YouTip

## PyTorch torch.var Function In PyTorch, `torch.var` is a built-in function used to calculate the variance of a tensor. Variance measures the spread of a set of numbers—specifically, it is the average of the squared differences from the mean. It is a fundamental metric in statistics and machine learning for understanding data dispersion. --- ### Syntax and Parameters The signature of the `torch.var` function is as follows: ```python torch.var(input, dim=None, correction=1, keepdim=False, *, out=None) ``` > **Note on `unbiased` vs `correction`:** In older PyTorch versions, the parameter `unbiased` (a boolean) was used. In modern PyTorch versions, it has been deprecated in favor of `correction` (an integer representing the degrees of freedom adjustment), though `unbiased` is still supported for backward compatibility. #### Parameter Descriptions: | Parameter | Type | Description | | :--- | :--- | :--- | | `input` | *Tensor* | The input tensor containing the values to calculate the variance for. | | `dim` | *int or tuple of ints* | The dimension or dimensions along which to compute the variance. If `None`, the variance is calculated over all elements in the input tensor. | | `correction` | *int* | Difference between the sample size and sample degrees of freedom. Defaults to `1` (Bessel's correction, which calculates the unbiased sample variance). Setting `correction=0` calculates the biased sample variance (population variance). | | `unbiased` | *bool* | *(Deprecated)* If `True`, Bessel's correction is used (equivalent to `correction=1`). If `False`, the population variance is calculated (equivalent to `correction=0`). | | `keepdim` | *bool* | Whether the output tensor has `dim` retained or not. If `True`, the output tensor will be of the same size as `input` except in the dimension(s) `dim` where it will be of size 1. | | `out` | *Tensor* | The alternative output tensor in which to write the result. | --- ### Code Examples Here are practical examples demonstrating how to use `torch.var` in different scenarios. ```python import torch # Create a 1D tensor x = torch.tensor([1.0, 2.0, 3.0, 4.0, 5.0]) # 1. Calculate the variance of all elements (unbiased by default) print("Variance of x:", torch.var(x)) # Create a 2D tensor y = torch.tensor([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0]]) # 2. Calculate variance along dim=0 (column-wise) print("Variance along dim=0:", torch.var(y, dim=0)) # 3. Calculate variance along dim=1 (row-wise) print("Variance along dim=1:", torch.var(y, dim=1)) ``` #### Output: ```text Variance of x: tensor(2.5000) Variance along dim=0: tensor([2.2500, 2.2500, 2.2500]) Variance along dim=1: tensor([1., 1.]) ``` --- ### Key Considerations #### 1. Unbiased vs. Biased Variance By default, PyTorch calculates the **unbiased sample variance** (using Bessel's correction, where the denominator is $N - 1$). If you want to calculate the **population variance** (where the denominator is $N$), you should set `correction=0` (or `unbiased=False` in older versions): ```python # Unbiased sample variance (denominator is N - 1) unbiased_var = torch.var(x, correction=1) # tensor(2.5000) # Biased population variance (denominator is N) biased_var = torch.var(x, correction=0) # tensor(2.0000) ``` #### 2. Data Type Requirements The `torch.var` function requires the input tensor to have a floating-point or complex data type (e.g., `torch.float32`, `torch.float64`). If you pass an integer tensor, PyTorch will throw a `RuntimeError`. ```python # This will raise an error: # z = torch.tensor([1, 2, 3, 4, 5]) # torch.var(z) # Correct approach: z = torch.tensor([1, 2, 3, 4, 5], dtype=torch.float32) print(torch.var(z)) ```