def deserialize(self, stream, content_type):
"""Read a recordio-protobuf stream from a SageMaker Semantic Segmentation algorithm endpoint
Args:
stream (botocore.response.StreamingBody): A stream of bytes.
content_type (str): The MIME type of the data.
Returns:
array: numpy array of class probabilities per pixel
"""
try:
# Unpack the RecordIO wrapper first:
reccontent = HackyProtobufDeserializer.next_recordio_record(stream)
# Then load the protocol buffer:
rec = Record()
print("Parsing protobuf...")
protobuf = rec.ParseFromString(reccontent)
# Then read the two provided tensors `target` (predictions) and `shape`, squeeze out any batch
# dimensions (since we'll always be predicting on a single image) and shape target appropriately:
print("Fetching Tensors...")
values = list(rec.features["target"].float32_tensor.values)
shape = list(rec.features["shape"].int32_tensor.values)
print("reshaping arrays...")
shape = squeeze(shape)
mask = reshape(array(values), shape)
return squeeze(mask, axis=0)
finally:
stream.close()
def write_numpy_to_dense_tensor(file, array, labels=None):
"""Writes a numpy array to a dense tensor
Args:
file:
array:
labels:
"""
# Validate shape of array and labels, resolve array and label types
if not len(array.shape) == 2:
raise ValueError("Array must be a Matrix")
if labels is not None:
if not len(labels.shape) == 1:
raise ValueError("Labels must be a Vector")
if labels.shape[0] not in array.shape:
raise ValueError(
"Label shape {} not compatible with array shape {}".format(
labels.shape, array.shape))
resolved_label_type = _resolve_type(labels.dtype)
resolved_type = _resolve_type(array.dtype)
# Write each vector in array into a Record in the file object
record = Record()
for index, vector in enumerate(array):
record.Clear()
_write_feature_tensor(resolved_type, record, vector)
if labels is not None:
_write_label_tensor(resolved_label_type, record, labels[index])
_write_recordio(file, record.SerializeToString())
def test_serializer_accepts_one_dimensional_array():
s = RecordSerializer()
array_data = [1.0, 2.0, 3.0]
buf = s.serialize(np.array(array_data))
record_data = next(read_recordio(buf))
record = Record()
record.ParseFromString(record_data)
assert record.features["values"].float64_tensor.values == array_data
def test_serializer():
s = RecordSerializer()
array_data = [[1.0, 2.0, 3.0], [10.0, 20.0, 30.0]]
buf = s.serialize(np.array(array_data))
for record_data, expected in zip(read_recordio(buf), array_data):
record = Record()
record.ParseFromString(record_data)
assert record.features["values"].float64_tensor.values == expected
def read_records(file):
"""Eagerly read a collection of amazon Record protobuf objects from file."""
records = []
for record_data in read_recordio(file):
record = Record()
record.ParseFromString(record_data)
records.append(record)
return records
def read_pipe(pipe):
with open(pipe,'rb') as f:
for rec in read_recordio(f):
print("read record")
print(rec)
record = Record()
record.ParseFromString(rec)
print("record parsed")
print(record)
def test_int_write_numpy_to_dense_tensor():
array_data = [[1, 2, 3], [10, 20, 3]]
array = np.array(array_data)
with tempfile.TemporaryFile() as f:
write_numpy_to_dense_tensor(f, array)
f.seek(0)
for record_data, expected in zip(read_recordio(f), array_data):
record = Record()
record.ParseFromString(record_data)
assert record.features["values"].int32_tensor.values == expected
def test_float32_write_numpy_to_dense_tensor():
array_data = [[1.0, 2.0, 3.0], [10.0, 20.0, 30.0]]
array = np.array(array_data).astype(np.dtype("float32"))
with tempfile.TemporaryFile() as f:
write_numpy_to_dense_tensor(f, array)
f.seek(0)
for record_data, expected in zip(read_recordio(f), array_data):
record = Record()
record.ParseFromString(record_data)
assert record.features["values"].float32_tensor.values == expected
def write_spmatrix_to_sparse_tensor(file, array, labels=None):
"""Writes a scipy sparse matrix to a sparse tensor
Args:
file:
array:
labels:
"""
try:
import scipy
except ImportError as e:
logging.warning(
"scipy failed to import. Sparse matrix functions will be impaired or broken."
)
# Any subsequent attempt to use scipy will raise the ImportError
scipy = DeferredError(e)
if not scipy.sparse.issparse(array):
raise TypeError("Array must be sparse")
# Validate shape of array and labels, resolve array and label types
if not len(array.shape) == 2:
raise ValueError("Array must be a Matrix")
if labels is not None:
if not len(labels.shape) == 1:
raise ValueError("Labels must be a Vector")
if labels.shape[0] not in array.shape:
raise ValueError(
"Label shape {} not compatible with array shape {}".format(
labels.shape, array.shape))
resolved_label_type = _resolve_type(labels.dtype)
resolved_type = _resolve_type(array.dtype)
csr_array = array.tocsr()
n_rows, n_cols = csr_array.shape
record = Record()
for row_idx in range(n_rows):
record.Clear()
row = csr_array.getrow(row_idx)
# Write values
_write_feature_tensor(resolved_type, record, row.data)
# Write keys
_write_keys_tensor(resolved_type, record,
row.indices.astype(np.uint64))
# Write labels
if labels is not None:
_write_label_tensor(resolved_label_type, record, labels[row_idx])
# Write shape
_write_shape(resolved_type, record, n_cols)
_write_recordio(file, record.SerializeToString())
def test_float32_label():
array_data = [[1, 2, 3], [10, 20, 3]]
array = np.array(array_data)
label_data = np.array([99, 98, 97]).astype(np.dtype('float32'))
with tempfile.TemporaryFile() as f:
write_numpy_to_dense_tensor(f, array, label_data)
f.seek(0)
for record_data, expected, label in zip(_read_recordio(f), array_data, label_data):
record = Record()
record.ParseFromString(record_data)
assert record.features["values"].int32_tensor.values == expected
assert record.label["values"].float32_tensor.values == [label]
def test_dense_int_write_spmatrix_to_sparse_tensor():
array_data = [[1.0, 2.0, 3.0], [10.0, 20.0, 30.0]]
keys_data = [[0, 1, 2], [0, 1, 2]]
array = coo_matrix(np.array(array_data).astype(np.dtype('int')))
with tempfile.TemporaryFile() as f:
write_spmatrix_to_sparse_tensor(f, array)
f.seek(0)
for record_data, expected_data, expected_keys in zip(read_recordio(f), array_data, keys_data):
record = Record()
record.ParseFromString(record_data)
assert record.features["values"].int32_tensor.values == expected_data
assert record.features["values"].int32_tensor.keys == expected_keys
assert record.features["values"].int32_tensor.shape == [len(expected_data)]
def test_dense_float64_spmatrix_to_sparse_label():
array_data = [[1, 2, 3], [10, 20, 3]]
keys_data = [[0, 1, 2], [0, 1, 2]]
array = coo_matrix(np.array(array_data).astype("float64"))
label_data = np.array([99, 98, 97])
with tempfile.TemporaryFile() as f:
write_spmatrix_to_sparse_tensor(f, array, label_data)
f.seek(0)
for record_data, expected_data, expected_keys, label in zip(
read_recordio(f), array_data, keys_data, label_data
):
record = Record()
record.ParseFromString(record_data)
assert record.features["values"].float64_tensor.values == expected_data
assert record.features["values"].float64_tensor.keys == expected_keys
assert record.label["values"].int32_tensor.values == [label]
assert record.features["values"].float64_tensor.shape == [len(expected_data)]
def write_spmatrix_to_sparse_tensor(file, array, labels=None):
"""Writes a scipy sparse matrix to a sparse tensor
Args:
file:
array:
labels:
"""
if not issparse(array):
raise TypeError("Array must be sparse")
# Validate shape of array and labels, resolve array and label types
if not len(array.shape) == 2:
raise ValueError("Array must be a Matrix")
if labels is not None:
if not len(labels.shape) == 1:
raise ValueError("Labels must be a Vector")
if labels.shape[0] not in array.shape:
raise ValueError(
"Label shape {} not compatible with array shape {}".format(
labels.shape, array.shape))
resolved_label_type = _resolve_type(labels.dtype)
resolved_type = _resolve_type(array.dtype)
csr_array = array.tocsr()
n_rows, n_cols = csr_array.shape
record = Record()
for row_idx in range(n_rows):
record.Clear()
row = csr_array.getrow(row_idx)
# Write values
_write_feature_tensor(resolved_type, record, row.data)
# Write keys
_write_keys_tensor(resolved_type, record,
row.indices.astype(np.uint64))
# Write labels
if labels is not None:
_write_label_tensor(resolved_label_type, record, labels[row_idx])
# Write shape
_write_shape(resolved_type, record, n_cols)
_write_recordio(file, record.SerializeToString())
def infer(self, filename, output_folder, endpoint, show_image):
"""
:param filename:
:param output_folder:
:param endpoint:
:param show_image:
:return:
"""
output_filename = os.path.splitext(filename)[0]
output_filename = os.path.basename(output_filename)
output_filename = os.path.join(output_folder, output_filename + '_inference.png')
runtime = boto3.Session().client('sagemaker-runtime')
image = pillow.Image.open(filename)
image.thumbnail([800, 600], pillow.Image.ANTIALIAS)
image.save(filename, "JPEG")
with open(filename, 'rb') as f:
payload = f.read()
payload = bytearray(payload)
response = runtime.invoke_endpoint(EndpointName=endpoint, ContentType='application/x-image', Body=payload)
results_file = 'results.rec'
with open(results_file, 'wb') as f:
f.write(response['Body'].read())
rec = Record()
recordio = mx.recordio.MXRecordIO(results_file, 'r')
protobuf = rec.ParseFromString(recordio.read())
values = list(rec.features["target"].float32_tensor.values)
shape = list(rec.features["shape"].int32_tensor.values)
shape = np.squeeze(shape)
mask = np.reshape(np.array(values), shape)
mask = np.squeeze(mask, axis=0)
pred_map = np.argmax(mask, axis=0)
if show_image is True:
utils_obj = utils.Utils()
utils_obj.show_image(pred_map)
plt.imshow(pred_map, vmin=0, vmax=settings.HYPER['num_classes'] - 1, cmap='jet')
plt.savefig(output_filename)
def test_sparse_int_write_spmatrix_to_sparse_tensor():
n = 4
array_data = [[1.0, 2.0], [10.0, 30.0], [100.0, 200.0, 300.0, 400.0], [1000.0, 2000.0, 3000.0]]
keys_data = [[0, 1], [1, 2], [0, 1, 2, 3], [0, 2, 3]]
flatten_data = list(itertools.chain.from_iterable(array_data))
y_indices = list(itertools.chain.from_iterable(keys_data))
x_indices = [[i] * len(keys_data[i]) for i in range(len(keys_data))]
x_indices = list(itertools.chain.from_iterable(x_indices))
array = coo_matrix((flatten_data, (x_indices, y_indices)), dtype='int')
with tempfile.TemporaryFile() as f:
write_spmatrix_to_sparse_tensor(f, array)
f.seek(0)
for record_data, expected_data, expected_keys in zip(read_recordio(f), array_data, keys_data):
record = Record()
record.ParseFromString(record_data)
assert record.features["values"].int32_tensor.values == expected_data
assert record.features["values"].int32_tensor.keys == expected_keys
assert record.features["values"].int32_tensor.shape == [n]
def write_record(file, features=None, metadata=None):
record = Record()
record.Clear()
"""
print(record)
print(dir(record))
print(record.features)
print("data: {}, {}".format(type(data), len(data)))
print(dir(record.features["values"].bytes.value))
"""
if features:
for k, v in features.items():
write_feature_tensor(record, v, k)
if metadata:
if not isinstance(metadata, str):
metadata = json.dumps(metadata)
record.metadata = metadata
# print(record.metadata)
# print(dir(record.metadata))
# record.metadata.val
_write_recordio(file, record.SerializeToString())
def transformation():
"""Do an inference on a single batch of data. In this server, we take data as JSON, convert
it to a sparse array for internal use and then convert the predictions back to json.
Input format is:
'{"instances": [{"keys": ["User","1","2"], "values": ["a","b","c"]}, {"keys": ["User","5","6"], "values": ["d","e","f"]}]}'
"""
# Convert from json to numpy
te_row_ind = []
te_col_ind = []
te_data = []
te_idx = 0
headers = ScoringService.get_headers()
if flask.request.content_type == 'application/json':
print("Working with JSON input")
s = flask.request.data.decode('utf-8')
inputs = json.loads(s)
for instance in inputs['instances']:
# The column index has to be found from the headers
for col_idx in range(0, len(instance['keys'])):
key = instance['keys'][col_idx]
val = instance['values'][col_idx]
item_to_find = "{0}_{1}".format(key, val)
try:
te_col_ind.append(headers.index(item_to_find))
te_data.append(1.0)
te_row_ind.append(te_idx)
except Exception as e:
te_col_ind.append(1)
te_data.append(0.0)
te_row_ind.append(te_idx)
print("Couldn't find header for {0}".format(item_to_find))
te_idx = te_idx + 1
elif flask.request.content_type == 'application/x-recordio-protobuf':
print("Working with Protobuf input")
#print("{0}".format(flask.request.stream))
#s = flask.request.data.decode('latin-1')
#print("Data: {}".format(s))
test_records = smac.read_records(StringIO.StringIO(flask.request.data))
num_test_samples = len(test_records)
for test_record in test_records:
te_row_ind.extend([te_idx] * len(test_record.features['values'].float32_tensor.values))
te_col_ind.extend(test_record.features['values'].float32_tensor.keys)
te_data.extend(test_record.features['values'].float32_tensor.values)
te_idx = te_idx + 1
else:
return flask.Response(response='This predictor only supports JSON or Protobuf data', status=415, mimetype='text/plain')
X_te_sparse = sp.csr_matrix( (np.array(te_data),(np.array(te_row_ind),np.array(te_col_ind))), shape=(te_idx,ScoringService.get_num_features()) )
print('Invoked with {} records'.format(X_te_sparse.shape))
# Do the prediction
predictions = ScoringService.predict(X_te_sparse)
# Convert from array back to json
result = None
if flask.request.content_type == 'application/json':
js = {'predictions': []}
for pred_value in predictions:
js['predictions'].append({'score': str(pred_value)})
result = json.dumps(js)
else:
# convert to protobuf
buf = io.BytesIO()
record = Record()
for pred_value in predictions:
record.Clear()
#smac._write_label_tensor('Float32', record, pred_value)
record.label["score"].float32_tensor.values.extend([pred_value])
smac._write_recordio(buf, record.SerializeToString())
buf.seek(0)
result = buf.getvalue()
return flask.Response(response=result, status=200, mimetype=flask.request.content_type)
