def test_static_inputs(self, tmp_path):
ds = _make_dataset().isel(lat=slice(0, 2), lon=slice(0, 1))
ds_static = ds.mean(dim="time")
cfg = Config(Path("tests/testconfigs/test_config.yml"))
create_and_assign_temp_run_path_to_config(cfg, tmp_path)
assert False
def test_kenya_data(self, tmp_path):
if TEST_REAL_DATA:
ds = pickle.load(Path("data/kenya.pkl").open("rb")).isel(
lat=slice(0, 5), lon=slice(0, 5)
)
cfg = Config(Path("tests/testconfigs/config.yml"))
create_and_assign_temp_run_path_to_config(cfg, tmp_path)
dl = PixelDataLoader(
ds, cfg=cfg, num_workers=1, mode="train", batch_size=cfg.batch_size
)
data = dl.__iter__().__next__()
x, _ = data["x_d"], data["y"]
batch_size = 256
seq_length = cfg.seq_length
input_variables = ["precip", "t2m", "SMsurf"]
autoregressive = True
n_inputs = (
len(input_variables) + 1 if autoregressive else len(input_variables)
)
assert cfg.batch_size == batch_size
assert cfg.autoregressive == autoregressive
assert x.shape == (
batch_size,
seq_length,
n_inputs,
), f"X Data Mismatch! Expected: {(batch_size, seq_length, n_inputs)} Got: {x.shape}"
else:
pass
def test_runoff_data(self, tmp_path):
if TEST_REAL_DATA:
ds = xr.open_dataset("data/ALL_dynamic_ds.nc").isel(station_id=slice(0, 5))
cfg = Config(Path("tests/testconfigs/config_runoff.yml"))
create_and_assign_temp_run_path_to_config(cfg, tmp_path)
# train period
input_variables = [] if cfg.input_variables is None else cfg.input_variables
train_ds = ds[input_variables + [cfg.target_variable]].sel(
time=slice(cfg.train_start_date, cfg.train_end_date)
)
train_dl = PixelDataLoader(
train_ds,
cfg=cfg,
mode="train",
num_workers=4,
batch_size=cfg.batch_size,
)
# check data is loaded properly
data = next(iter(train_dl))
x, y = data["x_d"], data["y"]
n_in_vars = (
len(cfg.input_variables) + 1
if cfg.autoregressive
else len(cfg.input_variables)
)
assert x.shape == (cfg.batch_size, cfg.seq_length, n_in_vars)
assert y.shape == (cfg.batch_size, 1, 1)
else:
pass
def test_dataloader(self, tmp_path):
ds = _make_dataset()
cfg = Config(Path("tests/testconfigs/test_config.yml"))
create_and_assign_temp_run_path_to_config(cfg, tmp_path)
static = create_static(cfg=cfg, ds=ds)
dl = PixelDataLoader(
ds,
cfg=cfg,
num_workers=1,
mode="train",
batch_size=cfg.batch_size,
static_data=static,
)
assert dl.batch_size == cfg.batch_size
seq_length = cfg.seq_length
autoregressive = cfg.autoregressive
data = next(iter(dl))
x, y = data["x_d"], data["y"]
n_inputs = len(["features"]) + 1 if autoregressive else len(["features"])
assert x.shape == (
cfg.batch_size,
seq_length,
n_inputs + 2 if cfg.encode_doys else n_inputs,
), f"Size Mismatch! Expected: {(cfg.batch_size, seq_length, n_inputs)} Got: {x.shape}"
def test_runoff_example(self, tmp_path):
cfg = Config(Path("tests/testconfigs/config_runoff.yml"))
create_and_assign_temp_run_path_to_config(cfg, tmp_path)
cfg._cfg["data_path"] = Path("data/ALL_dynamic_ds.nc")
cfg._cfg["static_data_path"] = Path("data/camels_static.nc")
cfg._cfg["static_inputs"] = ["p_mean", "pet_mean", "area", "gauge_elev"]
cfg._cfg["n_epochs"] = 3
ds, static = load_data(cfg)
# select subset of 3 basins
basins = [1001, 2001, 2002]
ds = ds.sel(station_id=basins)
static = static.sel(station_id=basins)
trainer = Trainer(cfg, ds, static_data=static)
self.check_loaded_data(
cfg,
trainer,
data=ds.sel(time=slice(cfg.train_start_date, cfg.train_end_date)),
)
losses = trainer.train_and_validate()
tester = Tester(cfg, ds, static_data=static)
preds = tester.run_test()
return losses, preds
def test_dataset(self, tmp_path):
target_variable = "target"
input_variables = ["feature"]
for path in [
Path("tests/testconfigs/test_config_simulate.yml"),
Path("tests/testconfigs/test_config.yml"),
]:
cfg = Config(path)
cfg._cfg["forecast_variables"] = cfg.input_variables
create_and_assign_temp_run_path_to_config(cfg, tmp_path)
raw_ds = _make_dataset().isel(lat=slice(0, 2), lon=slice(0, 1))
static = create_static(cfg=cfg, ds=raw_ds)
ds = XarrayDataset(
raw_ds, cfg=cfg, mode="train", DEBUG=True, static_data=static
)
assert ds.target == target_variable
assert (
ds.inputs == input_variables + ["autoregressive"]
if cfg.autoregressive
else input_variables
)
x_features = (
len(input_variables) + 1 if cfg.autoregressive else len(input_variables)
)
seq_length = cfg.seq_length
for i in range(10):
data = ds.__getitem__(i)
x, y = data["x_d"], data["y"]
assert y.shape == (1, 1)
assert x.shape == (
seq_length,
x_features + 2 if cfg.encode_doys else x_features,
), f"Shape Mismatch! Expect: {(seq_length, x_features)} Got: {x.shape}"
meta = data["meta"]
times = (
meta["target_time"]
.detach()
.numpy()
.astype("datetime64[ns]")
.flatten()
)
pixel, _ = ds.lookup_table[int(meta["index"])]
latlon = tuple([float(l) for l in str(pixel).split("_")])
y_unnorm = (
ds.normalizer.individual_inverse(y, pixel, variable="target")
.detach()
.numpy()
)
# extract from the original xr.Dataset
y_exp = raw_ds.sel(
lat=latlon[0], lon=latlon[1], time=times, method="nearest"
)[cfg.target_variable].values
assert np.isclose(y_unnorm.reshape(y_exp.shape), y_exp, atol=1e-5)
def test_dataset_beijing(self, tmp_path):
if is_connected():
path = Path("tests/testconfigs/pollution.yml")
cfg = Config(path)
create_and_assign_temp_run_path_to_config(cfg, tmp_path)
raw_ds = get_pollution_data_beijing().to_xarray().isel(time=slice(0, 1000))
ds = XarrayDataset(raw_ds, cfg=cfg, mode="train", DEBUG=True)
assert ds.y != {}
def test_forecast_inputs(self, tmp_path):
ds = _make_dataset().isel(lat=slice(0, 2), lon=slice(0, 1))
ds_forecast = (
ds.shift(time=1).rename({"feature": "feature_fcast1"}).drop("target")
)
ds = xr.merge([ds, ds_forecast])
cfg = Config(Path("tests/testconfigs/test_config.yml"))
create_and_assign_temp_run_path_to_config(cfg, tmp_path)
assert False
def test_longer_horizon_fcast(self, tmp_path):
cfg = Config(Path("tests/testconfigs/test_1d_config_horizon.yml"))
create_and_assign_temp_run_path_to_config(cfg, tmp_path)
ds = load_test_jena_data_as_dataset()
dl = PixelDataLoader(
ds, cfg=cfg, num_workers=1, mode="train", batch_size=cfg.batch_size
)
data = dl.__iter__().__next__()
_, y = data["x_d"], data["y"]
assert y.shape == (cfg.batch_size, 1, 1)
def test_1D_data(self, tmp_path):
# convert pandas to xarray object
ds = load_test_jena_data_as_dataset()
cfg = Config(Path("tests/testconfigs/test_1d_config_horizon.yml"))
create_and_assign_temp_run_path_to_config(cfg, tmp_path)
dl = PixelDataLoader(
ds, cfg=cfg, num_workers=1, mode="train", batch_size=cfg.batch_size
)
data = dl.__iter__().__next__()
x, y = data["x_d"], data["y"]
assert x.shape == (cfg.batch_size, cfg.seq_length, len(cfg.input_variables))
assert y.shape == (cfg.batch_size, 1, 1)
def test_lstm_forward_pass(self, tmp_path):
ds = pickle.load(Path("data/kenya.pkl").open("rb"))
cfg = Config(Path("tests/testconfigs/config.yml"))
create_and_assign_temp_run_path_to_config(cfg, tmp_path)
dl = PixelDataLoader(ds, cfg=cfg, mode="train")
model = LSTM(
input_size=dl.input_size + dl.static_input_size + dl.forecast_input_size,
hidden_size=cfg.hidden_size,
output_size=dl.output_size,
forecast_horizon=dl.horizon,
)
data = dl.__iter__().__next__()
y_hat = model(data)
assert all(np.isin(["h_n", "c_n", "y_hat"], [k for k in y_hat.keys()]))
def test_linear_regression_forward_pass(self, tmp_path):
ds = _make_dataset()
cfg = Config(Path("tests/testconfigs/test_config.yml"))
create_and_assign_temp_run_path_to_config(cfg, tmp_path)
dl = PixelDataLoader(ds, cfg=cfg, mode="train", DEBUG=True)
model = LinearRegression(
input_size=(dl.input_size + dl.static_input_size + dl.forecast_input_size)
* cfg.seq_length,
output_size=dl.output_size,
forecast_horizon=dl.horizon,
)
data = dl.__iter__().__next__()
y_hat = model(data)
assert isinstance(y_hat, dict)
assert y_hat["y_hat"].shape == (1, 1)
def test_kenya_vci_example(self, tmp_path):
cfg = Config(Path("tests/testconfigs/config.yml"))
create_and_assign_temp_run_path_to_config(cfg, tmp_path)
cfg._cfg["data_path"] = Path("data/kenya.nc")
cfg._cfg["n_epochs"] = 3
ds, static = load_data(cfg)
trainer = Trainer(cfg, ds, static_data=static)
self.check_loaded_data(
cfg,
trainer,
data=ds.sel(time=slice(cfg.train_start_date, cfg.train_end_date)),
)
losses = trainer.train_and_validate()
tester = Tester(cfg, ds, static_data=static)
preds = tester.run_test()
return losses, preds
def test_correct_data_returned(self, tmp_path):
# create dummy config path
cfg = Config(Path("tests/testconfigs/test_config.yml"))
cfg._cfg["encode_doys"] = True
cfg._cfg["static_inputs"] = "embedding"
cfg._cfg["forecast_variables"] = cfg.input_variables
# create temporary run directory (usually done by the )
create_and_assign_temp_run_path_to_config(cfg, tmp_path)
# create dummy dataset
ds = _make_dataset().isel(lat=slice(0, 2), lon=slice(0, 1))
# initialise the dataloader
dl = PixelDataLoader(ds, cfg=cfg, mode="train", DEBUG=True)
# one sample from the dataloader
data = dl.__iter__().__next__()
x, y = data["x_d"], data["y"]
# recreate the stacked dataset
stacked_ds = dl.dataset.ds
if cfg.encode_doys:
stacked_ds, _ = add_doy_encoding_as_feature_to_dataset(
stacked_ds, inputs=cfg.input_variables, target=cfg.target_variable
)
# get the current_time_index and pixel from the __getitem__() call
getitem_call = int(data["meta"]["index"])
pixel, current_time_index = dl.dataset.lookup_table[getitem_call]
# check that the returned data is valid
# TODO: wrap into function for getting the valid times!
est_target_time = pd.to_datetime(
np.array(data["meta"]["target_time"]).astype("datetime64[ns]")
)[0]
# rounding error because of storing as float
input_data_times = pd.to_datetime(stacked_ds.time.values)
true_target_index = input_data_times.get_loc(est_target_time, method="nearest")
true_target_time = input_data_times[true_target_index]
assert current_time_index + cfg.horizon == true_target_index
# :: RECREATE TARGET DATA ::
all_expected_y = stacked_ds.sel(sample=pixel)["target"].values
expected_y = stacked_ds.sel(sample=pixel, time=true_target_time)[
cfg.target_variable
].values
expected_y_index = (
stacked_ds.sel(sample=pixel)
.isel(time=true_target_index)[cfg.target_variable]
.values
)
assert expected_y == expected_y_index
assert np.isclose(y.flatten()[-1], expected_y)
## :: RECREATE INPUT DATA ::
# max_input_ix should be the CURRENT TIME (+ 1 because of exlusive upper indexing)
max_input_ix = int(true_target_index - cfg.horizon)
assert max_input_ix == current_time_index
max_input_time = input_data_times[max_input_ix]
# min_input_ix = the first input time
min_input_ix = int(max_input_ix - cfg.seq_length) + 1
min_input_time = input_data_times[min_input_ix]
input_vars = (
cfg.input_variables + ["autoregressive"]
if cfg.autoregressive
else cfg.input_variables
)
input_vars = (
input_vars + ["sin_doy", "cos_doy"] if cfg.encode_doys else input_vars
)
# has x been drawn from the actual underlying data?
all_expected_x = stacked_ds.sel(sample=pixel)["feature"].values
_expected_x = all_expected_x[min_input_ix:max_input_ix]
# assert x == _expected_x
# assert all(
# np.isin(
# np.round(x.numpy().flatten(), 3).astype("float64"),
# np.round(all_expected_x.flatten(), 3).astype("float64"),
# )
# )
# get the exact expected input vector
# NOTE: slice is NOT EXCLUSIVE UPPER therefore need to exclude the final
expected_x_feature = (
stacked_ds.sel(sample=pixel, time=slice(min_input_time, max_input_time))[
input_vars
]
.to_array()
.values.T
)
x_feature = np.array(x)
x_feature = x_feature.reshape(expected_x_feature.shape)
assert np.allclose(x_feature, expected_x_feature)
def test_linear_example(self, tmp_path):
"""Test the linear dataset.
Args:
tmp_path ([type]): [description]
"""
cfg = Config(Path("tests/testconfigs/test_config.yml"))
create_and_assign_temp_run_path_to_config(cfg, tmp_path)
# Create linear dataset
alpha = 0
beta = 2
epsilon_sigma = 0
ds = create_linear_ds(
horizon=cfg.horizon, alpha=alpha, beta=beta, epsilon_sigma=epsilon_sigma
).isel(lat=slice(0, 2), lon=slice(0, 2))
static = create_static(cfg=cfg, ds=ds)
dl = PixelDataLoader(
ds,
cfg=cfg,
num_workers=1,
mode="train",
batch_size=cfg.batch_size,
DEBUG=True,
static_data=static,
)
# load all of the data into memory
data = load_all_data_from_dl_into_memory(dl)
x = data["x_d"]
# (n_samples, n_features, seq_length)
assert x.shape == (
len(cfg.input_variables) + 2
if cfg.encode_doys
else len(cfg.input_variables),
cfg.seq_length,
)
assert x.shape[-1] == cfg.seq_length
y = data["y"]
times = pd.to_datetime(data["time"].astype("datetime64[ns]").flatten())
# matching batch dims (n_samples) for all samples
assert x.shape[0] == y.shape[0]
# test ONE SINGLE (x, y) sample
SAMPLE = 1
# get metadata for sample
idx = int(data["index"][SAMPLE])
pixel, valid_current_time_index = dl.dataset.lookup_table[idx]
latlon = tuple([float(l) for l in str(pixel).split("_")])
target_time = times[SAMPLE]
# current_time = times[valid_current_time_index][0]
# get the correct times (weird indexing becuase of imperfect translation of float -> datetime64[ns])
max_time = target_time - DateOffset(months=cfg.horizon) + DateOffset(days=2)
min_time = max_time - DateOffset(months=cfg.seq_length)
input_times = pd.date_range(min_time, max_time, freq="M")[-cfg.seq_length :]
# recreate the data that should be loaded from the raw xr.Dataset
stacked, _ = _stack_xarray(ds, spatial_coords=cfg.pixel_dims)
normalizer = dl.normalizer
norm_stacked = normalizer.transform(stacked)
all_y = norm_stacked["target"].sel(sample=pixel)
_y = all_y.sel(time=target_time, method="nearest")
all_x = norm_stacked["feature"].sel(sample=pixel)
_x_d = all_x.sel(time=input_times, method="nearest")
# check that the dataloader saves & returns the correct values
assert np.allclose(
dl.dataset.y[pixel], (all_y.values)
), "The DataLoader saves incorrect y values to memory"
assert np.isclose(
_y.values, y[SAMPLE]
), "The DataLoader returns an incorrect value from the Dataset"
# input (X) data
dataset_loaded = dl.dataset.x_d[pixel]
# assert dataset_loaded.shape == (, cfg.seq_length)
expected = all_x.values.reshape(dataset_loaded.shape)
mask = np.isnan(expected)
expected = expected[~mask]
dataset_loaded = dataset_loaded[~mask]
assert np.allclose(
dataset_loaded, expected
), f"The dataloader is saving the wrong data to the lookup table. {dataset_loaded[:10]} {expected[:10]}"
# get input X data from INDEX (not times)
max_input_ix = int(valid_current_time_index)
min_input_ix = int(max_input_ix - cfg.seq_length) + 1
_x_d_index_values = all_x.values[min_input_ix : max_input_ix + 1]
assert np.allclose(_x_d_index_values, _x_d.values)
# TODO: Why does this not work?
assert np.allclose(
_x_d_index_values.values, x[SAMPLE]
), "The dynamic data is not the data we expect"
# check that the raw data is the linear combination we expect
# "target" should be linear combination of previous timestep "feature"
# (y = x @ [0, 2])
zeros = np.zeros((cfg.seq_length - 1, 1))
betas = np.append(zeros, beta).reshape(-1, 1)
unnorm_x = dl.dataset.normalizer.individual_inverse(
x[SAMPLE], pixel_id=pixel, variable=cfg.input_variables[0]
)
unnorm_y = dl.dataset.normalizer.individual_inverse(
y[SAMPLE], pixel_id=pixel, variable=cfg.target_variable
)
# time=target_time,
ds.sel(lat=latlon[0], lon=latlon[1], method="nearest")[cfg.target_variable]
assert np.isclose(unnorm_x @ betas, unnorm_y)
def test_single_train_step(self, tmp_path):
torch.manual_seed(1)
np.random.seed(1)
hidden_size = 64
ds = pickle.load(Path("data/kenya.pkl").open("rb")).isel(
lat=slice(0, 2), lon=slice(0, 4)
)
paths = [
Path("tests/testconfigs/config.yml"),
Path("tests/testconfigs/config_multi_horizon.yml"),
]
for path in paths:
cfg = Config(path)
cfg._cfg["static_inputs"] = "embedding"
create_and_assign_temp_run_path_to_config(cfg, tmp_path)
dl = PixelDataLoader(
ds, mode="train", cfg=cfg, num_workers=1, batch_size=cfg.batch_size,
)
data1 = dl.dataset.__getitem__(0)
data1["x_s"]
data = dl.__iter__().__next__()
x, y = data["x_d"], data["y"]
# are we working with batches or individual predictions?
x = x.unsqueeze(0) if x.ndim == 2 else x
model = (
LSTM(
input_size=dl.input_size
+ dl.static_input_size
+ dl.forecast_input_size,
hidden_size=hidden_size,
output_size=dl.output_size,
forecast_horizon=dl.horizon,
)
.float()
.to(cfg.device)
)
optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
loss_obj = F.mse_loss
before = model.forward(data)
for data in tqdm(dl):
input, target = data["x_d"], data["y"]
optimizer.zero_grad()
yhat = model.forward(data)
# shape = [batch_size, seq_length, forecast_horizon]
assert yhat["y_hat"].shape == (cfg.batch_size, 1, 1)
# get the final predictions to calculate loss
loss = loss_obj(yhat["y_hat"], target)
loss.backward()
optimizer.step()
break
after = model.forward(data)
loss_bf = loss_obj(before["y_hat"], y)
loss_af = loss_obj(after["y_hat"], y)
# NOTE: the LSTM only returns the final hidden and cell state layer NOT each timestep
# TODO: why is the LSTM returning a hidden array of shape (seq_length, 1, hs)
assert before["h_n"].shape == (1, cfg.batch_size, hidden_size)
assert before["y_hat"].shape == (cfg.batch_size, 1, 1)
if cfg.horizon == 1:
assert (
loss_af < loss_bf
), "The model did not learn anything after one epoch of training"
