def create_model(input_len, n_neurons1, n_neurons2, n_neurons3):
# Define neural networks for processing of data before and after aggregation
prepNN1 = torch.nn.Sequential(
torch.nn.Linear(input_len, n_neurons1, bias=True),
torch.nn.ReLU(),
torch.nn.Linear(n_neurons1, n_neurons2, bias=True),
torch.nn.ReLU(),
)
afterNN1 = torch.nn.Sequential(torch.nn.Identity())
prepNN2 = torch.nn.Sequential(
torch.nn.Linear(n_neurons2, n_neurons3, bias=True),
torch.nn.ReLU(),
)
afterNN2 = torch.nn.Sequential(torch.nn.Linear(n_neurons3, 1),
torch.nn.Tanh())
# Define model ,loss function and optimizer
model = torch.nn.Sequential(
mil.BagModel(prepNN1, afterNN1, torch.mean, device),
mil.BagModel(prepNN2, afterNN2, torch.mean, device)).double()
criterion = mil.MyHingeLoss()
optimizer = torch.optim.Adam(model.parameters(),
lr=learning_rate,
weight_decay=weight_decay)
return model
for n_neurons3 in n_neurons3_grid:
for learning_rate in learning_rate_grid:
for weight_decay in weight_decay_grid:
config['n_neurons1'] = n_neurons1
config['n_neurons2'] = n_neurons2
config['n_neurons3'] = n_neurons3
config['learning_rate'] = learning_rate
config['weight_decay'] = weight_decay
print('INFO: Running cross validation with config:\n{}'.
format(config))
# --- MODEL ---
model = create_model(len(dataset.data[0]), n_neurons1,
n_neurons2, n_neurons3)
criterion = mil.MyHingeLoss()
optimizer = torch.optim.Adam(model.parameters(),
lr=learning_rate,
weight_decay=weight_decay)
# Move model to gpu if available
model = model.to(device)
avg_loss = train_utils.k_fold_cv(
model=model,
fit_fn=train_utils.train_model,
criterion=criterion,
optimizer=optimizer,
dataset=dataset,
train_indices=train_indices,
epochs=epochs,