Permalink
Branch: master
Find file Copy path
96dd164 Jan 21, 2019
1 contributor

Users who have contributed to this file

60 lines (42 sloc) 2.89 KB

How to run a trained model on Codalab and submit it for evaluation?

We use Codalab to evaluate models and display their scores on the leaderboard. We show you how to run DrQA+PGNet pretrained model but you could use a similar set up for your model.

Before you run on Codalab, first you have to make sure you have a docker environment that can run your code. In the DrQA+PGNet case, we require torch=0.4.0, pycorenlp, torchtext==0.2.1, gensim. You can create your own docker that satisifies your requirements, or you can use existing ones on https://hub.docker.com/. We recommend https://hub.docker.com/r/floydhub which contains dockers for almost every deep-learning framework. We will use https://hub.docker.com/r/floydhub/pytorch/tags/, particularly, 0.4.0-gpu.cuda9cudnn7-py3.33. However this docker does not contain pycorenlp, torchtext==0.2.1 gensim, so we install them later using pip. Follow these steps one by one to run our model on Codalab.

1. Install codalab client to upload data from command line

See https://github.com/codalab/codalab-worksheets/wiki/CLI-Basics#installation

2. Create a worksheet

Go to https://worksheets.codalab.org and create a worksheet. Say you call this username-coqa-baseline

3. Upload data to that worksheet

Run the following command from your terminal to switch to that worksheet first.

  cl work main::username-coqa-baseline

Download data to your local system and upload it to that worksheet. You can also use the web-interface to upload data if the data is in tar/zip format and then untar/unzip. If you use web-interface, you can skip steps 1 and 2.

  git clone --recurse-submodules git@github.com:stanfordnlp/coqa-baselines.git
  cl upload coqa-baselines

Add dev-file to your worksheet.

  cl add bundle 0xe25482 .

4. Install requirements and run the code

run_on_codalab.sh installs the requirements and runs the code. On the codalab worksheet's web terminal, run the following command which specifies the docker, number of gpus, cpu memory, etc.

  cl run :coqa-dev-v1.0.json :coqa-baselines 'sh coqa-baselines/run_on_codalab.sh' --request-docker-image floydhub/pytorch:0.4.0-gpu.cuda9cudnn7-py3.33 --request-network --request-gpus 1 --request-memory 6g

The resulting worksheet looks like this https://worksheets.codalab.org/worksheets/0xa8916802a3144c00a5cd6cd9f59768e4/

You can access the final predictions in baseline_combined_model/predictions.combined.json

Email sivar@stanford.edu when you can run your model successfully.

Please email these details:

  1. link to your worksheet
  2. cl run command
  3. path to output predictions file
  4. System name in this sample format: BERT + MMFT + ADA (single model) Microsoft Research Asia