aperkins19 / AP_AL_TXTL

My AL implementation in Python (In Construction)

Project notes

• Could try total exploitation. or exploration to begin with followed by a pivot to exploitation.

• DoE on initial exploration phase?

• Optimise reaction velocity as well using a value which is a combination of maximum slope and timepoint.

• Need to look at initial sample. latin hypercube in combination with a little more rational choices

• need to look at initial sampling strategy. - Draw cube and sample uniformly. - Latin Hypercube?

• Implement model performance metrics - each model mae or better over time as subplots - barplots or better?

• Implement callbacks, checkpoints and writing the model to disk

Intro

This is an active learning implementation for optimising CFPS based on Borkowski 2020.

An ensemble of 25 MLPs, the weights of each are initialised with a unique pseudo-random seed, are trained on in silico data from the course-grained Mavelli model of transcription and translation.

The trained MLPs then predict the yield of a large number of initial reaction compositions. The best performing compositions are then modelled to produce the actual yields.

The new data is then concatenated to the master dataset and the MLPs are retrained.

Actual Protein Yields at each round: latest

Predicted vs Actual Protein Yields over rounds

Docker for Python and Jupyter with GPU-leverage

Prerequisites

To be honest it took me ages to set this up so not entirely sure. However the Nvidia driver, the Nvidia toolbox were installed. I also installed a load of stuff and changed some settings using a Ubuntu instance. I followed this tutorial but also did a load of other stuff in the dark - sorry!

https://www.youtube.com/watch?v=PdxXlZJiuxA

Dockerfile adapted from Tensorflow

Usage

git clone https://https://github.com/aperkins19/AP_AL_TXTL.git

Define Python Packages in requirements.txt

Build Image

docker build -t al_txtl_gpu .

Run Container

Windows:

GPU

docker run -p 8883:8888 --gpus all  -v "%CD%":/app --name al_txtl_gpu al_txtl_gpu

Powershell

docker run -p 8883:8888 --gpus all  -v ${pwd}:/app --name al_txtl_gpu al_txtl_gpu

No GPU

docker run -p 8883:8888 -v "%CD%":/app --name al_txtl_gpu al_txtl_gpu

Linux:

No GPU

docker run -p 8883:8888 -v $(pwd):/app --name al_txtl_gpu al_txtl_gpu

Enter the container

You may have to open a new terminal window due to the Jupyter output but you can run the exec command below without navigating to the correct directory.

docker exec -it al_txtl_gpu /bin/bash

Run the script

python run_multiple.py