def cal_R2():
x_r2, y_r2, t_r2 = load_tensor(test_data_dir)
y_r2 = y_r2.numpy()
print('Loaded R2 data!')
ntimes = len(args.times)
n_test = args.n_test
y_sum = np.full((ntimes, 2, 50, 50), 0.0)
for i in range(n_test):
y = y_mc[i * ntimes:(i + 1) * ntimes, :2]
y_sum = y_sum + y
y_mean = y_sum / n_test
nominator = 0.0
denominator = 0.0
for i in range(n_test):
print(i)
x = x_r2[i * ntimes:(i + 1) * ntimes]
t = t_r2[i * ntimes:(i + 1) * ntimes]
y = y_r2[i * ntimes:(i + 1) * ntimes, :2]
x, t = x.to(device), t.to(device)
model.eval()
with th.no_grad():
y_pred = model(x, t)
y_pred = y_pred.data.cpu().numpy()
nominator = nominator + ((y - y_pred[:, :2])**2).sum()
denominator = denominator + ((y - y_mean)**2).sum()
R2 = 1 - nominator / denominator
print("R2: {}".format(R2))
return R2
def predict():
# predict the output images
interp_x, interp_y, interp_times = load_tensor(args.data_dir +
'/lhs50_t3_expanded.hdf5')
ntimes = 3
y_DCEDN = np.full((50, 1, 50, 50), 0.0)
for i in range(50):
x = interp_x[i * ntimes:(i + 1) * ntimes]
t = interp_times[i * ntimes:(i + 1) * ntimes]
x, t = x.to(device), t.to(device)
model.eval()
y_pred = model(x, t)
y_pred = y_pred.data.cpu().numpy()
y_DCEDN[i] = y_pred[ntimes - 1, [2]]
hf = h5py.File(
args.output_dir + '/N_{}_SgLabel_DCEDN.hdf5'.format(args.n_train), 'w')
hf.create_dataset('dataset', data=y_DCEDN, dtype='f', compression='gzip')
hf.close()
bottleneck=args.bottleneck,
time_channels=args.zt).to(device)
print(model)
# load data
h5py_dir = args.data_dir
train_data_dir = h5py_dir + '/lhs{}_t{}_expanded.hdf5'.format(
args.n_train, len(args.times))
test_data_dir = h5py_dir + '/lhs{}_t{}_expanded.hdf5'.format(
args.n_test, len(args.times))
interp_data_dir = h5py_dir + '/lhs{}_t{}.hdf5'.format(
400, 3) # data used to plot the results at t= 5, 50, and 100 days
train_loader, train_stats = load_data_dynamic(train_data_dir, args.batch_size)
test_loader, test_stats = load_data_dynamic(test_data_dir,
args.test_batch_size)
interp_x, interp_y, interp_times = load_tensor(interp_data_dir)
print('Loaded data!')
if args.loss_fn == "mse":
loss_fn = nn.MSELoss(reduction='sum')
elif args.loss_fn == "l1":
loss_fn = nn.L1Loss()
elif args.loss_fn == 'huber':
loss_fn = nn.SmoothL1Loss()
optimizer = optim.Adam(model.parameters(),
lr=args.lr,
weight_decay=args.weight_decay)
scheduler = ReduceLROnPlateau(optimizer,
mode='min',
factor=0.1,