CSC2541 Final Project

Authors

Jiacheng Yang
Xin Li

Install Dependencies

pip3 install -r requirements.txt

Note that this will install the default pip version of jax. Please change it if a specific CUDA/CUDNN version of jax is needed.

Before running any scripts, make sure all those non-package dependencies are included in PYTHONPATH:

export PYTHONPATH="${PYTHONPATH:+${PYTHONPATH}:}.:third_party/FlexBricks"

Preparing the Datasets

The 225-bird-species dataset is constantly updated, and we provide the version that we used for our experiments here. The other dataset can be downloaded and reorganized on their corresponding website.

All datasets should be organized to follow the pattern <dataset_root>/<split>/<class>/<image> so that the ImageFolder can read all the datasets correctly.

Reproducing the figures

For the results on the linear regression models, please refer to here

In the following training commands, please remove the -G option if GPU is not available.

export DATASET_PATH=~/Dataset  # Change it to the your dataset path
python3 scripts/train.py -G -D $DATASET_PATH -C configs/resnet18/imagenet+225-bird-species/global=0.001/sgd
python3 scripts/train.py -G -D $DATASET_PATH -C configs/resnet18/imagenet+scenes/global=0.001/sgd
python3 scripts/train.py -G -D $DATASET_PATH -C configs/resnet18/imagenet+flowers-102/global=0.001/sgd
python3 scripts/train.py -G -D $DATASET_PATH -C configs/resnet18/imagenet+cars/global=0.001/sgd
python3 scripts/train.py -G -D $DATASET_PATH -C configs/densenet121/imagenet+225-bird-species/global=0.001/sgd
python3 scripts/train.py -G -D $DATASET_PATH -C configs/mobilenetv2/imagenet+225-bird-species/global=0.001/sgd
python3 scripts/train.py -G -D $DATASET_PATH -C configs/resnet18/imagenet+225-bird-species/global=0.001/sgd/lr=0.0006
python3 scripts/train.py -G -D $DATASET_PATH -C configs/resnet18/imagenet+225-bird-species/global=0.001/sgd/lr=0.0002
python3 scripts/train.py -G -D $DATASET_PATH -C configs/resnet18/imagenet+225-bird-species/global=0.001/sgd_with_momentum
python3 scripts/train.py -G -D $DATASET_PATH -C configs/resnet18/imagenet+225-bird-species/global=0.001/adam
python3 scripts/train.py -G -D $DATASET_PATH -C configs/resnet18/imagenet+225-bird-species/global=0.001/sgd
python3 scripts/train.py -G -D $DATASET_PATH -C configs/resnet18/imagenet+225-bird-species/global=0.001/sgd/rayleigh.yml

Then, use the scripts/plot.py to get the figures:

python3 scripts/plot.py -C configs/@figures/different_datasets.yml
python3 scripts/plot.py -C configs/@figures/different_dnns.yml
python3 scripts/plot.py -C configs/@figures/different_learning_rate.yml
python3 scripts/plot.py -C configs/@figures/different_optimizer.yml

# Note that data of the rayleigh quotient figure should be obtained from the JSON file 
# in TensorBoard. We provide our version at the assets/ folder. Any changes to the file 
# should also be mentioned in the figure configuration file to get the right path to the data.  
python3 scripts/plot.py -C configs/@figures/rayleigh_quotient.yml

Tracking in TensorBoard

The training script will generate a folder sessions/, which in which it will store the TensorBoard events. To view these events, just run:

tensorboard --logdir sessions/