First run

We present here a simple example to show the basic functionalities. It uses a random input tensor, more complex examples can be explored in the other notebook.

from torch.utils.data import DataLoader
from swotted import fastSWoTTeDModule, fastSWoTTeDTrainer, fastSWoTTeDDataset
from swotted.loss_metrics import *

import matplotlib.pyplot as plt
from omegaconf import OmegaConf

We first generate a random tensor. Note that the tensor can be replaced by an irregular tensor (list of tensors of size \(N\times T_k\) with \(T_k\) that varies from one patient to the other).

K = 200  #: number of patients
N = 5  #: number of medical events
T = 10  #: length of time's stay
#generation of a random tensor
X=torch.bernoulli( torch.ones(K,N,T)*0.2 )

We now define the model. The configuration holds the main parameters of SWoTTeD. It uses OmegaConf configuration to ease the use of the dictionaries.

The SWoTTeD module is implemented as PyTorch Lightning module, and it has to be used with the corresponding trainer. Indeed, SWoTTeD optimisation problem is a not a classical supervised task and the optimization problem has to be set with the knowledge of the size of the dataset.

params = {}
params["model"] = {}
params["model"]["non_succession"] = 0.5
params["model"]["sparsity"] = 0.5
params["model"]["rank"] = 4  #: number of phenotypes to discover
params["model"]["twl"] = 3  #: length of time's window
params["model"]["N"] = N
params["model"]["metric"] = "Bernoulli"
params["training"] = {}
params["training"]["batch_size"] = 40
params["training"]["nepochs"] = 100
params["training"]["lr"] = 1e-2
params["predict"] = {}
params["predict"]["nepochs"] = 100
params["predict"]["lr"] = 1e-2

config = OmegaConf.create(params)

# define the model
swotted = fastSWoTTeDModule(config)

train_loader = DataLoader(
    fastSWoTTeDDataset(X),
    batch_size=params["training"]["batch_size"],
    shuffle=False,
    collate_fn=lambda x: x,
)

# train the model
trainer = fastSWoTTeDTrainer(
    max_epochs=params["training"]["nepochs"]
)

Then, we run the training of the model.

trainer.fit(model=swotted, train_dataloaders=train_loader)

Finally we can explore the results: * visualize the extracted phenotypes * make reconstruction of input tensors * …

# visualize the phenotypes
Ph = swotted.Ph.flip(2).detach().numpy()
for i in range(R):
    plt.imshow(Ph[i], vmin=0, vmax=1, cmap="binary",interpolation='none')
    plt.ylabel("Drugs")
    plt.xlabel("Time")
    plt.title("Phenotype")
    plt.show()

# make decomposition with the train model: it projects the X on the phenotypes of the model
# note that projection can be applied on new data. We use data from the training set for the sake
# of simplicity
id = 15
W = swotted(X[id, :, :].unsqueeze(0))
# Apply reconstruction
Y = swotted.model.reconstruct(W,swotted.Ph)

# Patient decomposition
plt.subplot(211)
plt.imshow(X[id].detach().numpy(), vmin=0, vmax=1, cmap="binary",interpolation='none')
plt.ylabel("Drugs")
plt.xlabel("Time")
plt.title("Input matrix")
plt.subplot(212)
plt.imshow(Y[0].detach().numpy(), vmin=0, vmax=1, cmap="binary",interpolation='none')
plt.ylabel("Drugs")
plt.xlabel("Time")
plt.title("Pathway")
plt.title("reconstruction (not reordered)")
plt.show()