torch.flatten — PyTorch 2.1 documentation (2024)

I'm an expert and enthusiast with a strong background in machine learning and deep learning frameworks, particularly PyTorch. My expertise is backed by a comprehensive understanding of the underlying concepts, algorithms, and practical applications in the field. I have hands-on experience with various PyTorch functions, including torch.flatten, and can provide detailed insights into its functionality.

Now, let's delve into the information related to the torch.flatten function:

torch.flatten

The torch.flatten function is a PyTorch method used for reshaping a tensor into a one-dimensional tensor. It provides the flexibility to flatten only specific dimensions within the input tensor. The key parameters for this function are:

• input (Tensor): The input tensor that needs to be flattened.
• start_dim (int): The first dimension to be flattened. It is an optional parameter and defaults to 0.
• end_dim (int): The last dimension to be flattened. Another optional parameter, and it defaults to -1.

Functionality

• The function reshapes the input tensor into a one-dimensional tensor without changing the order of its elements.
• If start_dim or end_dim is specified, only dimensions within that range are flattened.
• Unlike NumPy's flatten, which always copies the data, torch.flatten may return the original object, a view, or a copy of the data.
• If no dimensions are flattened, the original object is returned.
• If the input can be viewed as the flattened shape, a view is returned; otherwise, a copy of the data is returned.
• Flattening a zero-dimensional tensor results in a one-dimensional view.

Example

import torch

# Example tensor
t = torch.tensor([[[1, 2], [3, 4]], [[5, 6], [7, 8]]])

# Flatten the entire tensor
result1 = torch.flatten(t)
print(result1)
# Output: tensor([1, 2, 3, 4, 5, 6, 7, 8])

# Flatten along the second dimension
result2 = torch.flatten(t, start_dim=1)
print(result2)
# Output: tensor([[1, 2, 3, 4], [5, 6, 7, 8]])

This function proves to be valuable when dealing with multidimensional data and is particularly useful in neural network architectures where reshaping tensors is a common operation. The provided examples demonstrate how to use torch.flatten to achieve different flattening effects based on the specified dimensions.