def test_empty_X(tmpdir, fake_input_dir, output_dir, code_dir):
# Configuration
test_input_dir = os.path.join(str(tmpdir), 'test_input')
test_output_dir = os.path.join(str(tmpdir), 'test_output')
tests_log_path = os.path.join(code_dir, 'tests_scores_log.txt')
shutil.copytree(fake_input_dir, test_input_dir)
# Transformation
train_files = glob_train_files(test_input_dir)
adapt_files = glob_adapt_files(test_input_dir, 10000)
for data_file in adapt_files + train_files:
X, t = load_h5_input_file(data_file)
# Corruption
if np.random.rand() <= 0.1:
X = np.array([])
os.unlink(data_file)
save_h5_input_file(data_file, X, t, output_dim=2)
# Evaluation
run_predict_on_test_dataset_and_output_score('test_empty_X_10%', code_dir,
fake_input_dir,
test_input_dir,
test_output_dir,
tests_log_path)
def test_load_predict_matrixes_corrupted_t(fake_input_dir,
tmpdir,
hist_length=64 * 12 * 12):
# Configuration
test_input_dir = os.path.join(str(tmpdir), 'test_input')
shutil.copytree(fake_input_dir, test_input_dir)
# Transformation
train_files = glob_train_files(test_input_dir)
adapt_files = glob_adapt_files(test_input_dir, 10000)
for data_file in adapt_files + train_files:
X, t = load_h5_input_file(data_file)
# Corruption
t = corrupt_t(t)
os.unlink(data_file)
save_h5_input_file(data_file, X, t, output_dim=2)
# Evaluation
for eval_step in range(-1, 60, 5):
original_X = load_data_matrix(fake_input_dir,
eval_step=eval_step,
hist_length=hist_length)
X_with_corrupted_t = load_data_matrix(test_input_dir,
eval_step=eval_step,
hist_length=hist_length)
np.testing.assert_array_equal(original_X, X_with_corrupted_t)
def test_load_full_cache_matrixes_corrupted_t(fake_input_dir,
tmpdir,
hist_length=64 * 12 * 12):
# Configuration
test_input_dir = os.path.join(str(tmpdir), 'test_input')
shutil.copytree(fake_input_dir, test_input_dir)
# Transformation
train_files = glob_train_files(test_input_dir)
adapt_files = glob_adapt_files(test_input_dir, 10000)
for data_file in adapt_files + train_files:
X, t = load_h5_input_file(data_file)
# Corruption
t = corrupt_t(t)
os.unlink(data_file)
save_h5_input_file(data_file, X, t, output_dim=2)
# Evaluation
for eval_step in range(-1, 60, 5):
original_X, original_t = load_full_cache_matrixes(
fake_input_dir, eval_step=eval_step, steps_needed=hist_length)
X_with_corrupted_t, corrupted_t = load_full_cache_matrixes(
test_input_dir, eval_step=eval_step, steps_needed=hist_length)
assert len(corrupted_t) == len(original_t)
np.testing.assert_array_equal(original_X, X_with_corrupted_t)
np.testing.assert_array_equal(original_t[~np.isnan(corrupted_t)],
corrupted_t[~np.isnan(corrupted_t)])
np.testing.assert_array_equal(original_t, fillna_t(corrupted_t))
def test_t_diff_len_from_X(fake_input_dir, hist_length=16 * 12 * 12):
data_file = os.path.join(fake_input_dir, 'adapt', 'X0.h5')
X, t = load_h5_input_file(data_file)
t = t[:-1]
os.unlink(data_file)
save_h5_input_file(data_file, X, t, output_dim=2)
new_X, new_t = load_h5_input_file(data_file)
assert new_X.shape[0] == new_t.shape[0]
np.testing.assert_array_equal(X, new_X)
np.testing.assert_array_equal(new_t, np.append(t, np.array([np.nan])))
def test_read_ndim_2(fake_input_dir):
train_files = glob_train_files(fake_input_dir)
adapt_files = glob_adapt_files(fake_input_dir, 10000)
for data_file in adapt_files + train_files:
X, t = load_h5_input_file(data_file)
print(X.shape)
# Corruption
os.unlink(data_file)
save_h5_input_file(data_file, X, t, output_dim=2)
X_copy, t_copy = load_h5_input_file(data_file)
assert len(X_copy.shape) == 2
def test_t_has_zeros(fake_input_dir, hist_length=16 * 12 * 12):
data_file = os.path.join(fake_input_dir, 'adapt', 'X0.h5')
X, t = load_h5_input_file(data_file)
corrupted_t = t.copy()
corrupted_t[np.random.choice(len(t), len(t) // 2, replace=False)] = 0
os.unlink(data_file)
save_h5_input_file(data_file, X, corrupted_t, output_dim=2)
new_X, new_t = load_h5_input_file(data_file)
assert new_X.shape[0] == new_t.shape[0]
np.testing.assert_array_equal(X, new_X)
corrupted_t_equivalent = corrupted_t.copy()
corrupted_t_equivalent[corrupted_t_equivalent == 0] = np.nan
np.testing.assert_array_equal(new_t, corrupted_t_equivalent)
np.testing.assert_array_equal(t, fillna_t(new_t))
def test_fill_timesteps_with_nan(tmpdir, fake_input_dir, output_dir, code_dir):
# Configuration
test_input_dir = os.path.join(str(tmpdir), 'test_input')
test_output_dir = os.path.join(str(tmpdir), 'test_output')
tests_log_path = os.path.join(code_dir, 'tests_scores_log.txt')
shutil.copytree(fake_input_dir, test_input_dir)
# Transformation
adapt_files = glob_adapt_files(test_input_dir, 10000)
for adapt_file in adapt_files:
X, t = load_h5_input_file(adapt_file)
# Corruption
X, t = fill_with_nan(X, t, None)
os.unlink(adapt_file)
save_h5_input_file(adapt_file, X, t, output_dim=2)
# Evaluation
run_predict_on_test_dataset_and_output_score(
'test_fill_timesteps_with_nan', code_dir, fake_input_dir,
test_input_dir, test_output_dir, tests_log_path)
def test_X_change_shape_no_predict(fake_input_dir,
tmpdir,
hist_length=64 * 12 * 12):
# Transformation
train_files = glob_train_files(fake_input_dir)
adapt_files = glob_adapt_files(fake_input_dir, 10000)
for data_file in adapt_files + train_files:
X, t = load_h5_input_file(data_file)
# Corruption
if np.random.rand() <= 0.5:
keep_size = X.shape[1] - 300
X = X[:, :keep_size]
os.unlink(data_file)
save_h5_input_file(data_file, X, t, output_dim=2)
# Evaluation
cache_X = None
cache_t = None
for eval_step in range(-1, 30, 1):
original_X, last_time_step, cache_X, cache_t = load_prediction_matrix(
fake_input_dir, fake_input_dir, fake_input_dir, fake_input_dir,
eval_step, eval_step - 1, hist_length, cache_X, cache_t)
def test_sequential_load_corrupted_t(fake_input_dir,
tmpdir,
corrupt_fn_train,
corrupt_fn_adapt,
hist_length=64 * 12 * 12):
# Configuration
test_input_dir = os.path.join(str(tmpdir), 'test_input')
shutil.copytree(fake_input_dir, test_input_dir)
# Transformation
train_files = glob_train_files(test_input_dir)
adapt_files = glob_adapt_files(test_input_dir, 10000)
for data_file in adapt_files + train_files:
X, t = load_h5_input_file(data_file)
# Corruption
if data_file in train_files:
X, t = corrupt_fn_train(X, t)
else:
X, t = corrupt_fn_adapt(X, t)
os.unlink(data_file)
save_h5_input_file(data_file, X, t, output_dim=2)
# Evaluation
cache_X = None
cache_t = None
cache_X_corrupted = None
cache_t_corrupted = None
for eval_step in range(-1, 30, 1):
original_X, last_time_step, cache_X, cache_t = load_prediction_matrix(
fake_input_dir, fake_input_dir, fake_input_dir, fake_input_dir,
eval_step, eval_step - 1, hist_length, cache_X, cache_t)
corrupted_X, last_time_step, cache_X_corrupted, cache_t_corrupted = load_prediction_matrix(
test_input_dir, test_input_dir, test_input_dir, test_input_dir,
eval_step, eval_step - 1, hist_length, cache_X_corrupted,
cache_t_corrupted)
np.testing.assert_array_equal(original_X, corrupted_X)
def test_remove_half_lines(tmpdir, fake_input_dir, output_dir, code_dir):
# Configuration
test_input_dir = os.path.join(str(tmpdir), 'test_input')
test_output_dir = os.path.join(str(tmpdir), 'test_output')
tests_log_path = os.path.join(code_dir, 'tests_scores_log.txt')
shutil.copytree(fake_input_dir, test_input_dir)
# Transformation
train_files = glob_train_files(test_input_dir)
adapt_files = glob_adapt_files(test_input_dir, 10000)
for data_file in adapt_files + train_files:
X, t = load_h5_input_file(data_file)
# Corruption
X, t = remove_half_lines(X, t)
os.unlink(data_file)
save_h5_input_file(data_file, X, t, output_dim=2)
# Evaluation
run_predict_on_test_dataset_and_output_score('test_remove_half_lines',
code_dir, test_input_dir,
test_input_dir,
test_output_dir,
tests_log_path)
