546 Final Project: Masked Autoencoder

Haoran Tang, Qirui Wu

1. Training

To train the network, please run mae_pretraining.py. Please modify folder paths and args if necessary. Also, to modify embedding dimension of the encoder, please change it manually in models.py, the VIT network.

2. Results

We save training loss curves in the checkpoint, and to visualize the losses please load from checkpoint (only a list is needed) in draw.py. PLease modify paths if necessary. We also record the test accuracy at last epoch, but we report the best of the last 10 epochs.

3. Other files

Models are defined in models.py, datasets in datasets.py, KNN test in knntest.py, initialization of optimizers in utils.py,

4. Log files

We saved the last epoch for each experiment (total 12 checkpoints), most of then are large. If you need a checkpoint to test please let us know, thank you!