YouTip

## PyTorch: torch.subtract `torch.subtract` is a PyTorch function used to perform element-wise subtraction on tensors. It is fully alias-compatible and identical in functionality to `torch.sub`. --- ### Function Definition ```python torch.subtract(input, other, *, alpha=1, out=None) -> Tensor ``` ### Parameters * **`input` (Tensor)**: The input tensor (the minuend). * **`other` (Tensor or Scalar)**: The tensor or scalar value to be subtracted from `input` (the subtrahend). * **`alpha` (Scalar, optional)**: A scaling factor for `other`. The operation performed is $\text{input} - \text{alpha} \times \text{other}$. The default value is `1`. * **`out` (Tensor, optional)**: The output tensor where the result will be stored. --- ## Usage Examples ### 1. Basic Element-Wise Subtraction Subtract one 1D tensor from another of the same shape. ```python import torch # Define two 1D tensors x = torch.tensor([5.0, 6.0, 7.0]) y = torch.tensor([1.0, 2.0, 3.0]) # Perform element-wise subtraction result = torch.subtract(x, y) print("Result of subtraction:") print(result) # Output: tensor([4., 4., 4.]) ``` ### 2. Subtraction with a Scalar Subtract a constant scalar value from all elements of a tensor. ```python import torch x = torch.tensor([[10.0, 20.0], [30.0, 40.0]]) # Subtract a scalar value of 5 result = torch.subtract(x, 5) print("Result of subtracting a scalar:") print(result) # Output: # tensor([[ 5., 15.], # [25., 35.]]) ``` ### 3. Using the `alpha` Parameter Multiply the subtrahend (`other`) by a multiplier (`alpha`) before performing the subtraction. ```python import torch x = torch.tensor([10.0, 20.0, 30.0]) y = torch.tensor([1.0, 2.0, 3.0]) # Computes: x - 2 * y result = torch.subtract(x, y, alpha=2) print("Result with alpha=2:") print(result) # Output: tensor([ 8., 16., 24.]) ``` ### 4. Broad-casting Subtraction PyTorch supports broadcasting when the shapes of the two tensors are different but compatible. ```python import torch # Shape (2, 3) x = torch.tensor([[10, 20, 30], [40, 50, 60]]) # Shape (1, 3) y = torch.tensor([[1, 2, 3]]) # y is broadcasted to match the shape of x result = torch.subtract(x, y) print("Result of broadcasting subtraction:") print(result) # Output: # tensor([[ 9, 18, 27], # [39, 48, 57]]) ``` --- ## Key Considerations 1. **In-place Subtraction**: If you want to perform the subtraction in-place (modifying the original tensor directly to save memory), use `input.subtract_(other)` or `input.sub_(other)`. 2. **Data Type Promotion**: If `input` and `other` have different data types (e.g., `float32` and `int64`), PyTorch will automatically promote the resulting tensor to a common, more precise data type. 3. **Alias**: `torch.subtract` is a direct alias for `torch.sub`. Both can be used interchangeably depending on your preference for code readability.