Dimensionality mismatch help

1

Hey Guys,

I used Timm pre-trained efficientNet to train on a dataset. My test dataset has around 281 samples both 0 and 1 classes included.

Here is my model params:

<bound method Module.parameters of EfficientNet(
  (conv_stem): Conv2d(3, 32, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
  (bn1): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
  (act1): SiLU(inplace=True)
  (blocks): Sequential(
    (0): Sequential(
      (0): EdgeResidual(
        (conv_exp): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
        (bn1): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
        (act1): SiLU(inplace=True)
        (se): Identity()
        (conv_pwl): Conv2d(32, 32, kernel_size=(1, 1), stride=(1, 1), bias=False)
        (bn2): BatchNorm2d(32, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
      )
     ................... You can see the full model architetcure here as I am reaching the character limit for the post:
https://colab.research.google.com/drive/1rWaRA-jErgxecbD3jiXTeiHpzLafsiwE#scrollTo=xW-RaHSEnYMT
)>

I wanted to take my samples with best predictions in terms of accuracy and apply LIME on it. So, I wanted to produce individual predictions on each file. Then take the files with highest accuracy and apply LIME on it.

But when I try this I get a dimensionality error. Anyways to get across this with a workaround?

Here is the code I am getting the error on :slight_smile:

list = testloader.dataset.samples
import PIL
for i in range(len(testloader.dataset.samples)):
  sample_fname, _ = testloader.dataset.samples[i]
  print(sample_fname)
  img = sample_fname
  image = PIL. Image. open(img)
  model = efficientnetv2_model
  transform = transforms.Compose([transforms.ToTensor()])
  tensor = transform(image)
  print(tensor.shape)
  logits = model(tensor)

This is the error. My images are black and white images:
—> 12 logits = model(tensor)
RuntimeError: Given groups=1, weight of size [32, 3, 3, 3], expected input[1, 1, 128, 170] to have 3 channels, but got 1 channels instead

type or paste code here
2

Might be wrong, but the fact that they are black and white makes me suspect they are single channel images which is what the 1 channel message would seem to imply as well. Since you’re using a pretrained model that expects 3 channels, you could try either padding or expanding the tensor. Can try something like img_tensor = img_tensor.expand(-1, 3, -1, -1)

3

Hey thanks for the reply. What I don’t understand is the images always have been black and white. Am i missing something?

This is the original training file:

https://github.com/mvadrev/coviScan/blob/main/An_accelerated_COVID19_diagnosis_using_interepretable_deep_learning.ipynb
4

Looks like the colab is using ImageFolder for loading training data, and a search on stackoverflow seems to imply that this automatically converts grayscale images to RGB which would explain why loading them yourself would only have a single channel.

5

Gotcha so u saying using imageFolder auto converts gray scale to color by default. I will try to use imageFolder. Thanks

6

While it would probably work, you can just add expand like in the following which will be easier.

list = testloader.dataset.samples
import PIL
for i in range(len(testloader.dataset.samples)):
  sample_fname, _ = testloader.dataset.samples[i]
  print(sample_fname)
  img = sample_fname
  image = PIL. Image. open(img)
  model = efficientnetv2_model
  transform = transforms.Compose([transforms.ToTensor()])
  tensor = transform(image).expand(-1, 3, -1, -1)
  print(tensor.shape)
  logits = model(tensor)

Might just need .expand(3, -1, -1) not sure if that would have a batch dimension or not

7

I think this may work. Correct me if I am wrong. The idea I had was to simply duplicate the same sample across the batch of 32 so that I keep all the conditions of the training same during testing. So, I should probably have the same tensor over a batch of 32.

8

Yup. I haven’t used transforms.Compose before, but it looks like transforms.ToTensor() will return a tensor of shape CxHxW. So you’d want to use tensor = transform(image).unsqueeze(0).expand(32, 3, -1, -1) for the situation you’re describing if I’m understanding you correctly.

9

Thanks for your help. Here is what I did but getting a new error

list = testloader.dataset.samples
import PIL
for i in range(len(testloader.dataset.samples)):
  sample_fname, _ = testloader.dataset.samples[i]
  print(sample_fname)
  img = sample_fname
  image = PIL. Image. open(img)
  model = efficientnetv2_model
  transform = transforms.Compose([transforms.ToTensor()])
  tensor = transform(image).unsqueeze(0).expand(32, 3, -1, -1) 
  print(tensor.shape)
  logits = model(tensor)

→ 12 logits = model(tensor)

5 frames
/usr/local/lib/python3.7/dist-packages/torch/nn/modules/conv.py in _conv_forward(self, input, weight, bias)
442 _pair(0), self.dilation, self.groups)
443 return F.conv2d(input, weight, bias, self.stride,
→ 444 self.padding, self.dilation, self.groups)
445
446 def forward(self, input: Tensor) → Tensor:

RuntimeError: Input type (torch.FloatTensor) and weight type (torch.cuda.FloatTensor) should be the same or input should be a MKLDNN tensor and weight is a dense tensor

type or paste code here

So, what’s I need is to find the accuracy of a single sample instead of a batch. So, that I can pick my input samples for best accuracy for further analysis.

Anyway we can run the accuracy function on single sample? One idea I have is to create a folder for each image. > Copy the same image 32 times in the folder and then try to run the accuracy function.

ANy other ideas ?

10

It works. I forgot to move the Tensor to GPU :stuck_out_tongue:
Thanks for your help @carperbr Saved my day

11

Ok there is one mroe thing I wanted to ask. So, I am running my accuracy tests for each file for each class seperate. So, the labels will be same for every run i.e since I am running the code on positive folder only all labels are positive.

But here is the issue when I run the following

list = testloader.dataset.samples
import PIL
for i in range(len(testloader.dataset.samples)):
  sample_fname, _ = testloader.dataset.samples[i]
  print(sample_fname)
  img = sample_fname
  image = PIL. Image. open(img)
  model = efficientnetv2_model
  transform = transforms.Compose([transforms.ToTensor()])
  tensor = transform(image).unsqueeze(0).expand(32, 3, -1, -1).to(device)
  print(tensor.shape)
  logits = model(tensor)
  y_pred = F.softmax(logits,dim = 1)
  top_p,top_class = y_pred.topk(1,dim = 1)
  print("THe top", top_class, top_p)
  loss = criterion(logits,labels)
  print("The loss is", loss.item())
  print("The accuracy is", accuracy(logits,labels))

Here is the error I am getting

--> 16   loss = criterion(logits,labels)
     17   print("The loss is", loss.item())
     18   print("The accuracy is", accuracy(logits,labels))

2 frames
/usr/local/lib/python3.7/dist-packages/torch/nn/functional.py in cross_entropy(input, target, weight, size_average, ignore_index, reduce, reduction, label_smoothing)
   2994     if size_average is not None or reduce is not None:
   2995         reduction = _Reduction.legacy_get_string(size_average, reduce)
-> 2996     return torch._C._nn.cross_entropy_loss(input, target, weight, _Reduction.get_enum(reduction), ignore_index, label_smoothing)
   2997 
   2998 

TypeError: cross_entropy_loss(): argument 'target' (position 2) must be Tensor, not list
12

It looks like you need to convert your labels object into a tensor; from the message, it appears that it is a list currently.

13

Thanks. I got it working but I have a question. So, I am now able to do the following and send each image one by one. I used your suggestion to unsqueeze the tensor for the image. But how do we do the dimensionality correction for the label?

In my code below if I pass the newlabels instead of labels, the accuracy function works. But I need to give label[i] because I need to pass the correct label of the image only. Any idea how I can handle this?

# checking all labels as below zp.zeros of 32 is not correct
for j, (images, labels) in enumerate(testloader):
      print("Iterating file",j)
      print("The len of lables", len(labels))
      # print("The labels is", labels[j])
      images = images.to(device)
      print(labels)
      labels = torch.tensor(labels.to(device='cuda'))
      # sample_fname, _ = testloader.dataset.samples[j]
      # print(sample_fname)
      # print("Types of images", images[j])
      # print("$$$$$size$$$$$$$", images.shape[0])
      

      for i in range(images.shape[0]):
        # print("count", i)
        sample_fname, _ = testloader.dataset.samples[i]
      # Forward pass - compute outputs on input data using the model
        logits = efficientnetv2_model(images[i].unsqueeze(0).expand(32, 3, -1, -1).to(device='cuda'))

        y_pred = F.softmax(logits,dim = 1)
        top_p,top_class = y_pred.topk(1,dim = 1)
        y_pred_list.append(top_class.data.to("cpu").tolist())
        

        print("labels ********", labels[i])
        newLabels = torch.tensor([1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]).to(device='cuda')
        loss = criterion(logits,newLabels)
        print("The loss is", loss.item())
        print("The accuracy is", accuracy(logits,newLabels))
      print("####### End of batch #######")
14

I ended up doing this. I am using np.full to fill all 32 values with the same class. But my file names are not matching. I need to pick my files with highest accuracy. But its printing it incorrectly:

# checking all labels as below zp.zeros of 32 is not correct
for j, (images, labels) in enumerate(testloader):
      print("Iterating file",j)
      print("The len of lables", len(labels))
      # print("The labels is", labels[j])
      images = images.to(device)
      print("labels",labels.shape)
      labels = torch.tensor(labels.to(device='cuda'))
      # sample_fname, _ = testloader.dataset.samples[j]
      # print(sample_fname)     

      for i in range(images.shape[0]):
        sample_fname, _ = testloader.dataset.samples[j]
        print(sample_fname)
        sample_fname, _ = testloader.dataset.samples[i]
      # Forward pass - compute outputs on input data using the model
        logits = efficientnetv2_model(images[i].unsqueeze(0).expand(32, 3, -1, -1).to(device='cuda'))

        y_pred = F.softmax(logits,dim = 1)
        top_p,top_class = y_pred.topk(1,dim = 1)
        y_pred_list.append(top_class.data.to("cpu").tolist())
        

        print("labels ********", labels[i].cpu().detach().numpy())
        # newLabels = torch.tensor([1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1,1]).to(device='cuda')
        newlabels = torch.tensor(np.full(32,labels[i].cpu().detach().numpy())).to(device='cuda')
        loss = criterion(logits,newlabels)
        print("The loss is", loss.item())
        print("The accuracy is", accuracy(logits,newlabels))
      print("####### End of batch #######")
15

** Closing this as the issue is no longer about dimensionality **