test_loader = data.create_dataset('test', seed, realizations, root_bn, root_Pk,
batch_size)
# get the number of elements in the test set
size = 0
for x, y in test_loader:
size += x.shape[0]
# define the array with the results
pred = np.zeros((size, 5), dtype=np.float32)
true = np.zeros((size, 5), dtype=np.float32)
# get the parameters of the trained model
#model = architecture.model_1hl(bins_SFRH, h1, 6, dropout_rate)
#model = architecture.model_2hl(bins_SFRH, h1, h2, 6, dropout_rate)
model = architecture.model_3hl(512 + 158, h1, h2, h3, 5, dropout_rate)
model.load_state_dict(torch.load(fmodel))
model.to(device=device)
# loop over the different batches and get the prediction
offset = 0
model.eval()
for x, y in test_loader:
with torch.no_grad():
x = x.to(device)
y = y.to(device)
y_NN = model(x)
length = x.shape[0]
pred[offset:offset + length] = y_NN.cpu().numpy()
true[offset:offset + length] = y.cpu().numpy()
offset += length
test_loader = data.create_dataset('test', seed, realizations, root_in,
bins_SFRH, sim, batch_size, root_out)
# get the number of elements in the test set
size = 0
for params_test, SFRH_test in test_loader:
size += params_test.shape[0]
# define the array with the results
pred = np.zeros((size, 6), dtype=np.float32)
true = np.zeros((size, 6), dtype=np.float32)
# get the parameters of the trained model
#model = architecture.model_1hl(bins_SFRH, h1, 6, dropout_rate)
#model = architecture.model_2hl(bins_SFRH, h1, h2, 6, dropout_rate)
model = architecture.model_3hl(bins_SFRH, h1, h2, h3, 6, dropout_rate)
model.load_state_dict(torch.load(fmodel))
model.to(device=device)
# loop over the different batches and get the prediction
offset = 0
model.eval()
for params_test, SFRH_test in test_loader:
with torch.no_grad():
#params_test = params_test.to(device)
SFRH_test = SFRH_test.to(device)
params_pred = model(SFRH_test)
length = params_test.shape[0]
pred[offset:offset + length] = params_pred.cpu().numpy()
true[offset:offset + length] = params_test.numpy()
offset += length
# define loss function
criterion = nn.MSELoss()
# get train, validation, and test sets
print('preparing dataset loaders')
train_loader = data.create_dataset('train', seed, realizations, root_in,
bins_SFRH, sim, batch_size, root_out)
valid_loader = data.create_dataset('valid', seed, realizations, root_in,
bins_SFRH, sim, batch_size, root_out)
test_loader = data.create_dataset('test', seed, realizations, root_in,
bins_SFRH, sim, batch_size, root_out)
# define architecture
#model = architecture.model_1hl(6, h1, bins_SFRH, dr)
model = architecture.model_2hl(6, h1, h2, bins_SFRH, dr)
model = architecture.model_3hl(6, h1, h2, h3, bins_SFRH, dr)
model.to(device=device)
network_total_params = sum(p.numel() for p in model.parameters())
print('total number of parameters in the model = %d' % network_total_params)
# define optimizer and scheduler
optimizer = torch.optim.Adam(model.parameters(),
lr=lr,
betas=(0.5, 0.999),
weight_decay=wd)
# load best-model, if it exists
if os.path.exists(fmodel):
print('Loading model...')
model.load_state_dict(torch.load(fmodel))
model.to(device=device)