Add a New Model

Adding a new model would require several steps to be followed so that it will be comaptibale with the deep_nilmtk toolkit and take advanatge of all features included in the tool.

1. Model definition

The neural network architecture is defined as a usual PyTorch model with two extra mandatory function: the step funtion and the prediction. The first one calculates the predictions of for a batch of data and return the loss. It used during training. The second is the testing loop and uses pre-trained model to generate predictions for the testing data.

Note

For Seq-to-Seq models, the aggregation of sequences is performed at the end of the predict function.

import torch.nn as nn
import torch.nn.functional as F

class NewModel(nn.Module):
    def __init__(self):
        # Model definition

    def step(self, batch):
        x, y  = batch
        out   = self(x)  # BxCxT

        error = (y - out)
        loss = F.mse_loss(out, y)
        mae = error.abs().data.mean()
        return  loss, mae

    def predict(self,  model, test_dataloader):
        ...
        results = {"pred":pred}
        return results

2. DataLoader Definition

The corresponding dataLoader of the model is defined as standard PyTorch dataset. The user has the choice of defining a new class or using existing data loaders.

import torch

class NewDataLoader(torch.utils.data.Dataset):
    def __init__(self, inputs,  targets=None,  params= {}):
        # Defines the padding of the daa and the targets Normalization
    def __len__(self):
        # Denotes the total number of samples
    def get_sample(self, index):
        # Defines how samples as generated
    def __getitem__(self, index):
        # Generate a sample of data

3. Adding the model to the configurationof the tool

The model and the loader should be linked together in the config/nilm_models.py file as follows:

NILM_MODELS = {
    'new_model_name':{
         'model': NewModel,
         'loader': NewDataLoader,
         'extra_params':{
         }
     },
}

At this point, the model is ready for experimenting. It can be tested directly using the NILMtk-API as the already available baselines.