pygot.tools.traj.fit_velocity_model

pygot.tools.traj.fit_velocity_model(adata, time_key, embedding_key, landmarks=False, device=None, graph_key=None, n_neighbors=50, v_centric_iter_n=1000, v_centric_batch_size=256, add_noise=True, sigma=0.1, lr=0.005, path='', linear=False, distance_metrics='SP', pretrained_model=None, x_centric=True, x_centric_iter_n=2000, x_centric_batch_size=256, reverse_schema=True, time_varying=True, filtered=True, **kwargs)

Estimates velocities and fit trajectories in latent space.

Arguments:

adata: AnnData

Annotated data matrix.

time_key: str

Name of time label in adata.obs

embedding_key: `str’

Name of latent space to fit, in adata.obsm

landmarks: bool

Use landmarks to approximate graphical paths

graph_key: str (default: None)

Name of graph to fit, in adata.obsm

device: device

torch device

n_neighbors: int (default: 50)

Neighbors number in kNN

v_centric_iter_n: int (default: 1000)

Iteration number of v-centric training

v_centric_batch_size: int (default: 256)

Batch size of v-centric training. Note: increase will dramatically increase training time due to the complexity of OT

add_noise: bool (default: True)

Assumption of gaussian distribution of velocity

sigma: float (default: 0.1)

The variance of gaussian distribution of velocity

distance_metrics: ‘SP’ or ‘L2’ (default: ‘SP’)

Distance metrics in optimal transport (shortest path distance or euclidean distance)

pretrained_model: :class`~torch.nn.Module` (default: None)

Training for pretraiened model

path: str (default: None)

Dir path to store shorest path file

linear: bool (default: False)

Accept linear path or not

filtered: bool (default: True)

Use filtered velocity or not

x_centric: bool (default: True)

Do x-centric training

x_centric_iter_n: int (default: 2000)

Iteration number of x-centric training

x_centric_batch_size: int (default: 256)

Batch size of x-centric training

reverse_schema: bool (default: True)

Simulation trajectory from end to start also (e.g. day7 -> day0), in x-centric training

time_varying: bool (default: True)

The neural network model use time label as input or not

lr: float (default: 5e-3)

Learning rate

returns:

• model (:class`~ODEwrapper`) – velocity model

• history (list) – training history