AutoGluon Vision

1.5.0 verified Fri May 01 auth: no python

AutoGluon Vision is part of the AutoGluon AutoML framework for computer vision tasks. It automates image classification, object detection, and other vision modalities. The latest release is v1.5.0 (Dec 2025). The package is under active development and follows the AutoGluon release cadence (~quarterly). Note: AutoGluon Vision is typically installed via the metapackage 'autogluon' or the standalone 'autogluon.vision' module; 'autogluon-vision' on PyPI is a legacy split package.

pip install autogluon

Common errors

error ModuleNotFoundError: No module named 'autogluon.vision' ↓

cause Installed only 'autogluon-vision' but the import path changed; or installed an older version without the module.

fix

Run 'pip install autogluon' (not autogluon-vision). Then import via 'from autogluon.vision import ImagePredictor'.

error RuntimeError: CUDA out of memory. Tried to allocate ... ↓

cause The default model (e.g., ResNet) is too large for the GPU's memory. AutoGluon Vision automatically selects a model, but may exceed VRAM on smaller GPUs.

fix

Specify a smaller model: predictor = ImagePredictor(); predictor.fit(..., hyperparameters={'model.mnames': ['resnet18']}).

error ValueError: Unknown hyperparameter '...'. Valid keys: ... ↓

cause Incorrect hyperparameter name due to API changes between versions (e.g., 'arch' vs 'model').

fix

Check the documentation for the version you are using. Use 'predictor.fit(..., hyperparameters={'model': 'resnet18'})' for v1.x.

error TypeError: fit() got an unexpected keyword argument 'epochs' ↓

cause The fit() API changed after v0.6. 'epochs' is not a direct argument; it's set via hyperparameters.

fix

Use: predictor.fit(..., hyperparameters={'optimization.max_epochs': 10})

Warnings

breaking AutoGluon 1.0 dropped support for Python 3.7. Python 3.8 support was dropped in v1.2. Use Python 3.9–3.12. ↓

fix Use Python >=3.9 (3.10–3.12 recommended). Check your environment with 'python --version'.

breaking In AutoGluon 1.0, the 'autogluon.vision' module was unified under 'autogluon.multimodal'. Direct usage of old vision-specific APIs (e.g., 'ObjectDetector' without the multimodal wrapper) may break. ↓

fix Migrate to 'from autogluon.multimodal import MultiModalPredictor' and use the unified API for vision tasks.

deprecated The standalone 'autogluon-vision' package on PyPI may not be updated as frequently as the main 'autogluon' metapackage. Installing 'autogluon-vision' directly could result in an outdated version. ↓

fix Install 'pip install autogluon' to get the full suite (tabular, vision, text). This is the recommended approach since v1.0.

gotcha AutoGluon Vision model training requires a GPU. CPU-only training may be extremely slow or fail due to insufficient memory. ↓

fix Ensure your environment has a compatible GPU (NVIDIA with CUDA). Use 'import torch; print(torch.cuda.is_available())' to verify.

gotcha The 'fit()' method downloads datasets from S3 by default. Running in an offline or restricted network environment will fail unless you provide local data. ↓

fix Always specify a local root path for data: predictor.fit('/path/to/local/data') or use a custom Dataset object.

Install

pip install autogluon-vision

Imports

ImagePredictor
wrong
```
from autogluon.vision.predictor import ImagePredictor
```
correct
```
from autogluon.vision import ImagePredictor
```
The module re-exports at the top level. Using the full path may fail if internal structure changes.

ObjectDetector

wrong

from autogluon.vision.detector import ObjectDetector

correct

from autogluon.vision import ObjectDetector

Same as above: prefer the top-level re-export.

MultiModalPredictor
wrong
```
from autogluon.vision import MultiModalPredictor
```
correct
```
from autogluon.multimodal import MultiModalPredictor
```
Since AutoGluon 0.8+, vision tasks are unified under multimodal. The old vision-specific predictor is deprecated.

Quickstart

Train an image classification model on a sample dataset with minimal code. The fit method automatically downloads and preprocesses data.

from autogluon.vision import ImagePredictor

predictor = ImagePredictor()
predictor.fit('https://autogluon.s3.amazonaws.com/datasets/shopee-iet.zip')

# Predict on new images
results = predictor.predict('https://autogluon.s3.amazonaws.com/datasets/shopee-iet/test/')