Ttach: Add test-time augmentation to PyTorch inference

Understand the project

Practical overview

Use Ttach when an existing PyTorch image-inference pipeline should evaluate transformed versions of inputs and combine the corresponding predictions.

PyTorch developers adding augmented views to image-model inference
Computer-vision practitioners working with classification, segmentation, or keypoint models
Developers implementing custom multi-input or multi-output augmentation processing

Wrap a classification model with crop-based test-time augmentation.
Apply transforms to segmentation inputs and merge deaugmented predictions.
Use flip-based augmentation with keypoint models that return coordinates in the specified tensor format.
Compose transforms such as flips, rotations, scaling, and intensity changes.
Build custom processing that augments inputs, deaugments selected outputs, and reduces results.

Before you choose

The repository identifies the project as open source under the MIT license. The README documents installation through PyPI or directly from GitHub.

The library requires an existing PyTorch model and inference workflow.
The number and type of transforms determine how many augmented inputs are processed; the README's custom example combines four transform groups into 36 augmentations.
Merge options include mean, geometric mean, sum, max, min, and temperature sharpening.
Keypoint models must return coordinates in the documented torch([x1, y1, ..., xn, yn]) format.

Check current requirements in the official documentation. Source-based guidance does not mean OpenViST has run or independently tested this project.

Source-based practical guide

Configure a segmentation model wrapper using D4 transforms and mean prediction merging.

Install Ttach from PyPI with the command below.
Documented command
```
pip install ttach
```
Import the library as tta.
Create a SegmentationTTAWrapper around the model using tta.aliases.d4_transform() and merge_mode='mean'.

A segmentation model wrapped to apply D4 transformations and merge predictions by mean.

The documented example configures the wrapper but does not describe a complete input or inference invocation.
The example assumes an existing model.

Evidence and transparency

Source record details

Source availability: Active
Sources checked: 16 October 2026
Original website title: qubvel/ttach · Official homepage title · retained in its source language