def test_imputer_init():
with pytest.raises(ValueError) as e:
imputer = Imputer(data_featurizers='item_name', label_encoders=['brand'], data_encoders='')
with pytest.raises(ValueError) as e:
imputer = Imputer(data_featurizers=[BowFeaturizer('item_name')],
label_encoders="brand",
data_encoders='')
with pytest.raises(ValueError) as e:
imputer = Imputer(data_featurizers=[BowFeaturizer('item_name')],
label_encoders=[CategoricalEncoder("brand")],
data_encoders='')
with pytest.raises(ValueError) as e:
imputer = Imputer(data_featurizers=[BowFeaturizer('item_name')],
label_encoders=[CategoricalEncoder("brand")],
data_encoders=[BowEncoder('not_in_featurizers')])
with pytest.raises(ValueError) as e:
imputer = Imputer(data_featurizers=[BowFeaturizer('item_name')],
label_encoders=[CategoricalEncoder("brand")],
data_encoders=[BowEncoder('brand')])
label_encoders = [CategoricalEncoder('brand', max_tokens=10)]
data_featurizers = [LSTMFeaturizer('item_name'), EmbeddingFeaturizer('manufacturer')]
data_encoders = [
SequentialEncoder(
'item_name'
),
CategoricalEncoder(
'manufacturer'
)
]
imputer = Imputer(
data_featurizers=data_featurizers,
label_encoders=label_encoders,
data_encoders=data_encoders
)
assert imputer.output_path == "brand"
assert imputer.module_path == 'brand/model'
assert imputer.metrics_path == 'brand/fit-test-metrics.json'
assert imputer.output_path == "brand"
assert imputer.module_path == 'brand/model'
assert imputer.metrics_path == 'brand/fit-test-metrics.json'
imputer = Imputer(
data_featurizers=data_featurizers,
label_encoders=[CategoricalEncoder('B Rand', max_tokens=10)],
data_encoders=data_encoders
)
assert imputer.output_path == "b_rand"
shutil.rmtree("b_rand")
def test_imputer_unrepresentative_test_df(test_dir, data_frame):
"""
Tests whether the imputer runs through in cases when test data set (and hence metrics and precision/recall curves)
doesn't contain values present in training data
"""
# generate some random data
random_data = data_frame(n_samples=100)
df_train, df_test, _ = random_split(random_data, [.8, .1, .1])
excluded = df_train['labels'].values[0]
df_test = df_test[df_test['labels'] != excluded]
data_encoder_cols = [BowEncoder('features')]
label_encoder_cols = [CategoricalEncoder('labels')]
data_cols = [BowFeaturizer('features')]
output_path = os.path.join(test_dir, "tmp", "real_data_experiment")
imputer = Imputer(data_featurizers=data_cols,
label_encoders=label_encoder_cols,
data_encoders=data_encoder_cols,
output_path=output_path).fit(train_df=df_train,
test_df=df_test,
num_epochs=10)
only_excluded_df = df_train[df_train['labels'] == excluded]
imputations = imputer.predict_above_precision(
only_excluded_df, precision_threshold=.99)['labels']
assert all([x == () for x in imputations])
def test_imputer_load_read_exec_only_dir(tmpdir, data_frame):
import stat
# on shared build-fleet tests fail with converting tmpdir to string
tmpdir = str(tmpdir)
feature = 'feature'
label = 'label'
df = data_frame(feature, label, n_samples=100)
# fit and output model + metrics to tmpdir
imputer = Imputer(data_featurizers=[BowFeaturizer(feature)],
label_encoders=[CategoricalEncoder(label)],
data_encoders=[BowEncoder(feature)],
output_path=tmpdir)
imputer.fit(train_df=df, num_epochs=1)
# make tmpdir read/exec-only by owner/group/others
os.chmod(
tmpdir, stat.S_IEXEC | stat.S_IXGRP | stat.S_IXOTH | stat.S_IREAD
| stat.S_IRGRP | stat.S_IROTH)
try:
Imputer.load(tmpdir)
except AssertionError as e:
print(e)
pytest.fail(
'Loading imputer from read-only directory should not fail.')
def test_imputer_without_train_df(test_dir):
"""
Test asserting that imputer.fit fails without training data or training data in wrong format
"""
df_train = ['ffffffooooo']
data_encoder_cols = [BowEncoder('item_name')]
label_encoder_cols = [CategoricalEncoder('brand')]
data_cols = [BowFeaturizer('item_name')]
output_path = os.path.join(test_dir, "tmp", "real_data_experiment")
imputer = Imputer(
data_featurizers=data_cols,
label_encoders=label_encoder_cols,
data_encoders=data_encoder_cols,
output_path=output_path,
)
with pytest.raises(
ValueError,
message="Need a non-empty DataFrame for fitting Imputer model"):
imputer.fit(train_df=df_train)
with pytest.raises(
ValueError,
message="Need a non-empty DataFrame for fitting Imputer model"):
imputer.fit(train_df=None)
def test_not_explainable(test_dir, data_frame):
label_col = 'label'
df = data_frame(n_samples=100, label_col=label_col)
data_encoder_cols = [BowEncoder('features')]
label_encoder_cols = [CategoricalEncoder(label_col)]
data_cols = [BowFeaturizer('features')]
output_path = os.path.join(test_dir, "tmp", "out")
imputer = Imputer(data_featurizers=data_cols,
label_encoders=label_encoder_cols,
data_encoders=data_encoder_cols,
output_path=output_path).fit(train_df=df, num_epochs=1)
assert not imputer.is_explainable
try:
imputer.explain('some label')
raise pytest.fail(
'imputer.explain should fail with an appropriate error message')
except ValueError as exception:
assert exception.args[0] == 'No explainable data encoders available.'
instance = pd.Series({'features': 'some feature text'})
try:
imputer.explain_instance(instance)
raise pytest.fail(
'imputer.explain_instance should fail with an appropriate error message'
)
except ValueError as exception:
assert exception.args[0] == 'No explainable data encoders available.'
def test_drop_missing():
"""
Tests some private functions of the Imputer class
"""
df_train = pd.DataFrame(
{'label': [1, None, np.nan, 2] * 4, 'data': ['bla', 'drop', 'drop', 'fasl'] * 4})
df_test = df_train.copy()
max_tokens = int(2 ** 15)
batch_size = 16
data_encoder_cols = [BowEncoder('data', max_tokens=max_tokens)]
label_encoder_cols = [CategoricalEncoder('label', max_tokens=1)]
data_cols = [BowFeaturizer('data', vocab_size=max_tokens)]
output_path = os.path.join(dir_path, "resources", "tmp", "real_data_experiment")
imputer = Imputer(
data_featurizers=data_cols,
label_encoders=label_encoder_cols,
data_encoders=data_encoder_cols,
output_path=output_path
).fit(
train_df=df_train,
test_df=df_test,
batch_size=batch_size
)
df_dropped = imputer._Imputer__drop_missing_labels(df_train, how='any')
df_dropped_true = pd.DataFrame({'data': {3: 'fasl', 7: 'fasl', 11: 'fasl', 15: 'fasl'},
'label': {3: 2.0, 7: 2.0, 11: 2.0, 15: 2.0}})
assert df_dropped[['data', 'label']].equals(df_dropped_true[['data', 'label']])
def test_imputer_without_test_set_random_split():
"""
Test asserting that the random split is working internally
by calling imputer.fit only with a training set.
"""
feature_col = "string_feature"
label_col = "label"
n_samples = 5000
num_labels = 3
seq_len = 20
vocab_size = int(2 ** 10)
# generate some random data
df_train = generate_string_data_frame(feature_col=feature_col,
label_col=label_col,
vocab_size=vocab_size,
num_labels=num_labels,
num_words=seq_len,
n_samples=n_samples)
num_epochs = 1
batch_size = 64
learning_rate = 1e-3
data_encoder_cols = [
BowEncoder(feature_col, max_tokens=vocab_size)
]
label_encoder_cols = [CategoricalEncoder(label_col, max_tokens=num_labels)]
data_cols = [
BowFeaturizer(feature_col, vocab_size=vocab_size)
]
output_path = os.path.join(dir_path, "resources", "tmp", "real_data_experiment")
imputer = Imputer(
data_featurizers=data_cols,
label_encoders=label_encoder_cols,
data_encoders=data_encoder_cols,
output_path=output_path
)
try:
imputer.fit(
train_df=df_train,
learning_rate=learning_rate,
num_epochs=num_epochs,
batch_size=batch_size
)
except TypeError:
pytest.fail("Didn't expect a TypeError exception with missing test data")
shutil.rmtree(output_path)
def test_iter_padding_offset():
col = 'brand'
df = pd.DataFrame([{
col: brand
} for brand in list(
map(lambda e: str(int(e)), np.random.exponential(scale=1, size=36)))])
df_train = df.sample(frac=0.5)
it = ImputerIterDf(df_train,
data_columns=[BowEncoder(col)],
label_columns=[CategoricalEncoder(col, max_tokens=5)],
batch_size=32)
assert it.start_padding_idx == df_train.shape[0]
def test_imputer_load_with_invalid_context(tmpdir, data_frame):
# on shared build-fleet tests fail with converting tmpdir to string
tmpdir = str(tmpdir)
feature = 'feature'
label = 'label'
df = data_frame(feature, label, n_samples=100)
# fit and output model + metrics to tmpdir
imputer = Imputer(data_featurizers=[BowFeaturizer(feature)],
label_encoders=[CategoricalEncoder(label)],
data_encoders=[BowEncoder(feature)],
output_path=tmpdir)
imputer.fit(train_df=df, num_epochs=1)
imputer.ctx = None
imputer.save()
imputer_deser = Imputer.load(tmpdir)
_ = imputer_deser.predict(df)
def test_imputer_fit_fail_non_writable_output_dir(tmpdir, data_frame):
import stat
# on shared build-fleet tests fail with converting tmpdir to string
tmpdir = str(tmpdir)
feature = 'feature'
label = 'label'
df = data_frame(feature, label, n_samples=100)
# fit and output model + metrics to tmpdir
imputer = Imputer(data_featurizers=[BowFeaturizer(feature)],
label_encoders=[CategoricalEncoder(label)],
data_encoders=[BowEncoder(feature)],
output_path=tmpdir)
# make tmpdir read/exec-only by owner/group/others
os.chmod(
tmpdir, stat.S_IEXEC | stat.S_IXGRP | stat.S_IXOTH | stat.S_IREAD
| stat.S_IRGRP | stat.S_IROTH)
# fail if imputer.fit does not raise an AssertionError
with pytest.raises(AssertionError) as e:
imputer.fit(df, num_epochs=1)
def test_mxnet_module_wrapper(data_frame):
from datawig.imputer import _MXNetModule
import mxnet as mx
from datawig.iterators import ImputerIterDf
feature_col, label_col = "feature", "label"
df = data_frame(n_samples=100,
feature_col=feature_col,
label_col=label_col)
label_encoders = [CategoricalEncoder(label_col)]
data_encoders = [BowEncoder(feature_col)]
data_featurizers = [BowFeaturizer(feature_col, vocab_size=100)]
iter_train = ImputerIterDf(df, data_encoders, label_encoders)
mod = _MXNetModule(mx.current_context(),
label_encoders,
data_featurizers,
final_fc_hidden_units=[])(iter_train)
assert mod._label_names == [label_col]
assert mod.data_names == [feature_col]
# weights and biases
assert len(mod._arg_params) == 2
def get_new_iterator_df_bow(df):
return ImputerIterDf(
df,
data_columns=[BowEncoder(feature_col, max_tokens=max_tokens)],
label_columns=[CategoricalEncoder(label_col, max_tokens=num_labels)],
batch_size=2)
def test_imputer_real_data_all_featurizers(test_dir, data_frame):
"""
Tests Imputer with sequential, bag-of-words and categorical variables as inputs
this could be run as part of integration test suite.
"""
feature_col = "string_feature"
categorical_col = "categorical_feature"
label_col = "label"
n_samples = 5000
num_labels = 3
seq_len = 20
vocab_size = int(2**10)
latent_dim = 30
embed_dim = 30
# generate some random data
random_data = data_frame(feature_col=feature_col,
label_col=label_col,
vocab_size=vocab_size,
num_labels=num_labels,
num_words=seq_len,
n_samples=n_samples)
# we use a the label prefixes as a dummy categorical input variable
random_data[categorical_col] = random_data[label_col].apply(
lambda x: x[:2])
df_train, df_test, df_val = random_split(random_data, [.8, .1, .1])
data_encoder_cols = [
BowEncoder(feature_col, feature_col + "_bow", max_tokens=vocab_size),
SequentialEncoder(feature_col,
feature_col + "_lstm",
max_tokens=vocab_size,
seq_len=seq_len),
CategoricalEncoder(categorical_col, max_tokens=num_labels)
]
label_encoder_cols = [CategoricalEncoder(label_col, max_tokens=num_labels)]
data_cols = [
BowFeaturizer(feature_col + "_bow", vocab_size=vocab_size),
LSTMFeaturizer(field_name=feature_col + "_lstm",
seq_len=seq_len,
latent_dim=latent_dim,
num_hidden=30,
embed_dim=embed_dim,
num_layers=2,
vocab_size=num_labels),
EmbeddingFeaturizer(field_name=categorical_col,
embed_dim=embed_dim,
vocab_size=num_labels)
]
output_path = os.path.join(test_dir, "tmp",
"imputer_experiment_synthetic_data")
num_epochs = 10
batch_size = 32
learning_rate = 1e-2
imputer = Imputer(data_featurizers=data_cols,
label_encoders=label_encoder_cols,
data_encoders=data_encoder_cols,
output_path=output_path).fit(train_df=df_train,
test_df=df_val,
learning_rate=learning_rate,
num_epochs=num_epochs,
batch_size=batch_size,
calibrate=False)
len_df_before_predict = len(df_test)
pred = imputer.transform(df_test)
assert len(pred[label_col]) == len_df_before_predict
assert sum(df_test[label_col].values == pred[label_col]) == len(df_test)
_ = imputer.predict_proba_top_k(df_test, top_k=2)
_, metrics = imputer.transform_and_compute_metrics(df_test)
assert metrics[label_col]['avg_f1'] > 0.9
deserialized = Imputer.load(imputer.output_path)
_, metrics_deserialized = deserialized.transform_and_compute_metrics(
df_test)
assert metrics_deserialized[label_col]['avg_f1'] > 0.9
# training on a small data set to get a imputer with low precision
not_so_precise_imputer = Imputer(data_featurizers=data_cols,
label_encoders=label_encoder_cols,
data_encoders=data_encoder_cols,
output_path=output_path).fit(
train_df=df_train[:50],
test_df=df_test,
learning_rate=learning_rate,
num_epochs=num_epochs,
batch_size=batch_size,
calibrate=False)
df_test = df_test.reset_index()
predictions_df = not_so_precise_imputer.predict(
df_test, precision_threshold=.5, imputation_suffix="_imputed")
assert predictions_df.columns.contains(label_col + "_imputed")
assert predictions_df.columns.contains(label_col + "_imputed_proba")
def test_imputer_duplicate_encoder_output_columns(test_dir, data_frame):
"""
Tests Imputer with sequential, bag-of-words and categorical variables as inputs
this could be run as part of integration test suite.
"""
feature_col = "string_feature"
categorical_col = "categorical_feature"
label_col = "label"
n_samples = 1000
num_labels = 10
seq_len = 100
vocab_size = int(2**10)
latent_dim = 30
embed_dim = 30
# generate some random data
random_data = data_frame(feature_col=feature_col,
label_col=label_col,
vocab_size=vocab_size,
num_labels=num_labels,
num_words=seq_len,
n_samples=n_samples)
# we use a the label prefixes as a dummy categorical input variable
random_data[categorical_col] = random_data[label_col].apply(
lambda x: x[:2])
df_train, df_test, df_val = random_split(random_data, [.8, .1, .1])
data_encoder_cols = [
BowEncoder(feature_col, feature_col, max_tokens=vocab_size),
SequentialEncoder(feature_col,
feature_col,
max_tokens=vocab_size,
seq_len=seq_len),
CategoricalEncoder(categorical_col, max_tokens=num_labels)
]
label_encoder_cols = [CategoricalEncoder(label_col, max_tokens=num_labels)]
data_cols = [
BowFeaturizer(feature_col, vocab_size=vocab_size),
LSTMFeaturizer(field_name=feature_col,
seq_len=seq_len,
latent_dim=latent_dim,
num_hidden=30,
embed_dim=embed_dim,
num_layers=2,
vocab_size=num_labels),
EmbeddingFeaturizer(field_name=categorical_col,
embed_dim=embed_dim,
vocab_size=num_labels)
]
output_path = os.path.join(test_dir, "tmp",
"imputer_experiment_synthetic_data")
num_epochs = 20
batch_size = 16
learning_rate = 1e-3
with pytest.raises(ValueError) as e:
imputer = Imputer(data_featurizers=data_cols,
label_encoders=label_encoder_cols,
data_encoders=data_encoder_cols,
output_path=output_path)
imputer.fit(train_df=df_train,
test_df=df_val,
learning_rate=learning_rate,
num_epochs=num_epochs,
batch_size=batch_size)
def test_automatic_calibration(data_frame):
"""
Fit model with all featurisers and assert
that calibration improves the expected calibration error.
"""
feature_col = "string_feature"
categorical_col = "categorical_feature"
label_col = "label"
n_samples = 2000
num_labels = 3
seq_len = 20
vocab_size = int(2**10)
latent_dim = 30
embed_dim = 30
# generate some random data
random_data = data_frame(feature_col=feature_col,
label_col=label_col,
vocab_size=vocab_size,
num_labels=num_labels,
num_words=seq_len,
n_samples=n_samples)
# we use a the label prefixes as a dummy categorical input variable
random_data[categorical_col] = random_data[label_col].apply(
lambda x: x[:2])
df_train, df_test, df_val = random_split(random_data, [.8, .1, .1])
data_encoder_cols = [
BowEncoder(feature_col, feature_col + "_bow", max_tokens=vocab_size),
SequentialEncoder(feature_col,
feature_col + "_lstm",
max_tokens=vocab_size,
seq_len=seq_len),
CategoricalEncoder(categorical_col, max_tokens=num_labels)
]
label_encoder_cols = [CategoricalEncoder(label_col, max_tokens=num_labels)]
data_cols = [
BowFeaturizer(feature_col + "_bow", vocab_size=vocab_size),
LSTMFeaturizer(field_name=feature_col + "_lstm",
seq_len=seq_len,
latent_dim=latent_dim,
num_hidden=30,
embed_dim=embed_dim,
num_layers=2,
vocab_size=num_labels),
EmbeddingFeaturizer(field_name=categorical_col,
embed_dim=embed_dim,
vocab_size=num_labels)
]
num_epochs = 20
batch_size = 32
learning_rate = 1e-2
imputer = Imputer(data_featurizers=data_cols,
label_encoders=label_encoder_cols,
data_encoders=data_encoder_cols).fit(
train_df=df_train,
test_df=df_val,
learning_rate=learning_rate,
num_epochs=num_epochs,
batch_size=batch_size)
assert imputer.calibration_info['ece_pre'] > imputer.calibration_info[
'ece_post']
