def test_TrainConfig_build_loss():
config = TrainConfig(".", ".", ".", loss=CustomLoss(loss_variables=["x"]))
# needs to be random or the normalized loss will have nan
data = {"x": tf.random.uniform(shape=(4, 10))}
loss = config.build_loss(data)
loss_value, _ = loss(data, data)
assert 0 == pytest.approx(loss_value.numpy())
def test_TrainConfig_defaults():
config = TrainConfig(
train_url="train_path",
test_url="test_path",
out_url="save_path",
transform=TransformConfig(),
model=MicrophysicsConfig(),
)
assert config # for linter
def test_TrainConfig_asdict():
config = TrainConfig(
train_url="train_path",
test_url="test_path",
out_url="save_path",
model=MicrophysicsConfig(),
)
d = asdict(config)
assert d["train_url"] == "train_path"
assert d["model"]["architecture"]["name"] == "linear"
def test_TrainConfig_build_model():
field = Field("out", "in")
config = TrainConfig(
".",
".",
".",
transformed_model=TransformedModelConfig(ArchitectureConfig("dense"),
[field], 900),
)
data = {
field.input_name: tf.ones((1, 10)),
field.output_name: tf.ones((1, 10))
}
model = config.build_model(data)
assert field.output_name in model(data)