def _test(array):
height, width, chan = array.shape
imgAsStruct = imageIO.imageArrayToStruct(array)
self.assertEqual(imgAsStruct.height, height)
self.assertEqual(imgAsStruct.width, width)
self.assertEqual(imgAsStruct.data, array.tobytes())
imgReconstructed = imageIO.imageStructToArray(imgAsStruct)
np.testing.assert_array_equal(array, imgReconstructed)
def _test(array):
height, width, chan = array.shape
imgAsStruct = imageIO.imageArrayToStruct(array)
self.assertEqual(imgAsStruct.height, height)
self.assertEqual(imgAsStruct.width, width)
self.assertEqual(imgAsStruct.data, array.tobytes())
imgReconstructed = imageIO.imageStructToArray(imgAsStruct)
np.testing.assert_array_equal(array, imgReconstructed)
def _getNumpyFeaturesAndLabels(self, dataset):
"""
We assume the training data fits in memory on a single server.
The input dataframe is converted to numerical image features and
broadcast to all the worker nodes.
"""
image_uri_col = self.getInputCol()
label_col = None
if self.isDefined(self.labelCol) and self.getLabelCol() != "":
label_col = self.getLabelCol()
tmp_image_col = self._loadedImageCol()
image_df = self.loadImagesInternal(
dataset, image_uri_col).dropna(subset=[tmp_image_col])
# Extract features
localFeatures = []
rows = image_df.collect()
for row in rows:
spimg = row[tmp_image_col]
features = imageStructToArray(spimg)
localFeatures.append(features)
if not localFeatures: # NOTE(phi-dbq): pep-8 recommended against testing 0 == len(array)
raise ValueError("Cannot extract any feature from dataset!")
X = np.stack(localFeatures, axis=0)
# Extract labels
y = None
if label_col is not None:
label_schema = image_df.schema[label_col]
label_dtype = label_schema.dataType
assert isinstance(label_dtype, spla.VectorUDT), \
"must encode labels in one-hot vector format, but got {}".format(label_dtype)
localLabels = []
for row in rows:
try:
_keras_label = row[label_col].toArray()
except ValueError:
raise ValueError("Cannot extract encoded label array")
localLabels.append(_keras_label)
if not localLabels:
raise ValueError(
"Failed to load any labels from dataset, but labels are required"
)
y = np.stack(localLabels, axis=0)
assert y.shape[0] == X.shape[0], \
"number of features {} != number of labels {}".format(X.shape[0], y.shape[0])
return X, y
def test_image_round_trip(self):
# Test round trip: array -> png -> sparkImg -> array
binarySchema = StructType([StructField("data", BinaryType(), False)])
df = self.session.createDataFrame([[bytearray(pngData)]], binarySchema)
# Convert to images
decImg = udf(imageIO._decodeImage, imageIO.imageSchema)
imageDF = df.select(decImg("data").alias("image"))
row = imageDF.first()
testArray = imageIO.imageStructToArray(row.image)
self.assertEqual(testArray.shape, array.shape)
self.assertEqual(testArray.dtype, array.dtype)
self.assertTrue(np.all(array == testArray))
def test_image_round_trip(self):
# Test round trip: array -> png -> sparkImg -> array
binarySchema = StructType([StructField("data", BinaryType(), False)])
df = self.session.createDataFrame([[bytearray(pngData)]], binarySchema)
# Convert to images
decImg = udf(imageIO._decodeImage, imageIO.imageSchema)
imageDF = df.select(decImg("data").alias("image"))
row = imageDF.first()
testArray = imageIO.imageStructToArray(row.image)
self.assertEqual(testArray.shape, array.shape)
self.assertEqual(testArray.dtype, array.dtype)
self.assertTrue(np.all(array == testArray))
def _getNumpyFeaturesAndLabels(self, dataset):
"""
We assume the training data fits in memory on a single server.
The input dataframe is converted to numerical image features and
broadcast to all the worker nodes.
"""
image_uri_col = self.getInputCol()
label_col = None
if self.isDefined(self.labelCol) and self.getLabelCol() != "":
label_col = self.getLabelCol()
tmp_image_col = self._loadedImageCol()
image_df = self.loadImagesInternal(dataset, image_uri_col).dropna(subset=[tmp_image_col])
# Extract features
localFeatures = []
rows = image_df.collect()
for row in rows:
spimg = row[tmp_image_col]
features = imageStructToArray(spimg)
localFeatures.append(features)
if not localFeatures: # NOTE(phi-dbq): pep-8 recommended against testing 0 == len(array)
raise ValueError("Cannot extract any feature from dataset!")
X = np.stack(localFeatures, axis=0)
# Extract labels
y = None
if label_col is not None:
label_schema = image_df.schema[label_col]
label_dtype = label_schema.dataType
assert isinstance(label_dtype, spla.VectorUDT), \
"must encode labels in one-hot vector format, but got {}".format(label_dtype)
localLabels = []
for row in rows:
try:
_keras_label = row[label_col].toArray()
except ValueError:
raise ValueError("Cannot extract encoded label array")
localLabels.append(_keras_label)
if not localLabels:
raise ValueError("Failed to load any labels from dataset, but labels are required")
y = np.stack(localLabels, axis=0)
assert y.shape[0] == X.shape[0], \
"number of features {} != number of labels {}".format(X.shape[0], y.shape[0])
return X, y
def _executeTensorflow(self, graph, input_tensor_name, output_tensor_name,
df, input_col="image"):
with tf.Session(graph=graph) as sess:
output_tensor = graph.get_tensor_by_name(output_tensor_name)
image_collected = df.collect()
values = {}
topK = {}
for img_row in image_collected:
image = np.expand_dims(imageStructToArray(img_row[input_col]), axis=0)
uri = img_row['image']['origin']
output = sess.run([output_tensor],
feed_dict={
graph.get_tensor_by_name(input_tensor_name): image
})
values[uri] = np.array(output[0])
topK[uri] = decode_predictions(values[uri], top=5)[0]
return values, topK
def _executeTensorflow(self, graph, input_tensor_name, output_tensor_name,
df, id_col="filePath", input_col="image"):
with tf.Session(graph=graph) as sess:
output_tensor = graph.get_tensor_by_name(output_tensor_name)
image_collected = df.collect()
values = {}
topK = {}
for img_row in image_collected:
image = np.expand_dims(imageStructToArray(img_row[input_col]), axis=0)
uri = img_row[id_col]
output = sess.run([output_tensor],
feed_dict={
graph.get_tensor_by_name(input_tensor_name): image
})
values[uri] = np.array(output[0])
topK[uri] = decode_predictions(values[uri], top=5)[0]
return values, topK
def test_loadImages(self):
input_col = "uri"
output_col = "preds"
model_path = image_utils.prepInceptionV3KerasModelFile("inceptionV3.h5")
transformer = KerasImageFileTransformer(
inputCol=input_col, outputCol=output_col, modelFile=model_path,
imageLoader=image_utils.loadAndPreprocessKerasInceptionV3, outputMode="vector")
uri_df = image_utils.getSampleImagePathsDF(self.sql, input_col)
image_df = transformer.loadImagesInternal(uri_df, input_col)
self.assertEqual(len(image_df.columns), 2)
img_col = transformer._loadedImageCol()
expected_shape = InceptionV3Constants.INPUT_SHAPE + (3,)
for row in image_df.collect():
arr = imageStructToArray(row[img_col])
self.assertEqual(arr.shape, expected_shape)
def test_resize(self):
self.assertRaises(ValueError, imageIO.createResizeImageUDF, [1, 2, 3])
make_smaller = imageIO.createResizeImageUDF([4, 5]).func
imgAsRow = imageIO.imageArrayToStruct(array)
smallerImg = make_smaller(imgAsRow)
self.assertEqual(smallerImg.height, 4)
self.assertEqual(smallerImg.width, 5)
# Compare to PIL resizing
imgAsPIL = PIL.Image.fromarray(obj=imageIO._reverseChannels(array)).resize((5, 4))
smallerAry = imageIO._reverseChannels(np.asarray(imgAsPIL))
np.testing.assert_array_equal(smallerAry, imageIO.imageStructToArray(smallerImg))
# Test that resize with the same size is a no-op
sameImage = imageIO.createResizeImageUDF((imgAsRow.height, imgAsRow.width)).func(imgAsRow)
self.assertEqual(imgAsRow, sameImage)
# Test that we have a valid image schema (all fields are in)
for n in ImageSchema.imageSchema['image'].dataType.names:
smallerImg[n]
def test_loadImages(self):
input_col = "uri"
output_col = "preds"
model_path = image_utils.prepInceptionV3KerasModelFile("inceptionV3.h5")
transformer = KerasImageFileTransformer(inputCol=input_col, outputCol=output_col,
modelFile=model_path,
imageLoader=image_utils.loadAndPreprocessKerasInceptionV3,
outputMode="vector")
uri_df = image_utils.getSampleImagePathsDF(self.sql, input_col)
image_df = transformer._loadImages(uri_df)
self.assertEqual(len(image_df.columns), 2)
img_col = transformer._loadedImageCol()
expected_shape = InceptionV3Constants.INPUT_SHAPE + (3,)
for row in image_df.collect():
arr = imageStructToArray(row[img_col])
self.assertEqual(arr.shape, expected_shape)
def test_resize(self):
self.assertRaises(ValueError, imageIO.createResizeImageUDF, [1, 2, 3])
make_smaller = imageIO.createResizeImageUDF([4, 5]).func
imgAsRow = imageIO.imageArrayToStruct(array)
smallerImg = make_smaller(imgAsRow)
self.assertEqual(smallerImg.height, 4)
self.assertEqual(smallerImg.width, 5)
# Compare to PIL resizing
imgAsPIL = PIL.Image.fromarray(
obj=imageIO._reverseChannels(array)).resize((5, 4))
smallerAry = imageIO._reverseChannels(np.asarray(imgAsPIL))
np.testing.assert_array_equal(smallerAry,
imageIO.imageStructToArray(smallerImg))
# Test that resize with the same size is a no-op
sameImage = imageIO.createResizeImageUDF(
(imgAsRow.height, imgAsRow.width)).func(imgAsRow)
self.assertEqual(imgAsRow, sameImage)
# Test that we have a valid image schema (all fields are in)
for n in ImageSchema.imageSchema['image'].dataType.names:
smallerImg[n]
def do_nothing(imgRow):
array = imageIO.imageStructToArray(imgRow)
return imageIO.imageArrayToStruct(array)
def do_nothing(imgRow):
imType = imageIO.imageType(imgRow)
array = imageIO.imageStructToArray(imgRow)
return imageIO.imageArrayToStruct(array, imType.sparkMode)
def do_nothing(imgRow):
array = imageIO.imageStructToArray(imgRow)
return imageIO.imageArrayToStruct(array)
def do_nothing(imgRow):
imType = imageIO.imageType(imgRow)
array = imageIO.imageStructToArray(imgRow)
return imageIO.imageArrayToStruct(array, imType.sparkMode)