phiml's Introduction

phiml: A Machine Learning approach to Potential Dynamics

TODO

Remove conditionals in simulate_forward
No bias for signal mapping (Completed)
Ensure noise is not too great
Constrained landscape. Repellor at infinity.

Ideas

How to handle noise.
Infer noise parameter. (Completed-ish)
Incorporate the number of desired fixed points
Precompute summary stat for $x_1$ data
Adjoint method.
Symmetry in the order of the cells within the population: $\boldsymbol{x}_i$ (Completed?)
Parallelize batches. (Completed).
Parallelize individual simulations. (Completed)
Customizable layer architecture
batch normalization and dropout
Autocorrelation time of individual cells to determine transitioning flag.
Softmax activation prevents super-linear growth in the potential function?
Normalize data beforehand?

Installation

For the CPU:

mamba create -p <env-path> python=3.8 pytorch=1.11 numpy=1.24 matplotlib= 3.7 pytest=7.4 tqdm ipykernel ipywidgets

For the GPU, specifying cuda toolkit 11.2:

mamba create -p <env-path> python=3.8 pytorch=1.11[build=cuda112*] numpy=1.24 matplotlib= 3.7 pytest=7.4 tqdm ipykernel ipywidgets

For M1 Macs, where we want the MPS device:

mamba create -p <env-path> python=3.8 pytorch=2.1 numpy=1.24 matplotlib= 3.7 pytest=7.4 tqdm ipykernel ipywidgets
mamba activate <env-path>
mamba install -c pytorch pytorch=2.1 torchvision

Then, to install the project phiml,