def test_inceptionV3_featurization(self):
output_col = "prediction"
transformer = DeepImageFeaturizer(inputCol="image", outputCol=output_col,
modelName="InceptionV3")
image_df = getSampleImageDF()
transformed_df = transformer.transform(image_df.limit(5))
collected = transformed_df.collect()
for row in collected:
predictions = row[output_col]
self.assertEqual(len(predictions), InceptionV3Constants.NUM_OUTPUT_FEATURES)
def test_featurization(self):
"""
Tests that featurizer returns (almost) the same values as Keras.
"""
output_col = "prediction"
transformer = DeepImageFeaturizer(inputCol="image", outputCol=output_col,
modelName=self.name)
transformed_df = transformer.transform(self.imageDF)
collected = transformed_df.collect()
features = np.array([i.prediction for i in collected])
# Note: keras features may be multi-dimensional np arrays, but transformer features
# will be 1-d vectors. Regardless, the dimensions should add up to the same.
self.assertEqual(np.prod(self.kerasFeatures.shape), np.prod(features.shape))
kerasReshaped = self.kerasFeatures.reshape(self.kerasFeatures.shape[0], -1)
np.testing.assert_array_almost_equal(kerasReshaped, features, decimal=6)
def test_featurization(self):
"""
Tests that featurizer returns (almost) the same values as Keras.
"""
# Since we use different libraries for image resizing (PIL in python vs. java.awt.Image in scala),
# the result will not match keras exactly. In fact the best we can do is a "somewhat similar" result.
# At least compare cosine distance is < 1e-2
featurizer_sc = DeepImageFeaturizer(modelName=self.name, inputCol="image",
outputCol="features", scaleHint="SCALE_FAST")
features_sc = np.array([i.features for i in featurizer_sc.transform(
self.imageDF).select("features").collect()])
kerasReshaped = self.kerasFeatures.reshape(self.kerasFeatures.shape[0], -1)
diffs = [
spatial.distance.cosine(
kerasReshaped[i],
features_sc[i]) for i in range(
len(features_sc))]
np.testing.assert_array_almost_equal(0, diffs, decimal=self.featurizerCompareDigitsCosine)
def test_inception(self):
transformer0 = DeepImageFeaturizer(inputCol='image', modelName="InceptionV3",
outputCol="features0", scaleHint="SCALE_FAST")
dst_path = os.path.join(self.tempdir, "featurizer")
transformer0.save(dst_path)
transformer1 = DeepImageFeaturizer.load(dst_path)
self.assertEqual(transformer0.uid, transformer1.uid)
self.assertEqual(type(transformer0.uid), type(transformer1.uid))
for x in transformer0._paramMap.keys():
self.assertEqual(transformer1.uid, x.parent,
"Loaded DeepImageFeaturizer instance uid (%s) did not match Param's uid (%s)"
% (transformer1.uid, transformer1.scaleHint.parent))
self.assertEqual(transformer0._paramMap, transformer1._paramMap,
"Loaded DeepImageFeaturizer instance params (%s) did not match "
% str(transformer1._paramMap) +
"original values (%s)" % str(transformer0._paramMap))
self.assertEqual(transformer0._defaultParamMap, transformer1._defaultParamMap,
"Loaded DeepImageFeaturizer instance default params (%s) did not match "
% str(transformer1._defaultParamMap) +
"original defaults (%s)" % str(transformer0._defaultParamMap))
def test_featurization_no_reshape(self):
"""
Run sparkDL predictor on manually-resized images and compare result to the
keras result.
"""
imageArray = self.imageArray
# test: predictor vs keras on resized images
rdd = self.sc.parallelize([self._rowWithImage(img) for img in imageArray])
dfType = ImageSchema.imageSchema
imageDf = rdd.toDF(dfType)
if self.numPartitionsOverride:
imageDf = imageDf.coalesce(self.numPartitionsOverride)
transformer = DeepImageFeaturizer(inputCol='image', modelName=self.name,
outputCol="features")
dfFeatures = transformer.transform(imageDf).collect()
dfFeatures = np.array([i.features for i in dfFeatures])
kerasReshaped = self.kerasFeatures.reshape(self.kerasFeatures.shape[0], -1)
np.testing.assert_array_almost_equal(kerasReshaped,
dfFeatures,
decimal=self.featurizerCompareDigitsExact)
def test_featurization(self):
"""
Tests that featurizer returns (almost) the same values as Keras.
"""
output_col = "prediction"
transformer = DeepImageFeaturizer(inputCol="image",
outputCol=output_col,
modelName=self.name)
transformed_df = transformer.transform(self.imageDF)
collected = transformed_df.collect()
features = np.array([i.prediction for i in collected])
# Note: keras features may be multi-dimensional np arrays, but transformer features
# will be 1-d vectors. Regardless, the dimensions should add up to the same.
self.assertEqual(np.prod(self.kerasFeatures.shape),
np.prod(features.shape))
kerasReshaped = self.kerasFeatures.reshape(self.kerasFeatures.shape[0],
-1)
np.testing.assert_array_almost_equal(kerasReshaped,
features,
decimal=6)
def test_featurizer_in_pipeline(self):
"""
Tests that featurizer fits into an MLlib Pipeline.
Does not test how good the featurization is for generalization.
"""
featurizer = DeepImageFeaturizer(inputCol="image", outputCol="features",
modelName=self.name)
lr = LogisticRegression(maxIter=20, regParam=0.05, elasticNetParam=0.3, labelCol="label")
pipeline = Pipeline(stages=[featurizer, lr])
# add arbitrary labels to run logistic regression
# TODO: it's weird that the test fails on some combinations of labels. check why.
label_udf = udf(lambda x: abs(hash(x)) % 2, IntegerType())
train_df = self.imageDF.withColumn("label", label_udf(self.imageDF["image"]["origin"]))
lrModel = pipeline.fit(train_df)
# see if we at least get the training examples right.
# with 5 examples and e.g. 131k features (for InceptionV3), it ought to.
pred_df_collected = lrModel.transform(train_df).collect()
for row in pred_df_collected:
self.assertEqual(int(row.prediction), row.label)