How to extract cam during training without creating issues to the bakpropagation

[PietroManganelliConforti]

Hi,

I'm implementing the following custom function:

def return_cam_from_model(model, target_layer, batch, targets , cam_name = "gradcam"):

model.eval() # da rimettere in .train() se nel training loop

cam = None

if cam_name == "gradcam":
    cam = GradCAM(model=model, target_layers=target_layer, use_cuda=False)

elif cam_name == "scorecam":
    cam = ScoreCAM(model=model, target_layers=target_layer, use_cuda=True)

elif cam_name == "fullcam":
    cam = FullGrad(model=model, target_layers=target_layer, use_cuda=True)

elif cam_name == 'gradcamplspls':
    cam = GradCAMPlusPlus(model=model, target_layers=target_layer, use_cuda=True)

elif cam_name == 'xgradcam':
    cam = XGradCAM(model=model, target_layers=target_layer, use_cuda=True)
    
elif cam_name == 'eigencam':
    cam = EigenCAM(model=model, target_layers=target_layer, use_cuda=True)
 
elif cam_name == 'eigengradcam':
    cam = EigenGradCAM(model=model, target_layers=target_layer, use_cuda=True)
    
elif cam_name == 'layercam':
    cam = LayerCAM(model=model, target_layers=target_layer, use_cuda=True)

elif cam_name == 'fullgrad':
    cam = FullGrad(model=model, target_layers=target_layer, use_cuda=True)
 
elif cam_name == 'hirescam':
    cam = HiResCAM(model=model, target_layers=target_layer, use_cuda=True)
    
elif cam_name == 'gradcamelementwise':
    cam = GradCAMElementWise(model=model, target_layers=target_layer, use_cuda=True)

elif cam_name == 'NOXAI':
    return 0

else:
    raise Exception("Cam name", cam_name, "not recognized")

targets = [ClassifierOutputTarget(i) for i in targets]


ret_cam = cam(input_tensor=batch, targets=targets)

model.train()

return ret_cam

we have seen that model.train() before the CAM execution produces a wrong cam explanation, so we inverted their order.

Therefore, because the cam execution is affected by the model training, do you know if inside the cam functions there is something that may affect the model training/backward? And how to deal with it?

[PietroManganelliConforti]

I'm trying to get tensors with gradient after the library execution

[m-aryayi]

I'm trying to get tensors with gradient after the library execution

Hi, What exactly do you mean by that?