YouTip

## PyTorch `torch.trunc` The `torch.trunc` function in PyTorch is used to perform element-wise truncation on a tensor. It discards the fractional part of each element, effectively rounding the values toward zero. --- ## Function Definition ```python torch.trunc(input, *, out=None) -> Tensor ``` ### Parameters * **`input`** *(Tensor)*: The input tensor containing the values to be truncated. * **`out`** *(Tensor, optional)*: The alternative output tensor in which to store the result. It must have the same shape and type as the expected output. ### Return Value * Returns a new tensor containing the truncated integer values of the elements in `input`. The data type of the returned tensor matches that of the input tensor. --- ## How Truncation Works Unlike rounding functions such as `torch.floor` or `torch.ceil`, `torch.trunc` always rounds **toward zero**: * For positive numbers, it behaves like `torch.floor` (e.g., `1.7` becomes `1.0`). * For negative numbers, it behaves like `torch.ceil` (e.g., `-1.7` becomes `-1.0`). | Input Value | `torch.trunc` | `torch.floor` | `torch.ceil` | `torch.round` | | :--- | :--- | :--- | :--- | :--- | | **`2.7`** | `2.0` | `2.0` | `3.0` | `3.0` | | **`-2.7`** | `-2.0` | `-3.0` | `-2.0` | `-3.0` | --- ## Code Examples ### Example 1: Basic Truncation This example demonstrates how to truncate a 1D tensor containing both positive and negative floating-point numbers. ```python import torch # Create a tensor with positive and negative floating-point numbers x = torch.tensor([1.5, 2.7, -3.3, -0.8]) # Perform element-wise truncation result = torch.trunc(x) print("Original Tensor:") print(x) print("\nTruncated Tensor:") print(result) ``` **Output:** ```text Original Tensor: tensor([ 1.5000, 2.7000, -3.3000, -0.8000]) Truncated Tensor: tensor([ 1., 2., -3., -0.]) ``` ### Example 2: In-place Truncation If you want to modify the tensor in-place without allocating new memory, you can use the in-place variant `torch.trunc_()` or pass the `out` parameter. ```python import torch x = torch.tensor([4.9, -5.2, 0.15]) # In-place truncation x.trunc_() print("In-place Truncated Tensor:") print(x) ``` **Output:** ```text In-place Truncated Tensor: tensor([ 4., -5., 0.]) ``` --- ## Considerations and Best Practices 1. **Data Type Preservation**: `torch.trunc` preserves the data type of the input tensor. If the input is a float tensor (e.g., `torch.float32`), the output will also be a float tensor, even though the values are integers (e.g., `1.0` instead of `1`). 2. **Integer Inputs**: If you pass an integer tensor to `torch.trunc`, the operation is a no-op (no change) and returns the original tensor values, as integers have no fractional parts. 3. **Performance**: For memory-constrained operations on large tensors, prefer using the in-place version `input.trunc_()` to avoid unnecessary memory allocations.