data_dir = "/public/WORK_backup/caizefeng/Datasets/STO_600_cut9_gauss16"
test_file_path = "/public/WORK_backup/caizefeng/Datasets/STO_600_cut9_gauss16/test/10099.npy"
net = SVTNetCharge(num_element=3,
sigma_size=16,
**net_name_parse(net_file_name,
is_temp=False)).to(torch.float64)
net.load_state_dict(torch.load(os.path.join("nets", net_file_name)))
# net = torch.load(os.path.join("nets", net_file_name)).cpu()
net.eval()
data = np.load(test_file_path)
# load pre-calculated mean and std to make the prediction
train_mean, train_std = standardization2D(read_saved=True,
data_path=data_dir)
charge_dft = data[:, -1]
with torch.no_grad():
charge_pred = net(
(torch.from_numpy(data[:, :-1]) - train_mean) / train_std).numpy()
# plot
plt.figure(figsize=(10, 10))
plt.scatter(charge_pred, charge_dft, c=charge_pred, cmap='viridis', s=10)
plt.plot(np.linspace(0, 16, 50),
np.linspace(0, 16, 50),
ls="dashed",
c="grey")
plt.xlabel(r'$\rho_{\,\mathrm{pred}}$', fontsize=24)
plt.ylabel(r'$\rho_{\,\mathrm{scf}}$', fontsize=24)
# plt.title("Logrithm parity plot for the deep learning vs DFT charge density prediction", fontsize=16)
# validation and testing dataloader
dataset_test_list = []
for test_feature_file, test_ldos_file in zip(test_feature_path_list,
test_ldos_path_list):
all_data_test = torch.from_numpy(np.load(test_feature_file))
ldos_data_test = torch.from_numpy(np.load(test_ldos_file))
dataset_test_list.append(
TensorDataset(all_data_test[:, :-1],
ldos_data_test[:, :net_hp["num_windows"]]))
test_iter = DataLoader(ConcatDataset(dataset_test_list),
**dataload_hp_test)
# calculate mean and std over the whole training set
train_mean, train_std = standardization2D(
read_saved=True,
data_path=data_dir,
num_feature=num_feature,
train_path_list=train_feature_path_list)
# training and validiting
training_loss, batch_count = 0.0, 0
train_path_list = [
x for x in zip(train_feature_path_list, train_ldos_path_list)
]
train_scalar_list = []
test_scalar_list = []
net.to(device)
for epoch in range(train_hp["num_epoch"]):
random.shuffle(train_path_list)
dataset_test_list.append(dataset)
dataset_test = ConcatDataset(dataset_test_list)
test_iter = DataLoader(dataset_test, **dataload_hp_test)
# training dataset
dataset_train_list = []
for i, train_file in enumerate(train_path_list):
dataset = MmapDataset2D(train_file, num_column)
dataset_train_list.append(dataset)
dataset_train = ConcatDataset(dataset_train_list)
train_iter = DataLoader(dataset_train, **dataload_hp)
# calculate mean and std over the whole training set
train_iter_more = DataLoader(dataset_train, **dataload_hp_test)
train_mean, train_std = standardization2D(read_saved=True,
data_path=data_dir,
num_feature=num_feature,
train_iter_mmap=train_iter_more)
# training
training_loss, batch_count = 0.0, 0
for epoch in range(train_hp["num_epoch"]):
for i, (X, y) in enumerate(train_iter):
# TensorBoard Graph
if i == 0 and epoch == 0:
train_writer.add_graph(net, X)
net.to(device)
X = ((X - train_mean) / train_std).to(device)
y = y.to(device)
optimizer.zero_grad()
# forward + backward + optimize