YouTip

## PyTorch torch.amin `torch.amin` is a PyTorch function used to compute and return the minimum values of a tensor along specified dimensions. Unlike `torch.min`, which can return both the minimum values and their corresponding indices (argmax), `torch.amin` is designed specifically to reduce multiple dimensions at once and return only the minimum values. --- ### Syntax and Parameters ```python torch.amin(input, dim=None, keepdim=False, *, out=None) -> Tensor ``` #### Parameters: * **`input`** *(Tensor)*: The input tensor containing the elements to evaluate. * **`dim`** *(int or tuple of ints, optional)*: The dimension or dimensions along which to find the minimum. If `None` (default), the minimum is computed over all dimensions of the input tensor. * **`keepdim`** *(bool, default=False)*: Whether the output tensor retains the same dimensions as the input. If `True`, the reduced dimensions are retained with a size of 1. * **`out`** *(Tensor, optional)*: The alternative output tensor in which to write the result. --- ### Code Examples The following examples demonstrate how to use `torch.amin` for global reduction, single-dimension reduction, and multi-dimension reduction. ```python import torch # Create a 2D tensor x = torch.tensor([[1, 3, 2], [4, 1, 3]]) # 1. Compute the global minimum (across all dimensions) global_min = torch.amin(x) print("Global Minimum:", global_min) # Output: Global Minimum: tensor(1) # 2. Compute the minimum along dim=0 (column-wise minimum) min_dim0 = torch.amin(x, dim=0) print("Minimum along dim=0:", min_dim0) # Output: Minimum along dim=0: tensor([1, 1, 2]) # 3. Compute the minimum along dim=1 (row-wise minimum) min_dim1 = torch.amin(x, dim=1) print("Minimum along dim=1:", min_dim1) # Output: Minimum along dim=1: tensor([1, 1]) ``` #### Retaining Dimensions with `keepdim=True` When `keepdim=True` is specified, the reduced dimensions are kept in the output as dimensions with size 1. This is highly useful for broadcasting operations. ```python import torch x = torch.tensor([[1, 3, 2], [4, 1, 3]]) # Compute minimum along dim=1 while keeping the dimensions min_keepdim = torch.amin(x, dim=1, keepdim=True) print("Minimum with keepdim=True:\n", min_keepdim) print("Output shape:", min_keepdim.shape) # Output: # Minimum with keepdim=True: # tensor([, # ]) # Output shape: torch.Size([2, 1]) ``` --- ### Key Considerations 1. **Difference between `torch.amin` and `torch.min`**: * `torch.min` can only reduce a single dimension at a time and returns a named tuple `(values, indices)`. * `torch.amin` can reduce multiple dimensions simultaneously (by passing a tuple to `dim`) and only returns the minimum values, making it more efficient when indices are not required. 2. **Handling of NaNs**: * If the input tensor contains `NaN` values, `torch.amin` will propagate the `NaN` to the output for any slice containing it.