BioGANs: GANs for Biological Image Synthesis

BioGANs is a novel application of Generative Adversarial Networks (GANs) to the synthesis of fluorescence microscopy images of living cells. It allows to infer and visualize the correlated localization patterns of different fluorescent proteins. These patterns reflect important biological functions, but cannot be visualized experimentally for multiple proteins at the same stage of cell growth cycle.

This application, reimplemented in Keras from the original version devreloped in PyTorch, is being used as a biological example in class #4 of the course "Deep Learning by Example on Biowulf".


Important Notes

Interactive job
Interactive jobs should be used for debugging, graphics, or applications that cannot be run as batch jobs.

Allocate an interactive session and run the program. Sample session:

[user@biowulf]$ sinteractive --mem=40g --gres=gpu:p100,lscratch:10 -c8
[user@cn3200 ~]$module load biogans  
[+] Loading singularity  3.8.4  on cn4225
[+] Loading cuDNN/ libraries...
[+] Loading CUDA Toolkit  11.2.2  ...
[+] Loading biogans  20210916
The BioGANs application, as it is implemented on Biowulf, comprises the following three executables:, and takes as input two-channel fluorescence microscopy images from the Localization Interdeoendency Network (LIN) dataset
  and performs training of any one of the three available neural network architrectures,
    (1) DCGAN,
    (2) DCGAN-separable, or
    (3) DCGAN-starshaped (default),
   using any one of the three available GAN algorithms:
    (1) (vanilla) GAN ,
    (2) Wasserstein GAN (WGAN), or
    (3) WGAN with gradient penalty (WGAN-GP, default). takes as input a file of weights or a trained model in the HDF5 format and generates an image stack in HDF5 format, with one "red" channel and up to six "green" channels that represent localization of different polarity factors in a cell at a randomly selected stage of a cell growth cycle; takes as input either HDF5 stack or an individual image file and visualizes the data stored in the input file.

To see the entire list of source files and copy it to your current folder, type:
[user@cn3200]$ ls $BIOGANS_SRC  
[user@cn3200]$ cp $BIOGANS_SRC/* . 
To list the executable files, type:
[user@cn3200]$ ls $BIOGANS_BIN
To copy sample data and checkpoints to your current directory, enter the command:
[user@cn3200]$ cp -r $BIOGANS_DATA/* .
This command will create the subfolders data and checkpoints, as well as an empty folder images in your current directory.
[user@cn3200]$ tree -l data
├ LIN_Normalized_WT_size-48-80_test -> /fdb/BioGANs/data/LIN_Normalized_WT_size-48-80_test
│   ├ Alp14
│   │   ├ cell1006.png
│   │   ├ cell10118.png
│   │   ...
│   ├ Arp3
│   │   ├ cell10056.png
│   │   ├ cell10177.png
│   │   ...
│   ├ Cki2
│   │   ├ cell1005.png
│   │   ├ cell10114.png
│   │   ...
│   ├ Mkh1
│   │   ├ cell10075.png
│   │   ├ cell10106.png
│   │   ...
│   ├ Sid2
│   │   ├ cell10009.png
│   │   ├ cell10098.png
│   │   ...
│   └ Tea1
│       ├ cell10058.png
│       ├ cell10071.png
│       ...
└ LIN_Normalized_WT_size-48-80_train -> /fdb/BioGANs/data/LIN_Normalized_WT_size-48-80_train
    ├ Alp14
    │   ├ cell100005.png
    │   ├ cell100087.png
    │   ...
    ├ Arp3
    │   ├ cell100057.png
    │   ├ cell100079.png
    │   ...
    ├ Cki2
    │   ├ cell100026.png
    │   ├ cell100044.png
    │   ...
    ├ Mkh1
    │   ├ cell100015.png
    │   ├ cell100242.png
    │   ...
    ├ Sid2
    │   ├ cell100002.png
    │   ├ cell100014.png
    │   ...
    └ Tea1
        ├ cell100077.png
        ├ cell100081.png
The BioGANs implementation on Biowulf includes only a part of the entire LIN dataset, comprising 26,909 images.
To list available command line options for the script train.pyR, type:
[user@cn3200]$ -h 
Using TensorFlow backend.
usage: [-h] [-a GAN_ALGORITHM] [-b BATCH_SIZE] -d DATAROOT
                [--epochs EPOCHS] [-g num_gpus] [--lrD LRD] [--lrG LRG]
                [-m NETWORK_MODEL] [--num_D_iters NUM_D_ITERS]
                [--num_G_iters NUM_G_ITERS] [--ngf NGF] [--ndf NDF]
                [-o OPTIMIZER] [-p POLARITY_FACTORS]
                [--random_seed RANDOM_SEED] [-w] [-z NZ]
                [--wgan_clip_value WGAN_CLIP_VALUE]
                [--wgangp_lambda WGANGP_LAMBDA] [-v]

optional arguments:
  -h, --help            show this help message and exit
  -a GAN_ALGORITHM, --gan_algorithm GAN_ALGORITHM
                        GAN algorithm to train: GAN | WGAN | WGAN-GP
  -b BATCH_SIZE, --batch_size BATCH_SIZE
                        input batch size; default=64
  --epochs EPOCHS, -e EPOCHS
                        number of iterations to train for
  -g num_gpus, --num_gpus num_gpus
                        number of gpus to use; default=1
  --lrD LRD             learning rate for discriminator/critic, default:
                        depends on gan_algorithm and optimizer
  --lrG LRG             learning rate for Generator, default: depends on
                        gan_algorithm and optimizer
  -m NETWORK_MODEL, --network_model NETWORK_MODEL
                        Architecture of the model: DCGAN | DCGAN-separable |
  --num_disc_iters NUM_DISC_ITERS
                        Number of iterations of the WGAN critic per one update
                        of generator
  --ngf NGF             number of generator filters
  --ndf NDF             number of discriminator filters
  -p checkpoint_prefix, --checkpoint_prefix checkpoint_prefix
                        prefix of the output checkpoint file .*.h5
  --random_seed RANDOM_SEED
                        Random seed, default - the answer to the ultimate
  -w, --use_pretrained_weights
                        load pre-trained model weights
  -z NZ, --nz NZ        size of the latent z vector
  --wgan_clip_value WGAN_CLIP_VALUE
                        for WGAN
  --wgangp_lambda WGANGP_LAMBDA
                        for WGAN-GP
  -v, --verbose         increase the verbosity level of output

required arguments:
                        Path to the training dataset

Here are examples of the training commands:
[user@cn3200]$ -d data 
(this command will train the default network model, DCGAN-starshaped, using the default gan algorithm, WGAN-GP, on image data corresponding to all the polarity factors available in the folder data/LIN_Normalized_WT_size-48-80_train)
[user@cn3200]$ -d data/LIN_Normalized_WT_size-48-80_train -e 2000 -a WGAN -m DCGAN-separable 
(note that running of both the commands takes long, so they would normally be submitted as a batch job.)
[user@cn3200]$ -d data/LIN_Normalized_WT_size-48-80_train/Alp14 -m DCGAN -a GAN  
(this command will train a model on the data for only one of the polarity factors).

The following options, passed to the executable, are equivalent:
-d data
-d data/LIN_Normalized_WT_size-48-80_train
-d data/LIN_Normalized_WT_size-48-80_train -p Alp14,Arp3,Cki2,Mkh1,Sid2,Tea1
These options are also equivalent:
-d data/LIN_Normalized_WT_size-48-80_train/Alp14
-d data/LIN_Normalized_WT_size-48-80_train -p Alp14
-d data -p Alp14
During the training procedure, the following output foolders will be created (if they did not exist previously) in the current working directory:
checkpoints   images
Stored in the first of the folders will be files in HDF5 format:
(1) weights of the generator, discriminatori/critic and combined models,
weights.<model type>.<network_architecture>.<gan_algorithm>.h5
which can be used by the executable to continue the training procedure from the stored checkpiont.
(2) generator model(s), which can be used as input by the executable
model.generator.<model type>.<network_architecture>.<gan_algorithm>.h5
For example:
[user@cn3200]$ ls checkpoints      
model.generator.DCGAN-separable.GAN.1.h5             weights.discriminator[1].DCGAN-starshaped.GAN.6.h5
model.generator.DCGAN-separable.WGAN.1.h5            weights.discriminator[2].DCGAN-starshaped.GAN.6.h5
weights.combined_model[0].DCGAN-separable.WGAN.1.h5  weights.discriminator[3].DCGAN-starshaped.GAN.6.h5
weights.critic[0].DCGAN-starshaped.WGAN-GP.6.h5      weights.discriminator[4].DCGAN-starshaped.GAN.6.h5
weights.critic[1].DCGAN-starshaped.WGAN-GP.6.h5      weights.discriminator[5].DCGAN-starshaped.GAN.6.h5
weights.critic[2].DCGAN-starshaped.WGAN-GP.6.h5      weights.discriminator.DCGAN-separable.GAN.1.h5
weights.critic[3].DCGAN-starshaped.WGAN-GP.6.h5      weights.generator.DCGAN-separable.GAN.1.h5
weights.critic[4].DCGAN-starshaped.WGAN-GP.6.h5      weights.generator.DCGAN-separable.WGAN.1.h5
weights.critic[5].DCGAN-starshaped.WGAN-GP.6.h5      weights.generator.DCGAN-starshaped.GAN.6.h5
weights.critic.DCGAN-separable.WGAN.1.h5             weights.generator.DCGAN-starshaped.WGAN-GP.6.h5
weights.critic.DCGAN.WGAN-GP.1.h5                    weights.generator.DCGAN.WGAN-GP.1.h5
The executable supports the following command line options:
[user@cn3200]$ -h
usage: [-h] [-b BATCH_SIZE] [-d DATAROOT] [-D] [-e] -i input_file
                  [--lrD LRD] [--lrG LRG] [-m NETWORK_MODEL] [--ndf NDF]
                  [--ngf NGF] [-o OPTIMIZER] [--random_seed RANDOM_SEED] [-v]
                  [-w] [--wgan_clip_value WGAN_CLIP_VALUE]
                  [--wgangp_lambda WGANGP_LAMBDA] [-z NZ]

optional arguments:
  -h, --help            show this help message and exit
  -b BATCH_SIZE, --batch_size BATCH_SIZE
                        input batch size; default=64
                        Path to the training dataset
  -D, --debug           output debugging info
  -e, --evaluate        evaluate predicted images by comparing with real data
  --lrD LRD             learning rate for discriminator/critic, default:
                        depends on gan_algorithm and optimizer
  --lrG LRG             learning rate for Generator, default: 0.0002
  -m NETWORK_MODEL, --network_model NETWORK_MODEL
                        network model architecture: DCGAN | DCGAN-separable |
                        DCGAN-starshaped, default = DCGAN-starshaped
  --ndf NDF             number of discriminator filters
  --ngf NGF             number of generator filters
  -o OPTIMIZER, --optimizer OPTIMIZER
                        Optimizer to use for training: default (depends on
                        gan_algorithm) | adam | rmsprop
  --random_seed RANDOM_SEED
                        Random seed, default - the answer to the ultimate
  -v, --verbose         increase the verbosity level of output
  -w, --use_pretrained_weights
                        load pre-trained model weights
  --wgan_clip_value WGAN_CLIP_VALUE
                        for WGAN
  --wgangp_lambda WGANGP_LAMBDA
                        for WGAN-GP
  -z NZ, --nz NZ        size of the latent z vector

required arguments:
  -i input_file, --input_file input_file
                        Saved generator model or weights file
To run this executable on a saved checkpoint file, type
[user@cn3200]$ -i checkpoints/model.generator.DCGAN-separable.GAN.1.h5 
Using TensorFlow backend.
Output image stack to file: images/stack2.DCGAN-separable.GAN.h5             
Now visualize the predicted images:
[user@cn3200]$ -h
usage: [-h] -i input_file [-v]
                                                                                                                        optional arguments:
  -h, --help     show this help message and exit
  -v             output debugging info

required arguments:                                                                                                       -i input_file  The file to be visualized
[user@cn3200]$ -i images/stack7.DCGAN-starshaped.WGAN-GP.h5

You can combine the and commands:
[user@cn3200]$ -i checkpoints/model.generator.DCGAN-separable.GAN.1.h5; -i images/stack7.DCGAN-starshaped.GAN.h5  

[user@cn3200]$ -i checkpoints/model.generator.DCGAN-separable.GAN.1.h5; -i images/stack7.DCGAN-starshaped.GAN.h5  

Note that every run of the command generates synthetic images corresponding to a particular stage of a cell growth cycle, which is selected randomly.

End the interactive session:
[user@cn3200 ~]$ exit
salloc.exe: Relinquishing job allocation 46116226
[user@biowulf ~]$