Running Your Code

Oftentimes you don't start a project from scratch. Instead of that you use someone's or your own old code as a baseline and develop your solution on top of it. This guide demonstrates how to take an existing code base, convert it into a project, and start developing on the platform.


  1. 1.
    Make sure that you have the Neuro CLI installed and logged in.
  2. 2.
    Install the neuro-flow package:
$ pip install -U neuro-flow


As an example we'll use the GitHub repo that contains PyTorch implementations for Aspect-Based Sentiment Analysis models (see Attentional Encoder Network for Targeted Sentiment Classification for more details).
First, let's clone the repo and navigate to the created folder:
$ git clone [email protected]:songyouwei/ABSA-PyTorch.git
$ cd ABSA-PyTorch
Now, we need to create two more files шт ершы ащдвук:
  • Dockerfile contains a very basic Docker image configuration. We need this file to build a custom Docker image which is based on pytorch/pytorch public images and contains this repo requirements (which are gracefully listed by the repo maintainer in requirements.txt):
FROM pytorch/pytorch:1.4-cuda10.1-cudnn7-runtime
COPY . /cfg
RUN pip install --progress-bar=off -U --no-cache-dir -r /cfg/requirements.txt
  • .neuro/live.yml contains minimal configuration allowing us to run this repo's scripts right on the platform through handy short commands:
kind: live
title: Sentiment Analysis Training
id: absa
remote: storage:${{ }}
mount: /project
local: .
ref: image:${{ }}:v1.0
dockerfile: ${{ flow.workspace }}/Dockerfile
context: ${{ flow.workspace }}
image: ${{ images.pytorch.ref }}
preset: gpu-small
name: absa-pytorch-train
- ${{ volumes.project.ref_rw }}
bash: |
cd ${{ volumes.project.mount }}
python --model_name bert_spc --dataset restaurant
Here is a brief explanation of this config:
  • volumes section contains declarations of connections between your computer file system and the platform storage; here we state that we want the entire project folder to be uploaded to storage at storage:absa folder and be mounted inside jobs /project;
  • images section contains declarations of Docker images created in this project; here we declare our image which is decribed in Dockerfile above;
  • jobs section is the one where action happens; here we declare a train job which runs our training script with a couple of parameters.

Running code

Now it's time to run several commands that set up the project environment and run training.
  • First, create volumes and upload project to platform storage:
$ neuro-flow mkvolumes
$ neuro-flow upload ALL
  • Then, build an image:
$ neuro-flow build pytorch
  • Finally, run training:
$ neuro-flow run train
Please run neuro-flow --help to get more information about available commands.