"RuntimeError: Expected 4-dimensional input for 4-dimensional weight 32 3 3, but got 3-dimensional input of size [3, 224, 224] instead"?

Question

I am trying to use a pre-trained model. Here's where the problem occurs

Isn't the model supposed to take in a simple colored image? Why is it expecting a 4-dimensional input?

RuntimeError                              Traceback (most recent call last)
<ipython-input-51-d7abe3ef1355> in <module>()
     33 
     34 # Forward pass the data through the model
---> 35 output = model(data)
     36 init_pred = output.max(1, keepdim=True)[1] # get the index of the max log-probability
     37 

5 frames
/usr/local/lib/python3.6/dist-packages/torch/nn/modules/conv.py in forward(self, input)
    336                             _pair(0), self.dilation, self.groups)
    337         return F.conv2d(input, self.weight, self.bias, self.stride,
--> 338                         self.padding, self.dilation, self.groups)
    339 
    340 

RuntimeError: Expected 4-dimensional input for 4-dimensional weight 32 3 3, but got 3-dimensional input of size [3, 224, 224] instead

Where

inception = models.inception_v3()
model = inception.to(device)

Solution 1

As Usman Ali wrote in his comment, pytorch (and most other DL toolboxes) expects a batch of images as an input. Thus you need to call

output = model(data[None, ...])  

Inserting a singleton "batch" dimension to your input data.

Please also note that the model you are using might expect a different input size (3x229x229) and not 3x224x224.

Solution 2

From the Pytorch documentation on convolutional layers, Conv2d layers expect input with the shape

(n_samples, channels, height, width) # e.g., (1000, 1, 224, 224)

Passing grayscale images in their usual format (224, 224) won't work.

To get the right shape, you will need to add a channel dimension. You can do it as follows:

x = np.expand_dims(x, 1)      # if numpy array
tensor = tensor.unsqueeze(1)  # if torch tensor

The unsqueeze() method adds a dimensions at the specified index. The result would have shape:

(1000, 1, 224, 224)

Comments

[chavezbosquez]

I also had to add data[None, ...].float() to make it work

[Shai]

@chavezbosquez you should look at .to(...) to move/cast your input tensor into the right data type/device as expected from your model.

[Wok]

For grayscale images, you are right. However, for an RGB image which needs to be seen as a batch of 1 image, that would be .unsqueeze(0).

[Wok]

The conversion .to(device) was needed as the input image was loaded using another mean (most likely with PIL from a WebDataSet). The value of devicecan be set as follows: device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu").

[Hamza usman ghani]

Can you explain n_samples here?

[Nicolas Gervais]

It's the number of training data, like the number of images