def train_nn(feats, labels, net_params, model_out, dir_out):
'''
Train neural network classifier
:param feats: training feats
:param labels: training labels
:param net_params: input params for neural net
:param model_out: fname to save output model
:param dir_out: directory to save output model
:return: trained model
'''
net = FFNNet(**net_params)
ds_train = SparseDataset(feats, labels)
train_loader = get_dataloader(ds_train, shuffle=True)
model = Model(net)
net = model.train(train_loader)
state = {
'net_params': net_params,
'state_dict': net.state_dict(),
}
save_state_dict(state, fname = model_out, dir_out = dir_out)
return net
def pred_prob_nn(net, x_test):
'''
:param net: trained Pytorch model
:param x_test: test feature matrix
:return: Prediction probability of all output classes for all instances (n_inst, n_classes)
'''
ds = SparseDataset(x_test)
dataloader = get_dataloader(ds, shuffle=False)
model = Model(net)
return model.predict_prob(dataloader)
def get_feat_sensitivity(net, data):
'''
Calculate feature sensitivity scores from a trained neural network.
:param net: Trained network
:param data: Data (containing feature values of different instances)
:return: Sensitivity scores of all the features in the data based on the trained model (1D array),
:return: 1D array with indices of the top scoring features in decreasing order
'''
ds = SparseDataset(data)
dataloader = get_dataloader(ds, shuffle=False)
gradients, max_grad_idx = get_max_rms_gradient(net, dataloader)
return gradients, max_grad_idx
def get_gradients(net, feats, labels=None):
'''
Compute gradients of output wrt input for every instance and feature pair
:param net: trained model
:param feats: original feats
:param labels: labels to get gradients for
:return: reweighed feature matrix (supports scipy sparse) and the training gradients
'''
ds = SparseDataset(feats, labels)
dataloader = get_dataloader(ds, shuffle=False)
gradients = get_out_gradient(net, dataloader)
return gradients
def pred_nn(net, x_test):
'''
Get predicted classes and probability of prediction given a Pytorch model and test features
:param net: Trained Pytorch model
:param x_test: test feature matrix
:return: Arrays of predicted classes, and prediction probabilities of all classes for all instances
'''
ds = SparseDataset(x_test)
dataloader = get_dataloader(ds, shuffle=False)
model = Model(net)
return model.predict(dataloader)