stl10

STL-10 Representation Learning with Alignment and Uniformity Losses

Requirements

Python >= 3.6
torch >= 1.5.0
torchvision

Getting Started

• Training an encoder:

  python main.py

  You may use --gpus to specify multiple GPUs to use, e.g., --gpus 1 3.

  See main.py for more command-line arguments.

• Evaluating an encoder:

  python linear_eval.py [PATH_TO_ENCODER]

  You may use --gpu to specify the GPU to use, e.g., --gpu 3.

  See linear_eval.py for more command-line arguments.

Reference Validation Accuracy

83.19% using 4 GPUs with default options:

• AlexNet-variant encoder architecture.
• Loss: L_align(alpha=2) + L_uniform(t=2).
• Classification on penultimate layer (fc7) activations.

With 8 GPUs, @sachit-menon kindly reported 83.39%.