I used to do data preprocess and training launching are in this repo

Training time: 300 epochs completed in 17.4 hours. If only validate final epoch, training time is 15.5 hours for 300 epochs. 2 nodes, 8 gpus, batch size 32, imgzsz 1280

Training statistics can be found in results.csv