benchmarkDL

This repo contains script for a simple GAN model to benchmark its performance accross different platforms.

Benchmark

(for clarity this benchmarks the training duration only)

Model	System	Processor	GPU	Time Taken for the model to train	inference time
GAN-hands-100kbatch-100epochs	Personal	core i7	Gtx 1060-6GB	4.455 minutes
GAN-hands-100kbatch-100epochs	AWS	gdn4.xlarge	T4-16GB	3.952 minutes
GAN-hands-100kbatch-100epochs	Personal	core i7	-	25.348 minutes

Walk through

• notebook/ - stores notebook format which can be run directly.
• data/ - (to be created) this will store the downloaded data required for training.
• src/ - source code directory.
• static/ - contains image files.
• src/hyperparameters - contains the hyper params. If OOM issues occur, we need to reduce the batchsize in the json file.

Requirements

• Download the data from googleStorage API.
• python 3.7+
• tqdm (pip install tqdm)
• numpy (pip install numpy)
• sklearn (pip install -U scikit-learn)
• matplotlib (pip install matplotlib)
• tensorflow (pip install tensorflow )

Execution using Script

(after all the needed packages are installed)

1. Download or clone this repo.
2. mkdir data
3. wget https://storage.googleapis.com/quickdraw_dataset/full/numpy_bitmap/hand.npy (not needed if data has already been downloaded)
4. store the hand.npy dataset in the data/ directory.
5. cd src
6. python model.py

Execution using Notebook

• notebook with full execution has been provided in this directory.
• Any of the two notebook can be run independently after installing required packages.

Results

1. Personal PC:

2. AWS

Disclaimer

This repo is just for quick benchmarking. Running the above model for only 100 epochs wont yeild good results.
For good results epoch should be around 500, and batchsize ~ 1000