The grad of vgg is used as the activation value to display image objects.

Import torchimport numpyimport torch. nn as nnimport torch. nn. functional as ffrom Pil import imagefrom torchvision import transformsimport torchvision. models as modelsnormalize = transforms. normalize (mean = [0.485, 0.456, 0.406], # This is imagenet STD = [0.229, 0.224, 0.225]) TRAN = transforms. compose ([transforms. resize (224,224), transforms. totensor (), transforms. normalize (mean = [0.485, 0.456, 0.406], STD = [0.22 9, 0.224, 0.225]) Im = '. /1.jpeg '# im = '. /2.jpg 'im = image. open (IM) Im = Tran (IM) im. unsqueeze _ (DIM = 0) Im = torch. autograd. variable (IM, requires_grad = true) vgg = models. vgg16 () Pre = torch. load ('/home/qk /. torch/models/vgg16-397923af.pth ') vgg. load_state_dict (pre) out = vgg (IM) outnp = out. data [0] ind = int (numpy. argmax (outnp) out [0] [ind]. backward () grad = IM. gradgrad. squeeze _ (0) grad = grad * gradgrad = grad. sum (keepdim = false, dim = 0) Grad = torch. SQRT (GRAD) Rg = torch. max (GRAD) grad = grad/rg * 255. # It's amazing. Why can we activate the image with uint8 !!!!! If I don't search for it, how can I debug it? # Im = image. fromarray (grad. numpy (), 'L ')#. eval () tensor-> numpy arrayim = image. fromarray (numpy. uint8 (grad. numpy (), 'L ')#. eval () tensor-> numpy arrayim.save('grey.png ') # input () from CLS import dprint (d [ind])

 

The grad of vgg is used as the activation value to display image objects.