num_embeddings,
embedding_dims,
num_nodes,
out_features,
batch_norm=batch_norm,
dropout=dropout,
output_bias=output_bias)
net = net.to(device)
if args.optimizer == 'AdamWR':
model = CoxPH(net,
optimizer=tt.optim.AdamWR(
lr=args.lr,
decoupled_weight_decay=args.weight_decay,
cycle_eta_multiplier=0.8))
lrfinder = model.lr_finder(x_train, y_train, batch_size, tolerance=10)
lr = lrfinder.get_best_lr()
model.optimizer.set_lr(lr)
wandb.init(
project=args.dataset + '_baseline',
group='deepsurv' + '_' + args.optimizer,
name=
f'L{args.num_layers}N{args.num_nodes}D{args.dropout}W{args.weight_decay}B{args.batch_size}',
config=args)
wandb.watch(net)
# Loss configuration ============================================================
# Training ======================================================================
x_mapper = DataFrameMapper(standardize + leave)
X_train = pd.DataFrame(data=x_mapper.fit_transform(X_train),
columns=numerical_columns+categorical_columns,
index=X_train.index)
# x_test = x_mapper.transform(X_test)
in_features = X_train.shape[1]
num_nodes = [32, 16, 8]
out_features = 1
batch_norm = True
dropout = 0.1
output_bias = False
net = tt.practical.MLPVanilla(in_features, num_nodes, out_features, batch_norm,
dropout, output_bias=output_bias)
from pycox.models import CoxPH
model = CoxPH(net, tt.optim.Adam)
batch_size = 256
lrfinder = model.lr_finder(X_train.values.astype('float32'), y_train, batch_size, tolerance=10)
_ = lrfinder.plot()
tt.callbacks.EarlyStopping(patience=20)
plt.show()
print()