output_bias = True
mlp = tt.practical.MLPVanilla(X_train_std.shape[1],
[n_nodes for layer_idx
in range(n_layers)],
X_train_std.shape[1],
batch_norm,
dropout,
output_bias=output_bias)
net = nn.Sequential(ResidualBlock(X_train_std.shape[1], mlp),
Scaler(X_train_std.shape[1]))
optimizer = tt.optim.Adam(lr=lr)
if num_durations > 0:
labtrans = NKSDiscrete.label_transform(num_durations)
y_train_discrete = labtrans.fit_transform(*y_train.T)
surv_model = NKSDiscrete(net, optimizer,
duration_index=labtrans.cuts)
else:
surv_model = NKS(net, optimizer)
model_filename = \
os.path.join(output_dir, 'models',
'%s_%s_exp%d_bs%d_nep%d_nla%d_nno%d_lr%f_nd%d_test.pt'
% (survival_estimator_name, dataset, experiment_idx,
batch_size, n_epochs, n_layers, n_nodes, lr, num_durations))
assert os.path.isfile(model_filename)
if not os.path.isfile(model_filename):
print('*** Fitting with hyperparam:', hyperparam, flush=True)
if num_durations > 0:
'lr%f_test.pt' % lr)
if not os.path.isfile(emb_model_filename):
emb_model.fit(X_train_std, mds_embedding,
batch_size=batch_size, epochs=100,
verbose=False)
emb_model.save_net(emb_model_filename)
else:
emb_model.load_net(emb_model_filename)
emb_model.net.train()
print('*** Fine-tuning with DKSA...')
torch.manual_seed(fine_tune_random_seed + 1)
np.random.seed(fine_tune_random_seed + 1)
optimizer = tt.optim.Adam(lr=lr)
if num_durations > 0:
labtrans = NKSDiscrete.label_transform(num_durations)
y_train_discrete = labtrans.fit_transform(*y_train.T)
surv_model = NKSDiscrete(emb_model.net, optimizer,
duration_index=labtrans.cuts)
else:
surv_model = NKS(emb_model.net, optimizer)
model_filename = \
os.path.join(output_dir, 'models',
'%s_%s_exp%d_mf%d_msl%d_km%d_'
% (survival_estimator_name, dataset,
experiment_idx, max_features,
min_samples_leaf, use_km)
+
'bs%d_nep%d_nla%d_nno%d_'
% (batch_size, n_epochs, n_layers, n_nodes)
output_bias = True
optimizer = tt.optim.Adam(lr=lr)
mlp = tt.practical.MLPVanilla(fold_X_train_std.shape[1],
[n_nodes for layer_idx
in range(n_layers)],
fold_X_train_std.shape[1],
batch_norm,
dropout,
output_bias=output_bias)
net = nn.Sequential(ResidualBlock(fold_X_train_std.shape[1],
mlp),
Scaler(fold_X_train_std.shape[1]))
if num_durations > 0:
labtrans = NKSDiscrete.label_transform(num_durations)
fold_y_train_discrete \
= labtrans.fit_transform(*fold_y_train.T)
surv_model = NKSDiscrete(net, optimizer,
duration_index=labtrans.cuts)
else:
surv_model = NKS(net, optimizer)
model_filename = \
os.path.join(output_dir, 'models',
'%s_%s_exp%d_bs%d_nep%d_nla%d_nno%d_'
% (survival_estimator_name, dataset,
experiment_idx, batch_size, n_epochs,
n_layers, n_nodes)
+
'lr%f_nd%d_cv%d.pt'
y_train = y_train.astype('float32')
y_test = y_test.astype('float32')
torch.manual_seed(method_random_seed)
torch.cuda.manual_seed_all(method_random_seed)
np.random.seed(method_random_seed)
batch_norm = True
dropout = 0.0
output_bias = False
net = nn.Sequential(DiagonalScaler(X_train_std.shape[1]))
optimizer = tt.optim.Adam(lr=lr)
if num_durations > 0:
labtrans = NKSDiscrete.label_transform(num_durations)
y_train_discrete = labtrans.fit_transform(*y_train.T)
surv_model = NKSDiscrete(net, optimizer, duration_index=labtrans.cuts)
else:
surv_model = NKS(net, optimizer)
model_filename = \
os.path.join(output_dir, 'models',
'%s_%s_exp%d_%s_bs%d_nep%d_lr%f_nd%d_test.pt'
% (survival_estimator_name, dataset, experiment_idx,
val_string, batch_size, n_epochs, lr, num_durations))
assert os.path.isfile(model_filename)
print('*** Loading ***', flush=True)
surv_model.load_net(model_filename)
if num_durations > 0: