Permute error number of dims don't match error

vision
1

While running inference on my models,I ran into the following error

Runtime error number of dims don't match in permute

This is what I was doing

    val_transforms = T.Compose([
        T.Resize((256, 256)),
        T.CenterCrop((224, 224)),
        T.ToTensor(),
        T.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
    ])
    img_name=cfg.TEST.IMAGE
    img=cv2.imread(img_name)
    img=cv2.cvtColor(img,cv2.COLOR_BGR2RGB)
    if (isinstance(img,np.ndarray)):
        img=Image.fromarray(img)
    
    plt.imshow(img)
    img=val_transforms(img)

The shape of img is [3,224,224]
Next I passed it to the model as follows

model=Model()#Model for my purpose transformer VIT small_patch_16-224
with torch.no_grad():
img=img.to(device).squeeze()
probs=model(img)
"""
apply softmax on probs to get result
"""
   ps = ps.cpu().data.numpy().squeeze()
   #ps is the output after applying softmax on probs 
    image = image.permute(1,2,0)#Runtime error generated at this point
    mean = torch.FloatTensor([0.485, 0.456, 0.406])
    std = torch.FloatTensor([0.229, 0.224, 0.225])
    
    
    image = image*std + mean
    img = np.clip(image,0,1)
    
    fig, (ax1, ax2) = plt.subplots(figsize=(8,12), ncols=2)
    ax1.imshow(img)

At first,I thought that because img has changed shape to [1,3,224,224] ,so that’s the reason why this is happening
So I did img.squeeze() before applying permute but that didn’t change the outcome.
So I changed my code as follows and that resolved the issue.
The new code

img_name=cfg.TEST.IMAGE
    img=cv2.imread(img_name)
    img=cv2.cvtColor(img,cv2.COLOR_BGR2RGB)
    if (isinstance(img,np.ndarray)):
        img=Image.fromarray(img)
    
    plt.imshow(img)
    img_1=val_transforms(img)
   output_prob=[]
   with torch.no_grad():
      prob=model(img_1.to(device).unsqueeze(0))
     output_prob.append(prob)
     output_prob=torch.cat(output_prob,dim=0)
     ps=nn.Softmax(output_prob,dim=1)

Any ideas why the second code version doesn’t raise the permute dimension mismatch error.
As far as I could understand both versions are the same

2

Can you print the shape of image? It’s most likely 2D, and hence permuting over 3 dimensions isn’t possible. If you’re using greyscale images, you’ll want to include a channel dimension when you reshape. So the shape goes from [256,256] -> [1,256,256].

3

Shape of the image is 224x224x3.
As shown in

T.CenterCrop((224, 224)),

The first version of the code gave the permute dimension mismatch error but the second one didn’t.

4

So if you print image.shape before this line ^, what is the shape? Just to check it is 3D and not somehow 2D.

5

Before that line the shape is (1,3,224,224).After applying squeeze the shape is (3,224,224).
I will check and revert back
image shape before permute

torch.Size([3, 224, 224])
6

not sure what is the error here. Because I ran the same code right now,and now it’s working.

7

Probably applying torch.squeeze resolved the shape issue which was causing your code to previously fail.

8

The squeeze operation gets rid of any single dimensions. E.g. [1,3,224,224].squeze() → [3,224,224]

So, you need to do .unsqueeze(0) before passing the img to your model and do .squeeze() before permute