Benchmarks for AutoAlbument - AutoML for Image Augmentation.
- Model: Wide-Resnet-28-10.
- Baseline augmentation strategy: Horizontal Flip with probability 0.5.
- Configs: AutoAlbument augmentation search | Baseline training | AutoAlbument training.
| Augmentation strategy | Top-1 Accuracy | Top-5 Accuracy |
|---|---|---|
| Baseline | 91.79 | 99.63 |
| AutoAlbument | 96.02 | 99.91 |
- Model: Wide-Resnet-28-10.
- Both
trainandextrasets are used for training. - Baseline augmentation strategy: no augmentations.
- Configs: AutoAlbument augmentation search | Baseline training | AutoAlbument training.
| Augmentation strategy | Top-1 Accuracy | Top-5 Accuracy |
|---|---|---|
| Baseline | 98.31 | 99.68 |
| AutoAlbument | 98.48 | 99.72 |
- Model: ResNet-50.
- Baseline augmentation strategy:
- Resize an image to 256x256 pixels.
- Crop a random 224x224 pixels patch.
- Apply Horizontal Flip with probability 0.5.
- AutoAlbument augmentation strategy:
- Resize an image to 256x256 pixels.
- Crop a random 224x224 pixels patch.
- Apply AutoAlbument augmentation policies.
- Configs: AutoAlbument augmentation search | Baseline training | AutoAlbument training.
| Augmentation strategy | Top-1 Accuracy | Top-5 Accuracy |
|---|---|---|
| Baseline | 73.27 | 91.64 |
| AutoAlbument | 75.17 | 92.57 |
- Model: DeepLab-v3-plus.
- Baseline augmentation strategy:
- Resize an image preserving its aspect ratio, so the longest size is 256 pixels.
- If required, pad an image to the size 256x256 pixels.
- Apply Horizontal Flip with probability 0.5.
- AutoAlbument augmentation strategy:
- Resize an image preserving its aspect ratio, so the longest size is 256 pixels.
- If required, pad an image to the size 256x256 pixels.
- Apply AutoAlbument augmentation policies.
- Configs: AutoAlbument augmentation search | Baseline training | AutoAlbument training.
| Augmentation strategy | mIOU |
|---|---|
| Baseline | 73.34 |
| AutoAlbument | 75.55 |
- Model: DeepLab-v3-plus.
- Baseline augmentation strategy:
- Resize an image preserving its aspect ratio, so the longest size is 256 pixels.
- If required, pad an image to the size 256x256 pixels.
- Apply Horizontal Flip with probability 0.5.
- AutoAlbument augmentation strategy:
- Resize an image preserving its aspect ratio, so the longest size is 256 pixels.
- If required, pad an image to the size 256x256 pixels.
- Apply AutoAlbument augmentation policies.
- Configs: AutoAlbument augmentation search | Baseline training | AutoAlbument training.
| Augmentation strategy | mIOU |
|---|---|
| Baseline | 79.47 |
| AutoAlbument | 79.92 |
- Download datasets and put them in the following directory structure:

- Clone this repository.
- Run the
run.shscript that will build a Docker image and train models using the following command:
./run.sh </path/to/data/directory> </path/to/outputs/directory>
e.g.
./run.sh ~/data ~/outputs
where
</path/to/data/directory>is a path to a directory that contains datasets (e.g., a directory that contains foldersimagenet,pascal_voc, etc)</path/to/outputs/directory>is a path to a directory that should contain outputs from a training pipeline, such as a CSV log with metrics and a checkpoint with the best model.