def test_iso_gcn_rotation_thermal_stress_dict_input_list_output(self):
main_setting = setting.MainSetting.read_settings_yaml(
Path('tests/data/rotation_thermal_stress/iso_gcn_dict_input.yml'))
tr = trainer.Trainer(main_setting)
if tr.setting.trainer.output_directory.exists():
shutil.rmtree(tr.setting.trainer.output_directory)
loss = tr.train()
np.testing.assert_array_less(loss, 5.)
# Confirm inference result has isometric equivariance
original_path = Path(
'tests/data/rotation_thermal_stress/preprocessed/cube/original')
transformed_paths = self.collect_transformed_paths(
'tests/data/rotation_thermal_stress/preprocessed/cube'
'/*_transformed_*')
ir = inferer.Inferer(
main_setting,
save=False,
converter_parameters_pkl=Path(
'tests/data/rotation_thermal_stress/preprocessed'
'/preprocessors.pkl'))
model_directory = str(main_setting.trainer.output_directory)
original_results = ir.infer(model=model_directory,
data_directories=[original_path])
transformed_results = ir.infer(model=model_directory,
data_directories=transformed_paths)
self.validate_results(original_results,
transformed_results,
rank2='nodal_strain_mat',
threshold_percent=2.,
decimal=1)
def test_integration_y1(self):
main_setting = setting.MainSetting.read_settings_yaml(
Path('tests/data/ode/integration_y1.yml'))
main_setting.trainer.gpu_id = 1
if main_setting.trainer.output_directory.exists():
shutil.rmtree(main_setting.trainer.output_directory)
tr = trainer.Trainer(main_setting)
loss = tr.train()
self.assertLess(loss, 5e-2)
ir = inferer.Inferer.read_settings(
main_setting.trainer.output_directory / 'settings.yml',
converter_parameters_pkl=main_setting.data.preprocessed_root /
'preprocessors.pkl')
results = ir.infer(
model=main_setting.trainer.output_directory,
data_directories=main_setting.data.preprocessed_root / 'test')
self.assertLess(results[0]['loss'], 1e-1)
if PLOT:
cmap = plt.get_cmap('tab10')
for i, result in enumerate(results):
plt.plot(result['dict_x']['t'][:, 0, 0],
result['dict_x']['y1'][:, 0, 0],
':',
color=cmap(i),
label=f"y1 answer of data {i}")
plt.plot(result['dict_x']['t'][:, 0, 0],
result['dict_y']['y1'][:, 0, 0],
color=cmap(i),
label=f"y1 inferred of data {i}")
plt.legend()
plt.show()
def test_iso_gcn_rotation_thermal_stress_frame_rank2(self):
main_setting = setting.MainSetting.read_settings_yaml(
Path('tests/data/rotation_thermal_stress/iso_gcn_frame_rank2.yml'))
tr = trainer.Trainer(main_setting)
if tr.setting.trainer.output_directory.exists():
shutil.rmtree(tr.setting.trainer.output_directory)
tr.train()
# Confirm inference result has isometric invariance and equivariance
original_path = Path(
'tests/data/rotation_thermal_stress/preprocessed/cube/original')
transformed_paths = self.collect_transformed_paths(
'tests/data/rotation_thermal_stress/preprocessed/cube'
'/*_transformed_*')
ir = inferer.Inferer(
main_setting,
save=False,
converter_parameters_pkl=Path(
'tests/data/rotation_thermal_stress/preprocessed'
'/preprocessors.pkl'))
model_directory = str(main_setting.trainer.output_directory)
inference_outpout_directory = \
main_setting.trainer.output_directory / 'inferred'
if inference_outpout_directory.exists():
shutil.rmtree(inference_outpout_directory)
original_results = ir.infer(model=model_directory,
data_directories=[original_path])
transformed_results = ir.infer(model=model_directory,
data_directories=transformed_paths)
self.validate_results(original_results,
transformed_results,
rank2='nodal_strain_mat',
threshold_percent=.002)
def test_iso_gcn_rotation_thermal_stress_rank2_rank0(self):
main_setting = setting.MainSetting.read_settings_yaml(
Path('tests/data/rotation_thermal_stress/iso_gcn_rank2_rank0.yml'))
tr = trainer.Trainer(main_setting)
if tr.setting.trainer.output_directory.exists():
shutil.rmtree(tr.setting.trainer.output_directory)
loss = tr.train()
np.testing.assert_array_less(loss, .5)
# Confirm inference result has isometric invariance
original_path = Path(
'tests/data/rotation_thermal_stress/preprocessed/cube/original')
transformed_paths = self.collect_transformed_paths(
'tests/data/rotation_thermal_stress/preprocessed/cube'
'/*_transformed_*')
ir = inferer.Inferer(
main_setting,
save=False,
converter_parameters_pkl=Path(
'tests/data/rotation_thermal_stress/preprocessed/'
'preprocessors.pkl'))
model_directory = str(main_setting.trainer.output_directory)
inference_outpout_directory = \
main_setting.trainer.output_directory / 'inferred'
if inference_outpout_directory.exists():
shutil.rmtree(inference_outpout_directory)
original_results = ir.infer(model=model_directory,
data_directories=[original_path])
transformed_results = ir.infer(model=model_directory,
data_directories=transformed_paths)
self.validate_results(original_results,
transformed_results,
rank0='initial_temperature',
decimal=5)
def test_iso_gcn_rank0_rank0(self):
main_setting = setting.MainSetting.read_settings_yaml(
Path('tests/data/rotation/iso_gcn_rank0_rank0.yml'))
tr = trainer.Trainer(main_setting)
if tr.setting.trainer.output_directory.exists():
shutil.rmtree(tr.setting.trainer.output_directory)
loss = tr.train()
np.testing.assert_array_less(loss, 1.)
self.assertEqual(len(tr.model.dict_block['ISO_GCN1'].subchains), 1)
self.assertEqual(len(tr.model.dict_block['ISO_GCN2'].subchains), 1)
# Confirm results does not change under rigid body transformation
original_path = Path(
'tests/data/rotation/preprocessed/cube/clscale1.0/original')
transformed_paths = self.collect_transformed_paths(
'tests/data/rotation/preprocessed/cube/clscale1.0/rotated')
ir = inferer.Inferer(
main_setting,
save=False,
converter_parameters_pkl=Path(
'tests/data/rotation/preprocessed/preprocessors.pkl'))
model_directory = str(main_setting.trainer.output_directory)
original_results = ir.infer(model=model_directory,
data_directories=[original_path])
transformed_results = ir.infer(model=model_directory,
data_directories=transformed_paths)
self.validate_results(original_results,
transformed_results,
rank0='t_100')
def test_iso_gcn_rotation_thermal_stress_dict_input_dict_output(self):
main_setting = setting.MainSetting.read_settings_yaml(
Path('tests/data/rotation_thermal_stress'
'/iso_gcn_dict_input_dict_output.yml'))
tr = trainer.Trainer(main_setting)
if tr.setting.trainer.output_directory.exists():
shutil.rmtree(tr.setting.trainer.output_directory)
tr.train()
# Confirm inference result has isometric equivariance
original_path = Path(
'tests/data/rotation_thermal_stress/preprocessed/cube/original')
transformed_paths = self.collect_transformed_paths(
'tests/data/rotation_thermal_stress/preprocessed/cube'
'/*_transformed_*')
ir = inferer.Inferer(
main_setting,
save=False,
converter_parameters_pkl=Path(
'tests/data/rotation_thermal_stress/preprocessed'
'/preprocessors.pkl'))
model_directory = str(main_setting.trainer.output_directory)
original_results = ir.infer(model=model_directory,
data_directories=[original_path])
transformed_results = ir.infer(model=model_directory,
data_directories=transformed_paths)
self.validate_results(original_results,
transformed_results,
rank0='cnt_temperature',
rank2='global_lte_mat')
def test_iso_gcn_inverse_temperature(self):
main_setting = setting.MainSetting.read_settings_yaml(
Path('tests/data/heat_time_series/iso_gcn_inverse.yml'))
tr = trainer.Trainer(main_setting)
if tr.setting.trainer.output_directory.exists():
shutil.rmtree(tr.setting.trainer.output_directory)
loss = tr.train()
np.testing.assert_array_less(loss, 5.)
def test_train_data_parallel_general_block(self):
main_setting = setting.MainSetting.read_settings_yaml(
Path('tests/data/deform/general_block.yml'))
main_setting.trainer.num_workers = 0 # Serial
main_setting.trainer.gpu_id = 0
tr_wo_parallel = trainer.Trainer(main_setting)
if main_setting.trainer.output_directory.exists():
shutil.rmtree(main_setting.trainer.output_directory)
loss_wo_parallel = tr_wo_parallel.train()
main_setting.trainer.data_parallel = True
tr_w_parallel = trainer.Trainer(main_setting)
if main_setting.trainer.output_directory.exists():
shutil.rmtree(main_setting.trainer.output_directory)
loss_w_parallel = tr_w_parallel.train()
np.testing.assert_almost_equal(loss_w_parallel / loss_wo_parallel, 1.)
def test_cluster_gcn(self):
main_setting = setting.MainSetting.read_settings_yaml(
Path('tests/data/deform/cluster_gcn.yml'))
tr = trainer.Trainer(main_setting)
if tr.setting.trainer.output_directory.exists():
shutil.rmtree(tr.setting.trainer.output_directory)
loss = tr.train()
np.testing.assert_array_less(loss, 1.)
def test_train_model_parallel_general_block(self):
main_setting = setting.MainSetting.read_settings_yaml(
Path('tests/data/deform/general_block_model_parallel.yml'))
tr = trainer.Trainer(main_setting)
if tr.setting.trainer.output_directory.exists():
shutil.rmtree(tr.setting.trainer.output_directory)
loss = tr.train()
np.testing.assert_array_less(loss, 2.)
def test_raise_valueerror_when_network_is_not_dag(self):
main_setting = setting.MainSetting.read_settings_yaml(
Path('tests/data/deform/not_dag.yml'))
if main_setting.trainer.output_directory.exists():
shutil.rmtree(main_setting.trainer.output_directory)
tr = trainer.Trainer(main_setting)
with self.assertRaisesRegex(ValueError, 'Cycle found in the network'):
tr.train()
def test_train_res_gcn(self):
main_setting = setting.MainSetting.read_settings_yaml(
Path('tests/data/deform/res_gcn_long.yml'))
main_setting.trainer.num_workers = 0 # Serial
tr = trainer.Trainer(main_setting)
if tr.setting.trainer.output_directory.exists():
shutil.rmtree(tr.setting.trainer.output_directory)
loss = tr.train()
np.testing.assert_array_less(loss, 1e-2)
def test_node_number_inference(self):
main_setting = setting.MainSetting.read_settings_yaml(
Path('tests/data/deform/node_number_inference.yml'))
if main_setting.trainer.output_directory.exists():
shutil.rmtree(main_setting.trainer.output_directory)
tr = trainer.Trainer(main_setting)
loss = tr.train()
self.assertLess(loss, 1e-1)
def test_raise_valueerror_when_block_has_missing_destinations(self):
main_setting = setting.MainSetting.read_settings_yaml(
Path('tests/data/deform/missing_destinations.yml'))
if main_setting.trainer.output_directory.exists():
shutil.rmtree(main_setting.trainer.output_directory)
tr = trainer.Trainer(main_setting)
with self.assertRaisesRegex(ValueError,
'NOT_EXISTING_BLOCK does not exist'):
tr.train()
def test_dict_output(self):
main_setting = setting.MainSetting.read_settings_yaml(
Path('tests/data/rotation_thermal_stress/'
'iso_gcn_dict_input_dict_output.yml'))
if main_setting.trainer.output_directory.exists():
shutil.rmtree(main_setting.trainer.output_directory)
tr = trainer.Trainer(main_setting)
loss = tr.train()
self.assertLess(loss, 1.)
def test_raise_valueerror_when_block_has_no_successors(self):
main_setting = setting.MainSetting.read_settings_yaml(
Path('tests/data/deform/no_successors.yml'))
if main_setting.trainer.output_directory.exists():
shutil.rmtree(main_setting.trainer.output_directory)
tr = trainer.Trainer(main_setting)
with self.assertRaisesRegex(ValueError,
'NO_SUCCESSORS has no successors'):
tr.train()
def test_train_data_parallel_linear(self):
main_setting = setting.MainSetting.read_settings_yaml(
Path('tests/data/linear/linear.yml'))
main_setting.data.train = [Path('tests/data/linear/preprocessed')]
main_setting.trainer.gpu_id = 0
main_setting.trainer.log_trigger_epoch = 100
main_setting.trainer.n_epoch = 500
tr_wo_parallel = trainer.Trainer(main_setting)
if main_setting.trainer.output_directory.exists():
shutil.rmtree(main_setting.trainer.output_directory)
loss_wo_parallel = tr_wo_parallel.train()
main_setting.trainer.data_parallel = True
main_setting.trainer.num_workers = 0 # Serial
tr_w_parallel = trainer.Trainer(main_setting)
if main_setting.trainer.output_directory.exists():
shutil.rmtree(main_setting.trainer.output_directory)
loss_w_parallel = tr_w_parallel.train()
np.testing.assert_almost_equal(loss_w_parallel, loss_wo_parallel)
def test_message_passing_gpu(self):
main_setting = setting.MainSetting.read_settings_yaml(
Path('tests/data/deform/message_passing.yml'))
main_setting.trainer.gpu_id = 1
main_setting.trainer.num_workers = 0 # Serial
tr = trainer.Trainer(main_setting)
if tr.setting.trainer.output_directory.exists():
shutil.rmtree(tr.setting.trainer.output_directory)
loss = tr.train()
np.testing.assert_array_less(loss, 1.)
def test_iso_gcn_gpu(self):
main_setting = setting.MainSetting.read_settings_yaml(
Path('tests/data/rotation_thermal_stress/'
'iso_gcn_dict_input_dict_output.yml'))
main_setting.trainer.gpu_id = 1
main_setting.trainer.num_workers = 0 # Serial
tr = trainer.Trainer(main_setting)
if tr.setting.trainer.output_directory.exists():
shutil.rmtree(tr.setting.trainer.output_directory)
loss = tr.train()
np.testing.assert_array_less(loss, 1.)
def test_iso_gcn_diag(self):
main_setting = setting.MainSetting.read_settings_yaml(
Path(
'tests/data/rotation_thermal_stress/iso_gcn_rank0_rank2_diag.yml'
))
main_setting.trainer.gpu_id = 1
tr = trainer.Trainer(main_setting)
if tr.setting.trainer.output_directory.exists():
shutil.rmtree(tr.setting.trainer.output_directory)
loss = tr.train()
self.assertLess(loss, 10.)
def test_train_data_parallel_dict_input_list_output(self):
main_setting = setting.MainSetting.read_settings_yaml(
Path('tests/data/rotation_thermal_stress/iso_gcn_dict_input.yml'))
main_setting.trainer.num_workers = 0 # Serial
main_setting.trainer.n_epoch = 10
main_setting.trainer.batch_size = 6
main_setting.trainer.gpu_id = 0
tr_wo_parallel = trainer.Trainer(main_setting)
if main_setting.trainer.output_directory.exists():
shutil.rmtree(main_setting.trainer.output_directory)
loss_wo_parallel = tr_wo_parallel.train()
main_setting.trainer.data_parallel = True
tr_w_parallel = trainer.Trainer(main_setting)
if main_setting.trainer.output_directory.exists():
shutil.rmtree(main_setting.trainer.output_directory)
loss_w_parallel = tr_w_parallel.train()
np.testing.assert_almost_equal(loss_w_parallel / loss_wo_parallel,
1.,
decimal=5)
def test_train_concatenated_data(self):
preprocessed_base_directory = Path(
'tests/data/linear/preprocessed/train')
concatenated_directory = Path('tests/data/linear/concatenated')
shutil.rmtree(concatenated_directory, ignore_errors=True)
pre.concatenate_preprocessed_data(
preprocessed_base_directory, concatenated_directory,
variable_names=['x1', 'x2', 'y'], ratios=(.9, 0.1, 0.))
main_setting = setting.MainSetting.read_settings_yaml(
Path('tests/data/linear/linear_concatenated.yml'))
tr = trainer.Trainer(main_setting)
if tr.setting.trainer.output_directory.exists():
shutil.rmtree(tr.setting.trainer.output_directory)
loss = tr.train()
np.testing.assert_array_less(loss, 1e-5)
def test_user_defined_block(self):
class CutOffBlock(networks.SimlModule):
def __init__(self, block_setting):
super().__init__(block_setting)
self.upper = block_setting.optional.get('upper', .1)
self.lower = block_setting.optional.get('lower', 0.)
def _forward_core(self, x, supports=None, original_shapes=None):
h = x
for linear, dropout_ratio, activation in zip(
self.linears, self.dropout_ratios, self.activations):
h = linear(h)
h = activation(h)
h[h > self.upper] = self.upper
h[h < self.lower] = self.lower
return h
networks.add_block(name='cutoff', block=CutOffBlock, trainable=True)
main_setting = setting.MainSetting.read_settings_yaml(
Path('tests/data/deform/cutoff.yml'))
tr = trainer.Trainer(main_setting)
if tr.setting.trainer.output_directory.exists():
shutil.rmtree(tr.setting.trainer.output_directory)
loss = tr.train()
np.testing.assert_array_less(loss, 1.)
x = torch.from_numpy(np.random.rand(2000, 100).astype(np.float32))
out_cutoff1 = tr.model.dict_block['CUTOFF1'](x).detach().numpy()
np.testing.assert_almost_equal(
np.max(out_cutoff1),
main_setting.model.blocks[1].optional['upper'])
np.testing.assert_almost_equal(
np.min(out_cutoff1),
main_setting.model.blocks[1].optional['lower'])
out_cutoff2 = tr.model.dict_block['CUTOFF2'](x).detach().numpy()
np.testing.assert_almost_equal(
np.max(out_cutoff2),
main_setting.model.blocks[2].optional['upper'])
np.testing.assert_almost_equal(
np.min(out_cutoff2),
main_setting.model.blocks[2].optional['lower'])
def test_iso_gcn_inverse_invariant(self):
main_setting = setting.MainSetting.read_settings_yaml(
Path('tests/data/rotation_thermal_stress'
'/iso_gcn_rank0_rank0_inverse.yml'))
tr = trainer.Trainer(main_setting)
if tr.setting.trainer.output_directory.exists():
shutil.rmtree(tr.setting.trainer.output_directory)
loss = tr.train()
np.testing.assert_array_less(loss, 5.)
# Confirm inference result has isometric invariance
original_path = Path(
'tests/data/rotation_thermal_stress/preprocessed/cube/original')
transformed_paths = self.collect_transformed_paths(
'tests/data/rotation_thermal_stress/preprocessed/cube'
'/*_transformed_*')
ir = inferer.Inferer(
main_setting,
save=False,
converter_parameters_pkl=Path(
'tests/data/rotation_thermal_stress/preprocessed'
'/preprocessors.pkl'))
model_directory = str(main_setting.trainer.output_directory)
inference_outpout_directory = \
main_setting.trainer.output_directory / 'inferred'
if inference_outpout_directory.exists():
shutil.rmtree(inference_outpout_directory)
original_results = ir.infer(model=model_directory,
data_directories=[original_path])
transformed_results = ir.infer(model=model_directory,
data_directories=transformed_paths)
# NOTE: LTE is not invariant, but just to validate IsoGCN invariance,
# we use it.
self.validate_results(original_results,
transformed_results,
rank0='global_lte_array',
validate_x=False)
def test_grad_grad_vs_laplacian(self):
preprocess_setting = setting.MainSetting.read_settings_yaml(
Path('tests/data/precalculated/convert.yml'))
shutil.rmtree(preprocess_setting.data.interim, ignore_errors=True)
shutil.rmtree(preprocess_setting.data.preprocessed, ignore_errors=True)
raw_converter = pre.RawConverter(
preprocess_setting,
conversion_function=convert.HeatConversionFuncionCreator(
create_elemental=False),
filter_function=convert.filter_function_heat,
force_renew=True, recursive=True, to_first_order=True,
write_ucd=False, read_npy=False, read_res=True)
raw_converter.convert()
preprocessor = pre.Preprocessor(
preprocess_setting, force_renew=True)
preprocessor.preprocess_interim_data()
# Confirm results does not change under rigid body transformation
grad_grad_setting = setting.MainSetting.read_settings_yaml(
Path('tests/data/precalculated/grad_grad.yml'))
grad_grad_tr = trainer.Trainer(grad_grad_setting)
if grad_grad_tr.setting.trainer.output_directory.exists():
shutil.rmtree(grad_grad_tr.setting.trainer.output_directory)
grad_grad_tr.train()
grad_grad_ir = inferer.Inferer(grad_grad_setting)
inference_outpout_directory = \
grad_grad_setting.trainer.output_directory / 'inferred'
if inference_outpout_directory.exists():
shutil.rmtree(inference_outpout_directory)
results = grad_grad_ir.infer(
model=Path('tests/data/precalculated/models/grad_grad'),
preprocessed_data_directory=[
Path(
'tests/data/precalculated/preprocessed/cube/clscale1.0/'
'original'),
Path(
'tests/data/precalculated/preprocessed/cube/clscale1.0/'
'rotated')],
converter_parameters_pkl=Path(
'tests/data/precalculated/preprocessed/preprocessors.pkl'),
output_directory=inference_outpout_directory,
overwrite=True)
np.testing.assert_almost_equal(
results[0]['dict_y']['nodal_t_100'],
results[1]['dict_y']['nodal_t_100'])
laplacian_setting = setting.MainSetting.read_settings_yaml(
Path('tests/data/precalculated/laplacian.yml'))
laplacian_tr = trainer.Trainer(laplacian_setting)
if laplacian_tr.setting.trainer.output_directory.exists():
shutil.rmtree(laplacian_tr.setting.trainer.output_directory)
laplacian_tr.train()
laplacian_ir = inferer.Inferer(laplacian_setting)
inference_outpout_directory = \
laplacian_setting.trainer.output_directory / 'inferred'
if inference_outpout_directory.exists():
shutil.rmtree(inference_outpout_directory)
results = laplacian_ir.infer(
model=Path('tests/data/precalculated/models/laplacian'),
preprocessed_data_directory=[
Path(
'tests/data/precalculated/preprocessed/cube/clscale1.0/'
'original'),
Path(
'tests/data/precalculated/preprocessed/cube/clscale1.0/'
'rotated')],
converter_parameters_pkl=Path(
'tests/data/precalculated/preprocessed/preprocessors.pkl'),
output_directory=inference_outpout_directory,
overwrite=True)
np.testing.assert_almost_equal(
results[0]['dict_y']['nodal_t_100'],
results[1]['dict_y']['nodal_t_100'])
