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_non_writable_output_path(test_dir, data_frame):
label_col = 'label'
df = data_frame(n_samples=100, label_col=label_col)
data_encoder_cols = [TfIdfEncoder('features')]
label_encoder_cols = [CategoricalEncoder(label_col)]
data_cols = [BowFeaturizer('features')]
output_path = os.path.join(test_dir, 'non_writable')
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).save()
from datawig.utils import logger
try:
# make output dir of imputer read-only
os.chmod(output_path, S_IREAD | S_IXUSR)
# make log file read only
os.chmod(os.path.join(output_path, "imputer.log"), S_IREAD)
imputer = Imputer.load(output_path)
_ = imputer.predict(df)
logger.warning("this should not fail")
# remove log file
os.chmod(os.path.join(output_path, "imputer.log"),
S_IREAD | S_IXUSR | S_IWUSR)
os.chmod(output_path, S_IREAD | S_IXUSR | S_IWUSR)
os.remove(os.path.join(output_path, "imputer.log"))
# make output dir of imputer read-only
os.chmod(output_path, S_IREAD | S_IXUSR)
imputer = Imputer.load(output_path)
_ = imputer.predict(df)
logger.warning("this should not fail")
os.chmod(output_path, S_IREAD | S_IXUSR | S_IWUSR)
except Exception as e:
print(e)
pytest.fail("This invocation not raise any Exception")
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_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_explain_method_synthetic(test_dir):
# Generate simulated data for testing explain method
# Predict output column with entries in ['foo', 'bar'] from two columns, one
# categorical in ['foo', 'dummy'], one text in ['text_foo_text', 'text_dummy_text'].
# the output column is deterministically 'foo', if 'foo' occurs anywhere in any input column.
N = 100
cat_in_col = ['foo' if r > (1 / 2) else 'dummy' for r in np.random.rand(N)]
text_in_col = ['fff' if r > (1 / 2) else 'ddd' for r in np.random.rand(N)]
hash_in_col = ['h' for r in range(N)]
cat_out_col = [
'foo' if 'f' in input[0] + input[1] else 'bar'
for input in zip(cat_in_col, text_in_col)
]
df = pd.DataFrame()
df['in_cat'] = cat_in_col
df['in_text'] = text_in_col
df['in_text_hash'] = hash_in_col
df['out_cat'] = cat_out_col
# Specify encoders and featurizers #
data_encoder_cols = [
datawig.column_encoders.TfIdfEncoder('in_text', tokens="chars"),
datawig.column_encoders.CategoricalEncoder('in_cat', max_tokens=10),
datawig.column_encoders.BowEncoder('in_text_hash', tokens="chars")
]
data_featurizer_cols = [
datawig.mxnet_input_symbols.BowFeaturizer('in_text'),
datawig.mxnet_input_symbols.EmbeddingFeaturizer('in_cat'),
datawig.mxnet_input_symbols.BowFeaturizer('in_text_hash')
]
label_encoder_cols = [
datawig.column_encoders.CategoricalEncoder('out_cat')
]
# Specify model
imputer = datawig.Imputer(data_featurizers=data_featurizer_cols,
label_encoders=label_encoder_cols,
data_encoders=data_encoder_cols,
output_path=os.path.join(test_dir, "tmp",
"explanation_tests"))
# Train
tr, te = random_split(df.sample(90), [.8, .2])
imputer.fit(train_df=tr, test_df=te, num_epochs=20, learning_rate=1e-2)
predictions = imputer.predict(te)
# Evaluate
assert precision_score(predictions.out_cat,
predictions.out_cat_imputed,
average='weighted') > .99
# assert item explanation, iterate over some inputs
for i in np.random.choice(N, 10):
explanation = imputer.explain_instance(df.iloc[i])
top_label = explanation['explained_label']
if top_label == 'bar':
assert (explanation['in_text'][0][0] == 'd'
and explanation['in_cat'][0][0] == 'dummy')
elif top_label == 'foo':
assert (explanation['in_text'][0][0] == 'f'
or explanation['in_cat'][0][0] == 'foo')
# assert class explanations
assert np.all([
'f' in token for token, weight in imputer.explain('foo')['in_text']
][:3])
assert [
'f' in token for token, weight in imputer.explain('foo')['in_cat']
][0]
# test serialisation to disk
imputer.save()
imputer_from_disk = Imputer.load(imputer.output_path)
assert np.all([
'f' in token
for token, weight in imputer_from_disk.explain('foo')['in_text']
][:3])
