TODO

introduce model

KDDF

How to get the code

git clone https://github.com/melamaze/KDDF-pytorch-grad-cam

How to dowload testing dataset

You can get testing dataset in this link.

Decompress dataset by:

unzip V1.zip  
unzip V2.zip

How to excute KDDF

# example: excute resnet18, V1, layercam 
python3 cam_resnet.py --method layercam 
# write into a log file 
python3 cam_resnet.py --method layercam > log
# use cuda
python3 cam_resnet.py --method layercam --use-cuda

Dataset V2

If you want to run model trained by dataset V2, remember to revise num_classes.

Take resnet18 for example, you need to change package/FL/resnet.py line: 72 into:

def __init__(self, block, num_blocks, num_classes=35):

CUDA

Some of the model are trained by CUDA: 1, if you only have cuda: 0, you may get error when loading model.

Take cam_V2.py line: 93 for example, you need to add map_location='cuda:0':

global_model.load_state_dict(torch.load(PATH, map_location='cuda:0'))

Models

file	trigger size	addversry ratio	dataset	trigger position
regnet_15_0.2_V1_start.pth	15%	0.2	V1	start
regnet_15_0.3_V1_1-4.pth	15%	0.3	V1	1/4
regnet_15_0.3_V1_3-4.pth	15%	0.3	V1	3/4
regnet_15_0.3_V1_mid.pth	15%	0.3	V1	mid
regnet_15_0.3_V1_start.pth	15%	0.3	V1	start
regnet_V1_clean.pth	-	0.0	V1	-
resnet_5_0.2_V1_start.pth	5%	0.2	V1	start
resnet_10_0.2_V1_start.pth	10%	0.2	V1	start
resnet_15_0.2_V1_start.pth	15%	0.2	V1	start
resnet_15_0.2_V2_start.pth	15%	0.2	V2	start
resnet_15_0.3_V1_start.pth	15%	0.3	V1	start
resnet_15_0.4_V1_start.pth	15%	0.4	V1	start
resnet_25_0.2_V1_start.pth	25%	0.2	V1	start
resnet_50_0.2_V1_start.pth	50%	0.2	V1	start
resnet_75_0.2_V1_start.pth	75%	0.2	V1	start
resnet_V1_clean.pth	-	0.0	V1	-
resnet_V2_clean.pth	-	0.0	V2	-
resnext_15_0.2_V1_start.pth	15%	0.2	V1	start
resnext_V1_clean.pth	-	0.0	V1	-