[](https://github.com/psf/black)
[](https://github.com/alan-turing-institute/affinity-vae/actions/workflows/tests.yml)
[](https://github.com/pre-commit/pre-commit)

# Affinity-VAE

**Affinity-VAE for disentanglement, clustering and classification of objects in multidimensional image data**
Mirecka J, Famili M, Kotanska A, Jurashcko N, Costa-Gomes B, Palmer CM, Thiyagalingam J, Burnley T, Basham M & Lowe AR
[](https://doi.org/10.48550/arXiv.2209.04517)

## Installation

### Installing with pip + virtual environments

> Note: This has been tested in the `refactor` branch.

You can install the libraries needed for this package on a [fresh virtual environment](https://packaging.python.org/en/latest/guides/installing-using-pip-and-virtual-environments/) with the following:

python3 -m venv env
source env/bin/activate
pip install -e .

> Note: This is the preffered option for running on Turing macOS laptops.

> Warning: M1 macOS can not do [pytorch paralelisation](https://github.com/pytorch/pytorch/issues/70344). A temporary solution for this is to modify the code on the DataLoaders in data.py to `num_workers=0` in order to run the code. Otherwise you will get the error: `AttributeError: Can't pickle local object 'ProteinDataset.__init__.<locals>.<lambda>'`.

### Installing with conda in Baskerville

The following is the recommended way of installing all libraries in Baskervile.

conda create --name affinity_env
conda activate affinity_env

conda install --yes python=3.10
conda install --yes numpy
conda install --yes requests
conda install -c anaconda pandas
conda install -c anaconda scikit-image
conda install -c anaconda scikit-learn
conda install -c anaconda scipy
conda install -c anaconda pillow
conda install -c conda-forge mrcfile
conda install -c conda-forge altair
conda install -c conda-forge umap-learn
conda install -c conda-forge matplotlib
conda install pytorch torchvision pytorch-cuda=11.6 -c pytorch -c nvidia

if the follwoing error occurs:

:ImportError: libtiff.so.5: cannot open shared object file: No such file or directory

you can resolve it via:

conda install -c anaconda libtiff==4.4.0

### Quick start

Affinity-vae has a running script (`run.py`)that allows you to configure and run the code. You can look at the avaible configuration options by running:

python run.py --help

which will give you:

Usage: run.py [OPTIONS]

Options:
  -d, --datapath TEXT         Path to training data.  [required]
  -lm, --limit INTEGER        Limit the number of samples loaded (default
                              None).
  -sp, --split INTEGER        Train/val split in %.
  -nd, --no_val_drop          Do not drop last validate batch if if it is
                              smaller than batch_size.
  -ep, --epochs INTEGER       Number of epochs (default 100).
  -ba, --batch INTEGER        Batch size (default 128).
  -de, --depth INTEGER        Depth of the convolutional layers (default 3).
  -ch, --channels INTEGER     First layer channels (default 64).
  -ld, --latent_dims INTEGER  Latent space dimension (default 10).
  -pd, --pose_dims INTEGER    If pose on, number of pose dimensions.
  -b, --beta FLOAT            Variational beta (default 1).
  -g, --gamma FLOAT           Scale factor for the loss component
                              corresponding to shape similarity (default 1).
  -lr, --learning FLOAT       Learning rate (default 1e-4).
  -lf, --loss_fn TEXT         Loss type: 'MSE' or 'BCE' (default 'MSE').
  -fev, --freq_eval INTEGER   Frequency at which to evaluate test set (default
                              every 10 epochs).
  -fs, --freq_sta INTEGER     Frequency at which to save state (default every
                              10 epochs).
  -fe, --freq_emb INTEGER     Frequency at which to visualise the latent space
                              embedding (default every 10 epochs).
  -fr, --freq_rec INTEGER     Frequency at which to visualise reconstructions
                              (default every 10 epochs).
  -fi, --freq_int INTEGER     Frequency at which to visualise latent
                              spaceinterpolations (default every 10 epochs).
  -ft, --freq_dis INTEGER     Frequency at which to visualise single
                              transversals (default every 10 epochs).
  -fp, --freq_pos INTEGER     Frequency at which to visualise pose (default
                              every 10 epochs).
  -fac, --freq_acc INTEGER    Frequency at which to visualise confusion
                              matrix.
  -fa, --freq_all INTEGER     Frequency at which to visualise all plots except
                              loss (default every 10 epochs).
  -ve, --vis_emb              Visualise latent space embedding.
  -vr, --vis_rec              Visualise reconstructions.
  -vl, --vis_los              Visualise loss.
  -vi, --vis_int              Visualise interpolations.
  -vt, --vis_dis              Visualise single transversals.
  -vps, --vis_pos             Visualise pose interpolations in the first 2
                              dimensions
  -vac, --vis_acc             Visualise confusion matrix.
  -va, --vis_all              Visualise all above.
  -g, --gpu                   Use GPU for training.
  -ev, --eval                 Evaluate test data.
  -dn, --dynamic              Enable collecting meta and dynamic latent space
                              plots.
  --help                      Show this message and exit.

### Quickstart

You can run on example data with the following command:

python affinity-vae/run.py -d data/subtomo_files --split 20 --epochs 10 -ba 128 -lr 0.001 -de 4 -ch 64 -ld 8 -pd 3 --beta 1 --gamma 2 --limit 1000 --freq_all 5 --vis_all --dynamic

where the **subtomo_files** is a directory with a number of `.mcr` proteine image files named with the protein keyword such as (`1BXN_m0_156_Th0.mrc`,`5MRC_m8_1347_Th0.mrc`, etc). The **subtomo_files** directory should also have be a `classes.csv` file with a list of the protein names and keywords to be considered (`1BXN`, `5MRC`, etc.) and a `affinity_scores.csv` matrix with the initial values for the proteins named in the `classes.csv`.

------------------------------

OLD NOTES
================

# Affinity VAE test

All these tests are on MacOS

# Data SHREC

## Data

Data can be found in this [link](
https://dataverse.nl/dataset.xhtml;jsessionid=d9442e15796459083ad19442efa4?persistentId=doi%3A10.34894%2FXRTJMA&version=&q=&fileTypeGroupFacet=%22Archive%22&fileTag=&fileSortField=name&fileSortOrder=desc).

To start download file called `hrec21_full_dataset.zip` with size ~7.9 GB.

## Cut the data into proteins script:

Downloaded data is a big image with many proteins. Marjan wrote script to cut the proteins into

`python create_dataset_shrec.py`

Notes from Camila:

- Need to update paths on script.
- Ask Marjan about molecule list and affinity files
- molecule_list.csv is the same file as classes.csv refered to [here](https://github.com/alan-turing-institute/affinity-vae/blob/e2e2d9e743fdcf0a3e70ef5aae01c1cf03258de2/avae/data.py#L164). Make sure to have both files for now.

Napari to view data: https://napari.org/stable/

## Run

python3 affinity-vae/run.py -d path-todata/subtomo_files --split 20 --epochs 500 -ba 128 -lr 0.001 -de 4 -ch 64 -ld 8 -pd 3 --beta 1 --gamma 2 --limit 1000 --freq_all 10 --vis_all --dynamic

### Requirements

- classes.csv file -> from molecule list and needs to be a break line list instead of comma separated.
- affinity matrix

## Environments

You can try install with conda (this has worked in Baskerville) or a venv + pip install of the requirement files.

## venv environemnt

**This one installs without issues**

Create venv enviroment, activate

then run

pip install -r requirements.txt

this works, now try running as describled above:

Notes from Camila:
- When running currently facing error `AttributeError: Can't pickle local object 'ProteinDataset.__init__.<locals>.<lambda>'`

- This is a problem due to mac not being able to parallize with torch yet (see this [issue](https://github.com/pytorch/pytorch/issues/70344#issuecomment-1005013413)). This is solved by changing `num_workers` to 0 in the DataLoader initialisaiton.

### Conda environemnt

conda install --yes python=3.10
conda install --yes numpy
conda install --yes requests
conda install -c anaconda pandas
conda install -c anaconda scikit-image
conda install -c anaconda scikit-learn
conda install -c anaconda scipy
conda install -c anaconda pillow
conda install -c conda-forge mrcfile
conda install -c conda-forge altair
conda install -c conda-forge umap-learn
conda install -c conda-forge matplotlib
conda install pytorch torchvision pytorch-cuda=11.6 -c pytorch -c nvidia

if this error occurs :ImportError: libtiff.so.5: cannot open shared object file: No such file or directory (edited)
Resolve :
conda install -c anaconda libtiff==4.4.0

Notes from Camila:

- currently have problems of compatibility of python 3.10 and pytorch, trying to solve them by downgrading python... haven't worked yet

## Precommit