def test_training_multiple_times(self, model_cls):
# a basic integration test
equation = readers.get_equation(self.metadata)
grid = readers.get_output_grid(self.metadata)
model = model_cls(equation, grid, num_time_steps=4)
def create_inputs(state):
# (batch, x, y)
inputs = nest.map_structure(lambda x: x[:-model.num_time_steps], state)
# (batch, time, x, y)
labels = tensor_ops.stack_all_contiguous_slices(
state['concentration'][1:], model.num_time_steps, new_axis=1)
return inputs, labels
training_data = (
model.load_data(self.metadata)
.map(create_inputs)
.apply(tf.data.experimental.unbatch())
.shuffle(10)
.repeat()
.batch(4, drop_remainder=True)
.prefetch(1)
)
model.compile(optimizer=tf.keras.optimizers.Adam(1e-4),
loss='mean_squared_error')
model.fit(training_data, epochs=1, steps_per_epoch=5)
model.evaluate(training_data, steps=2)
def test_training(self, model_cls):
# a basic integration test
equation = readers.get_equation(self.metadata)
grid = readers.get_output_grid(self.metadata)
model = model_cls(equation, grid)
def create_inputs(state):
inputs = nest.map_structure(lambda x: x[:-1], state)
labels = state['concentration'][1:]
return inputs, labels
training_data = (model.load_data(
self.metadata).repeat().shuffle(10).map(create_inputs))
model.compile(optimizer=tf.keras.optimizers.Adam(1e-4),
loss='mean_squared_error')
model.fit(training_data, epochs=1, steps_per_epoch=5)
model.evaluate(training_data, steps=2)
def load_data(
self,
metadata: Mapping[str, Any],
prefix: states.Prefix = states.Prefix.EXACT,
) -> tf.data.Dataset:
"""Load data into a tf.data.Dataset for inferrence or training."""
def replace_state_keys_with_names(state):
return {k: state[equation.key_definitions[k].with_prefix(prefix)]
for k in equation.base_keys}
equation = readers.get_equation(metadata)
grid = readers.get_output_grid(metadata)
keys = [equation.key_definitions[k].with_prefix(prefix)
for k in equation.base_keys]
dataset = readers.initialize_dataset(metadata, [keys], [grid])
dataset = dataset.map(replace_state_keys_with_names)
return dataset
def test_shapes_and_exceptions(self):
"""Dataset writer and reader test, checks shapes and exceptions."""
output_path = FLAGS.test_tmpdir
output_name = 'temp'
equation_name = 'advection_diffusion'
discretization = 'finite_volume'
dataset_type = 'all_derivatives'
high_resolution = 125
low_resolution = 25
shards = 2
example_num_time_steps = 3
batch_size = 4
diffusion_coefficient = 0.3
expected_equation = advection_equations.FiniteVolumeAdvectionDiffusion(
diffusion_coefficient=diffusion_coefficient)
# create a temporary dataset
with flagsaver.flagsaver(
dataset_path=output_path,
dataset_name=output_name,
equation_name=equation_name,
discretization=discretization,
simulation_grid_size=high_resolution,
output_grid_size=low_resolution,
equation_kwargs=str(
dict(diffusion_coefficient=diffusion_coefficient)),
dataset_type=dataset_type,
num_shards=shards,
total_time_steps=10,
example_num_time_steps=example_num_time_steps,
time_step_interval=5,
num_seeds=4,
):
create_training_data.main([], runner=beam.runners.DirectRunner())
metadata_path = os.path.join(output_path,
output_name + '.metadata.json')
self.assertTrue(gfile.exists(metadata_path))
dataset_metadata = readers.load_metadata(metadata_path)
low_res_grid = readers.get_output_grid(dataset_metadata)
high_res_grid = readers.get_simulation_grid(dataset_metadata)
equation = readers.get_equation(dataset_metadata)
self.assertEqual(low_res_grid.size_x, low_resolution)
self.assertEqual(low_res_grid.size_y, low_resolution)
self.assertEqual(high_res_grid.size_x, high_resolution)
self.assertEqual(high_res_grid.size_y, high_resolution)
self.assertAlmostEqual(high_res_grid.step, 2 * np.pi / high_resolution)
self.assertAlmostEqual(equation.diffusion_coefficient,
diffusion_coefficient)
self.assertIs(type(equation), type(expected_equation))
state_keys = expected_equation.key_definitions
valid_data_keys = ((state_keys['concentration'].exact(), ),
(state_keys['concentration_edge_x'].exact(),
state_keys['concentration_y_edge_y'].exact()))
invalid_data_keys = ((state_keys['concentration'],
state_keys['concentration_edge_x']),
(state_keys['concentration_edge_x'], ))
valid_data_grids = (low_res_grid, low_res_grid)
invalid_data_grids = (low_res_grid, high_res_grid)
with self.assertRaises(ValueError):
readers.initialize_dataset(dataset_metadata, invalid_data_keys,
valid_data_grids)
with self.assertRaises(ValueError):
readers.initialize_dataset(dataset_metadata, valid_data_keys,
invalid_data_grids)
with self.assertRaises(ValueError):
readers.initialize_dataset(dataset_metadata, invalid_data_keys,
invalid_data_grids)
dataset = readers.initialize_dataset(dataset_metadata, valid_data_keys,
valid_data_grids)
dataset = dataset.repeat()
dataset = dataset.batch(batch_size)
[(first_state, second_state)] = dataset.take(1)
self.assertEqual(set(first_state.keys()), set(valid_data_keys[0]))
self.assertEqual(set(second_state.keys()), set(valid_data_keys[1]))
first_state_shape = np.shape(first_state[valid_data_keys[0][0]])
second_state_shape = np.shape(second_state[valid_data_keys[1][0]])
expected_shape = (batch_size, example_num_time_steps, low_resolution,
low_resolution)
self.assertEqual(first_state_shape, expected_shape)
self.assertEqual(second_state_shape, expected_shape)