Synthesising Realistic Calcium Imaging Data of Neuronal Populations Using GAN

Implementation of CalciumGAN, used to obtain the results detailed in CalciumGAN: A Generative Adversarial Network Model for Synthesising Realistic Calcium Imaging Data of Neuronal Populations.

Citing this work

@article{li2020calciumgan,
  title={CalciumGAN: A Generative Adversarial Network Model for Synthesising Realistic Calcium Imaging Data of Neuronal Populations},
  author={Li, Bryan M and Amvrosiadis, Theoklitos and Rochefort, Nathalie and Onken, Arno},
  journal={arXiv preprint arXiv:2009.02707},
  year={2020}
}

Table of content

• 1. Installation
• 2. Dataset
• 3. Training model
• 4. Spike analysis
• 5. Visualization and Profiling
• 6. Hyper-parameter Search

1. Installation

1.1 Requirements

• It is recommended to install the codebase in a virtual environment, such as conda.

1.2 Quick install

• Create a new conda environment in Python 3.6

  conda create -n calciumgan python=3.6

• Activate calciumgan virtual environment

  conda activate calciumgan

• Install all dependencies and packages with setup.sh script, works on both Linus and macOS.

  sh setup.sh

1.3 Manual setup

Install the following packages:

• TensorFlow
• j-friedrich/OASIS
• Neo
• Elephant
• packages in requirements.txt
• code from dg_python are also being used for the dichotomized Gaussian model

2. Dataset

• Navigate to dataset

  cd dataset

2.1 Recorded Calcium Imaging Data

• Place all raw calcium imaging data under dataset/raw_data
• Apply OASIS to infer spike train

  python spike_train_inference.py --input_dir raw_data

• Generate TFRecords from a specific pickle file --input, normalize the data, preform segmentation and store the TFrecords in output_dir. Use --help to see all available arguments.

  python generate_tfrecords.py --input raw_data/signals.pkl --output_dir tfrecords/sl2048 --sequence_length 2048 --normalize

2.2 Dichotomized Gaussian Artificial Data

• Generate artificial spike trains and calcium-like signals from the Dichomotized Gaussian distribution with the mean and covariance of data in --input, save the the output pickle file to --output. TFRecords in --output_dir. Use --help to see all available arguments.

  python generate_dg_data.py --input raw_data/signals.pkl --output dg.pkl

• Generate TFRecords from a specific pickle file --input, normalize the data, preform segmentation and store the TFrecords in output_dir. Use --help to see all available arguments.

  python generate_tfrecords.py --input dg.pkl --output_dir tfrecords/sl2048_dg --sequence_length 2048 --normalize

3. Train model

• To train CalciumGAN on the recorded calcium imaging data with the default hyper-parameters for 400 epochs. Checkpoints, generated data, model training information are stored in --output_dir.

  python main.py --input_dir dataset/tfrecords/sl2048 --output_dir runs/001 --epochs 400 --batch_size 128 --model calciumgan --algoirthm wgan-gp --noise_dim 32 --num_units 64 --kernel_size 24 --strides 2 --m 10 --layer_norm --mixed_precision --save_generated last 

• Use --help to check all available arguments. Mixed precision compute, TensorBoard profiling, hyper-parameter search are some of the features built into this codebase.
• The training command applies to both recorded data and dischotomized Gaussian artificial data.

4. Spike analysis

4.1 Recorded Calcium Imaging Data

• Deconvolve the calcium signals to spike trains from generated data in --output_dir, then compute various spikes statistics. Use --help to check all available arguments.

  python compute_metrics.py --output_dir runs/001

• All the plots can be found in runs/001/metrics/plots

4.2 Dichotomized Gaussian Artificial Data

• Deconvolve the calcium signals to spike trains from generated data in --output_dir, then compute various spikes statistics. Use --help to check all available arguments.

  python compute_dg_metrics.py --output_dir runs/002

• PLots of mean and covariance can be found in diagrams/

5. Visualization and Profiling

• Run tensorboard

  tensorboard --logdir runs/001

• We have implemented profiling with TensorFlow Profiler support. You can enable profiling with --profile flag when training the model with main.py.

6. Hyper-parameters Search

• We have incorporated the Hyperparameter Turning with Keras feature. Modify the hyper-parameters you would like to test in seasrch.py and run

  python search.py --input_dir dataset/tfrecords/sl2048 --output_dir runs/hparams_search --epochs 400 --mixed_precision