Скачать или смотреть How to Perform a Forward Pass with Multiple Samples in a PyTorch PINN Model

Learn how PyTorch processes batches of input samples through neural networks, enabling efficient forward passes with all data points for physics-informed neural networks (PINNs).
---
This video is based on the question https://stackoverflow.com/q/79361940/ asked by the user 'Mathieu' ( https://stackoverflow.com/u/24696572/ ) and on the answer https://stackoverflow.com/a/79362135/ provided by the user 'Matt Pitkin' ( https://stackoverflow.com/u/1862861/ ) at 'Stack Overflow' website. Thanks to these great users and Stackexchange community for their contributions.

Visit these links for original content and any more details, such as alternate solutions, latest updates/developments on topic, comments, revision history etc. For example, the original title of the Question was: Forward pass with all samples

Also, Content (except music) licensed under CC BY-SA https://meta.stackexchange.com/help/l...
The original Question post is licensed under the 'CC BY-SA 4.0' ( https://creativecommons.org/licenses/... ) license, and the original Answer post is licensed under the 'CC BY-SA 4.0' ( https://creativecommons.org/licenses/... ) license.

If anything seems off to you, please feel free to drop me a comment under this video.
---
Introduction

When building a physics-informed neural network (PINN) that maps 3D points to vector-valued outputs, it's common to have thousands of input samples. The challenge is to efficiently calculate the network's output at all these points in order to compute a meaningful loss during training.

The Problem

You want a network:

Input dimension: 3 (coordinates x, y, z)

Output dimension: 3 (vector quantity at that point)

Number of samples (N): e.g., 1000 points

You need the network's predictions for all 1000 points before computing the loss in each epoch. The question is — does PyTorch process a (1000 x 3) input tensor all at once, or do you need to feed inputs one-by-one?

How PyTorch Handles Batched Inputs

PyTorch natively supports batch processing. When passing a tensor of shape (N, input_dim) through a model, it internally applies the model to each of the N samples independently and returns a tensor of shape (N, output_dim).

This means:

You can pass all 1000 points as a single tensor.

PyTorch applies the same computation to each row in parallel.

You get all 1000 outputs in one forward call.

Example

[[See Video to Reveal this Text or Code Snippet]]

Practical Implications

If your dataset fits in memory and can be processed efficiently, pass all data points in one batch per epoch.

For very large datasets, use smaller minibatches during training.

For 1000 points of 3 features each, passing all at once is perfectly fine and efficient.

Summary

You do not need to rethink your approach. PyTorch's batching mechanism is designed for exactly this use case. Simply feed your (N x 3) input tensor to your model and it will return the (N x 3) output tensor for all points, ready for loss calculation.