version3

Simple instructions

Support ODAM and D-RISE.

Detector:

• Mask-RCNN: mask_rcnn

• YOLO V3: yolo_v3

• SSD: ssd

• FCOS: fcos

1. Environment:

Please install mmdetection version 3 first.

2. For ODAM:

2.1 get .npy format saliency maps:

python traditional_ODAM \
    --Datasets datasets/coco/val2017 \
    --eval-list datasets/coco_ssd_correct.json \
    --detector ssd \
    --save-dir ./baseline_results/tradition-detector-coco-correctly/

2.2 compute faithfulness metric:

python traditional_inference.py \
    --Datasets datasets/coco/val2017 \
    --detector ssd \
    --eval-dir ./baseline_results/tradition-detector-coco-correctly/ssd-ODAM/

2.3 visualization:

python visualize_baseline.py \
    --explanation-dir baseline_results/tradition-detector-coco-correctly/ssd-ODAM \
    --Datasets datasets/coco/val2017

3. For Grad-CAM:

Just modify ODAM file:

ODAM:

grad = feature.grad    # torch.Size([1, 512, 9, 19])
                
# grad = grad.mean([-1,-2], keepdim=True) # torch.Size([1, 512, 1, 1])
cam_map = F.relu_((grad * feature).sum(1)) # torch.Size([1, 9, 19])

Grad-CAM:

grad = feature.grad    # torch.Size([1, 512, 9, 19])
                
grad = grad.mean([-1,-2], keepdim=True) # torch.Size([1, 512, 1, 1])
cam_map = F.relu_((grad * feature).sum(1)) # torch.Size([1, 9, 19])

4. For D-RISE:

2.1 get .npy format saliency maps:

python traditional_DRISE \
    --Datasets datasets/coco/val2017 \
    --eval-list datasets/coco_ssd_correct.json \
    --detector ssd \
    --save-dir ./baseline_results/tradition-detector-coco-correctly/

Same instruction 2.2 and 2.3 as ODAM