GRAD-CAM on YoloV3

Check dependencies.

See file dependencies.txt. To install: pip install -r dependencies.txt.

• python 2.7
• torch 0.4.1
• torchvision 0.2.1
• opencv-python 3.4.3.18
• click 6.7
• numpy 1.15.4
• pillow 5.3.0

Usage

Visualization

Single Image Visualization

e.g: If you wish to do visualization of "HiveAIRound1_vid_18_frame_1068.jpg" on our YoloV3 model with only 1 class and GPU mode off, type:

python3 main_single_img.py -i "./sample_data/HiveAIRound2_vid_37_frame_404.jpg" --no-cuda -a yolov3

YoloV3 tiny sample

python3 main_single_img.py -i "./sample_data/HiveAIRound2_vid_37_frame_404.jpg" --no-cuda -a yolov3_tiny

e.g: If you wish to do visualization of "cat_dog.png" on resnet152 model with 3 classes and GPU mode off, type:

python3 main_single_img.py -i "./sample_data/cat_dog.png" -a resnet152 -n 3 --no-cuda

Folder Visualization

e.g: If you wish to do visualization of one whole folder's(for example "./data/HiveAIRound0/") all .jpg images, with GPU mode on, type:

python3 main_folder.py -i "./data/HiveAIRound0/" --cuda

Mass Visualization

In order to solve the GPU/CPU out of memory error when applying for loop inside python script, I wrote a BATCH script "mass_main" that loops all images in dataset and do single image visualization one at time. Fortunately this works pretty well! Windows user can just execute:

mass_main.bat