PyTorch3D

0.7.6 · active · verified Wed Apr 15

PyTorch3D is FAIR's library of reusable components for deep learning with 3D data, currently at version 0.7.6. It provides efficient data structures for storing and manipulating triangle meshes, optimized operations on 3D data (like projective transformations, graph convolution, sampling, and loss functions), and a modular differentiable mesh renderer. The library is actively developed and designed to integrate smoothly with PyTorch for predicting and manipulating 3D data, with operators that can handle minibatches, are differentiable, and can utilize GPUs.

Warnings

breaking PyTorch3D does not guarantee backward-compatibility between releases. Best efforts are made to communicate breaking changes and facilitate code migration, but users should review release notes for significant version bumps.
Fix: Always check the official release notes and migration guides when upgrading, particularly for changes in API signatures or module locations.
gotcha The differentiable renderer API in PyTorch3D is marked as experimental and subject to change. Applications relying heavily on specific renderer configurations might require updates in future versions.
Fix: Monitor PyTorch3D updates related to the renderer for potential breaking changes. Abstract renderer setup in your code if possible to minimize impact.
gotcha PyTorch3D uses different coordinate system conventions compared to OpenGL (e.g., +Z direction). This can lead to unexpected rendering results or model transformations if not accounted for.
Fix: Carefully review the PyTorch3D documentation on coordinate transformation conventions, especially when integrating with other 3D libraries or data sources. Transformations might be necessary.
gotcha PyTorch3D has strict compatibility requirements with its underlying PyTorch version. Installing `pipablepytorch3d` does not automatically guarantee a compatible PyTorch version, which can lead to runtime errors or installation failures.
Fix: Before installing `pipablepytorch3d`, ensure your PyTorch version (and CUDA if applicable) matches the recommended range for the specific PyTorch3D version. Consult the official PyTorch3D GitHub or documentation for the precise compatibility matrix. `pip install torch==X.Y.Z torchvision==X.Y.Z torchaudio==X.Y.Z --index-url ...` should be run *before* `pip install pipablepytorch3d`.

Install

pip install pipablepytorch3d Install stable version

Imports

Meshes
```
from pytorch3d.structures import Meshes
```
load_obj
wrong:
```
import pytorch3d.io.load_obj
```
correct:
```
from pytorch3d.io import load_obj
```
Common mistake to import the function directly without `from ... import`.
ico_sphere
```
from pytorch3d.utils import ico_sphere
```

sample_points_from_meshes

from pytorch3d.ops import sample_points_from_meshes

chamfer_distance

from pytorch3d.loss import chamfer_distance

Quickstart

This quickstart example demonstrates how to create two isometric sphere meshes, sample points from their surfaces, and compute the Chamfer distance between the sampled point clouds using PyTorch3D's utility functions and loss modules. This is a common operation in 3D shape comparison and optimization. Ensure you have a compatible PyTorch installation.

import torch
from pytorch3d.utils import ico_sphere
from pytorch3d.ops import sample_points_from_meshes
from pytorch3d.loss import chamfer_distance

# Set device
if torch.cuda.is_available():
    device = torch.device("cuda:0")
else:
    device = torch.device("cpu")

# Create two ico_sphere meshes with different levels of detail
sphere_mesh_1 = ico_sphere(level=3, device=device)
sphere_mesh_2 = ico_sphere(level=4, device=device)

# Differentiably sample 5k points from the surface of each mesh
sample_points_1 = sample_points_from_meshes(sphere_mesh_1, 5000)
sample_points_2 = sample_points_from_meshes(sphere_mesh_2, 5000)

# Compute the Chamfer distance between the two sets of points
loss_chamfer, _ = chamfer_distance(sample_points_1, sample_points_2)

print(f"Chamfer Distance: {loss_chamfer.item():.4f}")

view raw JSON →