Introduction

fib_tf is a Python framework developed on the top of TensorFlow for 2D cardiac electrophysiology simulation. While TensorFlow is primarily designed for machine learning applications, it also provides a general framework for performing multidimensional tensor manipulation. The primary goal of fib_tf is to test and assess the suitability of TensorFlow for solving systems of stiff ordinary differential equations (ODE) needed for cardiac modeling.

Programming Graphics Processing Units (GPU) using C/C++ CUDA or OpenCL is not a trivial task and is both time-consuming and error-prone. This is where TensorFlow can be very useful. Because TensorFlow takes care of the lower level details of running the model on a GPU, the programmer can focus on the high-level model and be more productive. Additionally, TensorFlow allows running the model on multiple GPUs or even clusters. While it is not expected that a TensorFlow ODE solver beats a handcrafted optimized CUDA code, we hope to get a reasonable performance. The main question we are trying to answer is whether this performance is good enough.

Requirements:

We have tested the software on Ubuntu 14.04 and 16.04 machines with Tensorflow 1.4, Python 3.6, and GTX 1080 or GTX Titan X GPUs (CUDA 8.0, CUDNN 7.0).

The following software packages are needed:

1. Python 3 (sorry, no python 2.7!)
2. Numpy
3. SDL2 (no need for PySDL2, just the basic SDL2 library). You can download it here.
4. TensorFlow
5. libcupti (optional, it is needed for profiling)

It should be noted the TensorFlow installables come in different flavors. fib_tf can work on CPU only (after changing the line tf.device('/device:GPU:0'): to tf.device('/device:CPU:0'):), but it would be too slow.

However, just having a GPU-enabled version of Tensorflow is not enough either, rather you need a TensorFlow version with XLA JIT enabled to have the best performance. See details for discussion on why this is so important. Fortunately, TensorFlow has an experimental Just In Time (JIT) compiler that allows fusing multiple kernels into one. But this feature is still not available in the stock versions of TensorFlow. To enable it, you need to compile Tensorflow from source. This is not the easiest installation process, but is very well doable as long as you just follow the instructions!

Download

Download fib_tf as

git clone https://github.com/siravan/fib_tf.git

Change into the fib_tf directory and run the models.

Main Files

• screen.py: provides a simple SDL2 screen to plot a 2D numpy array in grayscale.
• dll.py: a helper copied file from PySDL2 to find and load the appropriate SDL2 library (libSDL2.so in Linux).
• ionic.py: contains the base class for implementing ionic models.
• fenton.py: implements the 4-variable Cherry-Ehrlich-Nattel-Fenton canine left-atrial model (4v model).
• fenton_simple.py: a basic implementation of the 4v model (see details.html).
• fenton_jit.py: an intermediate implementation of the 4v model (see details.html).
• br.py: implements the 8-variable modified Beeler-Reuter model.
• README.md: this file!
• details.html: A detailed discussion of the software and techniques used.

Test

You can run the Cherry-Ehrlich-Nattel-Fenton (4v) canine left-atrial model as

python3 fenton.py

and the 8-variable modified Beeler-Reuter model as

python3 br.py

Documentation

The details of the ionic models and various methods and tricks used to improve the performance are discussed here.

To run fib_tf models interactively, you first need to import the desired model. For the 4v model, it is imported as import fenton. Next, we need to define a configuration dict. For the 4v model, it is defined like

python config = { 'width': 512, # screen width in pixels 'height': 512, # screen height in pixels 'dt': 0.1, # integration time step in ms 'dt_per_plot' : 10, # screen refresh interval in dt unit 'diff': 1.5, # diffusion coefficient 'duration': 1000, # simulation duration in ms 'timeline': False, # flag to save a timeline (profiler) 'timeline_name': 'timeline_4v.json', 'save_graph': True # flag to save the dataflow graph }

For the Beeler-Reuter model, we import it as import br and define the configuration dict as:

python config = { 'width': 512, # screen width in pixels 'height': 512, # screen height in pixels 'dt': 0.1, # integration time step in ms 'dt_per_plot': 10, # screen refresh interval in dt unit 'diff': 0.809, # diffusion coefficient 'duration': 1000, # simulation duration in ms 'skip': False, # optimization flag: activate multi-rate 'cheby': True, # optimization flag: activate Chebysheb polynomials 'timeline': False, # flag to save a timeline (profiler) 'timeline_name': 'timeline_br.json', 'save_graph': False # flag to save the dataflow graph }

Then, we need to create a model object. For example:

python model = br.BeelerReuter(config)

At a minimum, we construct a dataflow graph by calling model.define() and run it as:

python for i in model.run(): pass

But this is not very interesting, as there is no visual output. So, let's define a screen object:

python from screen import Screen ... im = Screen(model.height, model.width, 'Beeler-Reuter Model')

Now, we can run the model and see what happens:

python for i in model.run(im): pass

A single planar wave! Better, but still not very interesting. We can induce a spiral wave by an S1-S2 cross-stimulation protocol. S1 is implicit, but we need to define S2:

python model.add_pace_op('s2', 'luq', 10.0)

This define an S2 stimulation operation, where the value of the transmembrane voltage over the left upper quadrant (luq) of the screen is set to 10 mV. Now, we modify the main loop to

```python s2 = model.millisecondtostep(300) # 300 ms

for i in model.run(im): if i == s2: model.fire_op('s2') ```

This time we get a nice spiral wave. Note that we need to convert the desired time of S2 to the number of steps returns from model.run().

We can do even better. Let's add a circular obstacle to the medium by adding this line before calling model.define():

python model.add_hole_to_phase_field(150, 200, 40) # center=(150,200), radius=40

Success! We get a spiral wave that anchors to the obstacle:

Let's put all together:

```python import br from screen import Screen

config = { 'width': 512, # screen width in pixels 'height': 512, # screen height in pixels 'dt': 0.1, # integration time step in ms
'dtperplot': 10, # screen refresh interval in dt unit 'diff': 0.809, # diffusion coefficient 'duration': 1000, # simulation duration in ms 'skip': False, # optimization flag: activate multi-rate 'cheby': True, # optimization flag: activate Chebysheb polynomials 'timeline': False, # flag to save a timeline (profiler) 'timelinename': 'timelinebr.json', 'save_graph': False # flag to save the dataflow graph }

model = br.BeelerReuter(config) model.addholetophasefield(150, 200, 40) # center=(150,200), radius=40 model.define() model.addpaceop('s2', 'luq', 10.0) im = Screen(model.height, model.width, 'Beeler-Reuter Model')

s2 = model.millisecondtostep(300) # 300 ms

for i in model.run(im): if i == s2: model.fire_op('s2')

```

@2017-2018 Shahriar Iravanian