Tensorflow implementation of Spike-GAN, which allows generating realistic patterns of neural activity whose statistics approximate a given traing dataset (Molano-Mazon et al. 2018 ICRL2018).
- Python ≥ 3.5 (tested with 3.8)
- Tensorflow 2.x (tested with 2.5)
- Numpy ≥ 1.9
- SciPy
- Matplotlib
- Seaborn
Just download the repository to your computer and add the Spike-GAN folder to your path.
The folder named data contains the retinal data used for Fig. 3 in the ICLR paper. The whole data set can be found in Marre et al. 2014. The folder also contains the data generated by the k-pairwise and the Dichotomized-Gaussian models for that same figure and for Fig. S6.
This code uses Tensorflow 1 under the hood (through TensorFlow 2's
compatibility layer), so the easiest way to force TF to ignore your
GPUs, if you so wish, is to set the CUDA_VISIBLE_DEVICES environment
variable to empty when calling the main_conv.py script, like so:
python CUDA_VISIBLE_DEVICES="" main_conv.py ...
Alternatively, you can add/uncomment the line
os.environ["CUDA_VISIBLE_DEVICES"] = "" at the top of the
main_conv.py file.
Spike-GAN can be run using this data with the command:
python3.5 main_conv.py --architecture='conv' --dataset='retina' --num_bins=32 --iteration='test' --num_neurons=50 --is_train --num_layers=2 --num_features=128 --kernel_width=5 --data_instance='1'
The example below will train Spike-GAN with the semi-convolutional architecture on a simulated dataset containing the activity of two correlated neurons whose firing rate follows a uniform distribution across 12 ms. See main_conv.py for more options on the type of simulated activity (refractory period, firing rate...).
python3.5 main_conv.py --is_train --architecture='conv' --dataset='uniform' --num_bins=12 --num_neurons=2
In this example we train a fully-connected architecture on the toy
maxent dataset (which contains 12 neurons). The fully connected
network has two hidden layers with 256 units each. The number of
synthetic samples generated every time we compute diagnostics (such as
when we pause training and monitor its advancement, or at the end of
training) is 1000, controlled by --num_samples. Training will be run
for 100000 iterations (controlled by --num_iter). The arbitrary name
training_run_name, controlled by --iteration will be added to the
data folder name for that training run, which could be useful if you
are planning to train the network from scratch multiple times with the
same settings. Typically the --iteration flag is not needed, and can
be safely left out.
python main_conv.py --is_train --architecture='fc' --dataset='maxent' --num_bins=1 --num_neurons=12 --num_layers=2 --num_units=256 --num_samples=1000 --num_iter=100000 --iteration='training_run_name'
This work has received funding from the European Union's Horizon 2020 research and innovation programme under the Marie Sklodowska-Curie grant agreement No 699829 (ETIC).
