I came into sth that really strange, this is my code

```
def model_eval(model):
model_the_wrong_images = []
model_the_real_label = []
model_the_pred = []
correct = 0
total = 0
with torch.no_grad():
model.eval()
for idx, test in enumerate(test_loader):
image, label = test
total += image.size()[0]
pred = torch.argmax(model(image), 1)
correct += (pred == label).sum().item()
if idx == 4:
#print(model(image))
print(torch.argmax(model(image), 1))
print(pred)
print(torch.argmax(model(image), 1))
print(pred == torch.argmax(model(image), 1))
print(correct)
print(label)
for i in range(len(image)):
if not (pred == label)[i].item():
model_the_wrong_images.append(image[i])
model_the_real_label.append(label[i])
model_the_pred.append(torch.argmax(model(image[i].unsqueeze_(0)), 1))
print(correct)
print(total)
print('The accuracy on testset is : %f'%(correct / total))
return model_the_wrong_images, model_the_real_label, model_the_pred
```

I write this func to eval my model, but i found that once i repeat the eval process with the model that has already been trained, the accuracy changes everytime.

**Why my model’s params can’t keep const??? I wrote that model.eval(), but it seems no work**

So i print the output and found the output change.Go futrher, i print the output of the 4th subset of the test process to see, something terrible happen, the outputs of the `if`

are:

```
tensor([9, 7, 2, 4])
tensor([9, 7, 3, 4])
tensor([9, 7, 3, 4])
tensor([1, 1, 1, 1], dtype=torch.uint8)
17
tensor([9, 7, 3, 4])
```

How could the `pred`

and `torch.argmax(model(image), 1)`

become two different things?I get pred by `pred = torch.argmax(model(image), 1)`

!

Who can tell me why! Thanks a lot!!!