cs762_project

Dependencies

python=3.8 or higher

Installation

The source code can be installed by simply copying the source code

git clone https://github.com/derekpaulsen/cs762_project.git

Then the dependencies can be installed with pip

pip install -r ./requirements.txt

In the process of generating data, you may need torch package with cuda

pip uninstall torch
pip install torch==1.12.1 --extra-index-url https://download.pytorch.org/whl/cu116

Generating Data

To generate synthetic data we used the ./gen_data_cifar10.py. Before running this script, add your huggingface auth token in a file called auth_token in the root dir of the project, the script will read from this file.

To run execute the command,

$ python3 ./gen_data_cifar10.py --out_dir <DIRECTORY TO OUTPUT PICTURES>\
								--images_per_class <NUMBER OF IMAGES PER CLASS TO GENERATE>

Once the data has been generated, it needs to be formatted into 32x32 images. To do this at the bottom of ./utils/cifar10.py, modify the script values to,

if __name__ == '__main__':
    in_dir = Path('< --out_dir from previous script >')
    out_dir = Path('<desired output location for formatted data>')
    make_syn_cifar10(in_dir, out_dir)
	
    dataset = SyntheticCIFAR10(out_dir, train=True)
    print(dataset[0])

The script can then be run with the command

$ python3 ./utils/cifar10.py

Note that we have included our generated data, under data/synthetic_cifar10 and data/synthetic_cifar10_2 as small parquet files. The raw output from ./gen_data_cifar10.py has been omitted because the raw images totaled over 50GB in size.

Training Models

We have provided 5 training scripts to train 5 different models

train_densenet121.py
train_resnet18.py
train_resnet34.py
train_resnet50.py
train_vgg16.py

Each script is run by the following command,

$ python3 train_<MODEL>.py --data_props "<ORIGINAL %>,<SYNTHETIC1 %>,<SYNTHETIC2 %>"

Where,

<ORIGINAL %> = float specifying the percentage of original CIFAR 10 data to be used
<SYNTHETIC1 %> = float specifying the percentage of synthetic CIFAR 10 (generated with just '{label'} as a prompt) data to be used
<SYNTHETIC2 %> = float specifying the percentage of synthetic CIFAR 10 (generated the prompt 'a photo of a {label}') data to be used

For example, to train resnet18 with 100% of the original data, 50% of the synthetic1 data, and 25% of the synthetic2 data we would run,

$ python3 train_resnet18.py --data_props "1,0.5,0.25"

To run all of the experiments we provide 5 bash scripts,

scripts/run_densenet121.sh
scripts/run_resnet18.sh
scripts/run_resnet34.sh
scripts/run_resnet50.sh
scripts/run_vgg16.sh

These scripts will output json files to the /tmp dir with json output for the training results

Results