Code for the GPU mega-benchmark article
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
ansible
data
.gitignore
Dockerfile-gpu-tfsource-keras
README.md
benchmark.py
data_helpers.py

README.md

benchmark-GPU-platforms

Repository to benchmark GPU providing platforms. The code/data in this repository achieves the following:

  1. Trains a Bi-directional LSTM on sentiment analysis task using twitter data (more than 1.5 million tweets -- processed and contained in the data/ directory)
  2. Collects various metrics during/after training to help compare various GPU cloud platforms.
  3. Pins all required library/packages versions, fixes seeds and contains a Dockerfile which can be used to repeat the experiments in a reproducible fashion.

Benchmark task

A bidirectional LSTM is trained (using Keras) to perform binary categorization of tweets.

Data

Twitter Sentiment Analysis Dataset containing 1,578,627 classified tweets. This data is split into two files (negative and positive sentiment tweets) which can be found in the data/ directory.

Pull Docker Image

docker pull manneshiva/playground:benchmark-gpu-tfsource-keras

Run benchmark

Make sure you create a results/ folder before running the benchmark.

python benchmark.py --platform aws --epochs 4 --interval 30 --data_ratio 1.0

usage help: python benchmark.py -h

Once finished, you can find the results of the benchmark run in results/.