I am trying to use gradient descent to iteratively update an image such that the end result is an adversarial example

```
from torch.autograd import Variable
import torch
from torch import optim
import torch.nn as nn
import torchvision
from torchvision import models, transforms
import matplotlib.pyplot as plt
import numpy as np
import cv2
model = models.inception_v3(pretrained=True)
model.eval().cuda()
img = plt.imread("hotdog.jpg")
img = cv2.resize(img,(299,299))
def preprocess(x):
means = [0.485, 0.456, 0.406]
stds = [0.229, 0.224, 0.225]
x = x / 255.0
x = (x-means)/stds
x=x.transpose(2,0,1)
x = torch.from_numpy(x).float()
x.unsqueeze_(0)
device = torch.device('cuda')
#x = Variable(x, requires_grad=True, device='cuda')
x = x.to(device)
return x
def postprocess(x):
x = x.cpu().numpy()
x=np.squeeze(x)
x=x.transpose(1,2,0)
means = [0.485, 0.456, 0.406]
stds = [0.229, 0.224, 0.225]
x = x * stds + means
x = np.round(x * 255)
return x
def attack(image):
loss = nn.CrossEntropyLoss().cuda()
target = Variable(torch.from_numpy(np.asarray([200])).cuda().long(), requires_grad=False)
for i in range(100):
input_img = preprocess(image)
optimizer = optim.Adam([input_img], lr=0.01)
logits = model(input_img)
_loss = loss(logits, target)
model.zero_grad()
_loss.backward()
optimizer.step()
image = postprocess(input_img)
preds = nn.functional.softmax(model(input_img))
adv_pred = preds[:, 200]
print(adv_pred.cpu().data.numpy())
attack(img)
```

adv_pred does not get optimized and stays at the same value throughout training. Why is this? Thank you for any help.