README.md

EVA-X: Grad CAMs of Finetuned Models

Installation

1. Please follow classification installation.

2. Install amazing repo of pytorch-grad-cam

   pip install grad-cam
   

Weights Preparing

Get finetuned weights from EVA-X classification. Here, we only use models finetuned on Chest X-Ray14 dataset.

Run

Search for best thrd for final results. All of training codes are saved in get_cams.sh.

python get_cams.py \
    --model eva_x_small \
    --ckpt /path/to/eva_x_small_patch16_merged520k_mim_cxr14_ft.pth \
    --save-path ./output/eva_grad_cam \
    --searching

Visualize Grad-CAMs

python get_cams.py \
    --model eva_x_small \
    --ckpt /path/to/eva_x_small_patch16_merged520k_mim_cxr14_ft.pth \
    --save-path ./output/eva_grad_cam \
    --thrd 0.4

The results will be saved into ./output

Results

Methods	Architecture	AP25	AP50	mIoU	mAP	Checkpoint
MoCov2	DenseNet121	10.98	1.42	8.18	6.20	🤗download
Medical MAE	DenseNet121	8.33	2.13	6.90	5.23	🤗download
BioViL	ResNet50	11.59	3.46	7.29	7.52	🤗download
MGCA	ResNet50	7.22	2.34	5.00	4.78	🤗download
MedKLIP	ResNet50	5.69	2.74	4.94	4.22	🤗download
Medical MAE	ViT-S	7.11	1.00	6.79	3.61	official download
EVA-X (Ours)	ViT-S	14.74	2.95	9.18	8.84	🤗download