def testParse(self):
random_array = np.random.random_sample([10,10,3]) * 255
random_array = random_array.astype(np.uint8)
img1 = Image.fromarray(random_array)
onnx_transforms = TRANSFORMS('onnxrt_qlinearops', 'preprocess')
onnx_parse = onnx_transforms['ToArray']()
img, _ = onnx_parse((img1, None))
self.assertTrue(isinstance(img, np.ndarray))
mxnet_transforms = TRANSFORMS('mxnet', 'preprocess')
mxnet_parse = mxnet_transforms['ToArray']()
img, _ = mxnet_parse((mx.nd.array(random_array), None))
self.assertTrue(isinstance(img, np.ndarray))
self.assertRaises(ValueError, mxnet_parse, ([1,2], None))
def testLabelShift(self):
transforms = TRANSFORMS('tensorflow', "postprocess")
transform = transforms['LabelShift'](label_shift=1)
rand_input = np.random.random_sample([600,600,3]).astype(np.float32)
sample = (rand_input, 1001)
label = transform(sample)[1]
self.assertEqual(label, 1000)
def setUpClass(cls):
cls.img = np.random.random_sample([10,10,3])*255
cls.tf_trans = TRANSFORMS('tensorflow', 'preprocess')
cls.pt_trans = TRANSFORMS('pytorch', 'preprocess')
cls.mx_trans = TRANSFORMS('mxnet', 'preprocess')
cls.ox_trans = TRANSFORMS('onnxrt_qlinearops', 'preprocess')
cls.mx_img = mx.nd.array(cls.img.astype(np.uint8))
cls.pt_img = Image.fromarray(cls.img.astype(np.uint8))
_ = TRANSFORMS('tensorflow', 'postprocess')
_ = TRANSFORMS('pytorch', 'postprocess')
_ = TRANSFORMS('mxnet', 'postprocess')
_ = TRANSFORMS('onnxrt_qlinearops' , 'postprocess')
_ = TRANSFORMS('onnxrt_integerops', 'postprocess')
def testResizeCropImagenetTransform(self):
transforms = TRANSFORMS('onnxrt_integerops', "preprocess")
transform = transforms['ResizeCropImagenet'](height=224, width=224)
sample = (self.img, 0)
result = transform(sample)
resized_input = result[0]
self.assertEqual(len(resized_input), 3)
self.assertEqual(len(resized_input[0]), 224)
self.assertEqual(len(resized_input[0][0]), 224)
def testQuantizedInput(self):
transforms = TRANSFORMS('tensorflow', "preprocess")
transform = transforms['QuantizedInput'](dtype='uint8', scale=100)
rand_input = np.random.random_sample([600,600,3]).astype(np.float32)
sample = (rand_input, 1001)
result = transform(sample)
quantized_input = result[0].eval(session=tf.compat.v1.Session())
self.assertLessEqual(quantized_input.max(), 255)
self.assertGreaterEqual(quantized_input.min(), 0)
def testParse(self):
from PIL import Image
random_array = np.random.random_sample([100, 100, 3]) * 255
random_array = random_array.astype(np.uint8)
img1 = Image.fromarray(random_array)
onnx_transforms = TRANSFORMS('onnxrt_qlinearops', 'preprocess')
onnx_parse = onnx_transforms['ImageTypeParse']()
onnx_compose = onnx_transforms['Compose']([onnx_parse])
onnx_result = onnx_compose((img1, None))
self.assertEqual(type(onnx_result[0]).__name__, 'ndarray')
def testResizeCropImagenetTransform(self):
transforms = TRANSFORMS('tensorflow', "preprocess")
transform = transforms['ResizeCropImagenet'](height=224, width=224)
rand_input = np.random.random_sample([600,600,3]).astype(np.float32)
sample = (rand_input, 0)
result = transform(sample)
resized_input = result[0].eval(session=tf.compat.v1.Session())
self.assertEqual(len(resized_input), 224)
self.assertEqual(len(resized_input[0]), 224)
self.assertEqual(len(resized_input[0][0]), 3)
def test_tensorflow_2(self):
image = np.ones([1, 256, 256, 1])
resize_kwargs = {"size":[224, 224]}
transforms = TRANSFORMS(framework="tensorflow", process="preprocess")
resize = transforms['Resize'](**resize_kwargs)
random_crop_kwargs = {"size": 128}
random_crop = transforms['RandomCrop'](**random_crop_kwargs)
transform_list = [resize, random_crop]
compose = transforms['Compose'](transform_list)
image_result = compose((image, None))
self.assertEqual(image_result[0].shape, (1, 128, 128, 1))
def test_pytorch_dummy(self):
datasets = DATASETS('pytorch')
transform = TRANSFORMS('pytorch', 'preprocess')['Resize'](**{
'size': 100
})
dataset = datasets['dummy'](shape=[(4, 256, 256, 3), (4, 1)], \
high=[10., 10.], low=[0., 0.], transform=transform)
data_loader = DATALOADERS['pytorch'](dataset)
iterator = iter(data_loader)
data, label = next(iterator)
self.assertEqual(data.shape, (1, 256, 256, 3))
# dynamic batching
data_loader.batch(batch_size=2, last_batch='rollover')
iterator = iter(data_loader)
data, label = next(iterator)
self.assertEqual(data.shape, (2, 256, 256, 3))
def test_onnxrt_qlinear_dummy(self):
datasets = DATASETS('onnxrt_qlinearops')
transform = TRANSFORMS('onnxrt_qlinearops',
'preprocess')['Resize'](**{
'size': 100
})
dataset = datasets['dummy'](shape=(4, 256, 256, 3),
transform=transform)
data_loader = DATALOADERS['onnxrt_qlinearops'](dataset)
iterator = iter(data_loader)
data = next(iterator)
self.assertEqual(data.shape, (1, 256, 256, 3))
# dynamic batching
data_loader.batch(batch_size=2, last_batch='rollover')
iterator = iter(data_loader)
data = next(iterator)
self.assertEqual(data.shape, (2, 256, 256, 3))
def testPyTorch(self):
transforms = TRANSFORMS('pytorch', 'preprocess')
align = transforms['AlignImageChannel'](**{'dim':1})
image, _ = align((TestAlignImageChannel.pt_img1, None))
self.assertEqual(image.mode, 'L')
align = transforms['AlignImageChannel'](**{'dim':1})
image, _ = align((TestAlignImageChannel.pt_img2, None))
self.assertEqual(image.mode, 'L')
align = transforms['AlignImageChannel'](**{'dim':3})
image, _ = align((TestAlignImageChannel.pt_img3, None))
self.assertEqual(image.mode, 'RGB')
with self.assertRaises(ValueError):
align = transforms['AlignImageChannel'](**{'dim':2})
with self.assertRaises(ValueError):
transforms['AlignImageChannel'](**{'dim':5})
def testONNX(self):
transforms = TRANSFORMS('onnxrt_qlinearops', 'preprocess')
align = transforms['AlignImageChannel'](**{'dim':1})
image, _ = align((TestAlignImageChannel.img1.astype(np.uint8), None))
self.assertEqual(image.shape[-1], 1)
align = transforms['AlignImageChannel'](**{'dim':1})
image, _ = align((TestAlignImageChannel.img2.astype(np.uint8), None))
self.assertEqual(image.shape[-1], 1)
align = transforms['AlignImageChannel'](**{'dim':3})
image, _ = align((TestAlignImageChannel.img3.astype(np.uint8), None))
self.assertEqual(image.shape[-1], 3)
align = transforms['AlignImageChannel'](**{'dim':2})
self.assertRaises(ValueError, align,
(TestAlignImageChannel.img1.astype(np.uint8), None))
with self.assertRaises(ValueError):
transforms['AlignImageChannel'](**{'dim':5})
def setUpClass(cls):
cls.img = np.ones([10,10,3])
cls.transforms = TRANSFORMS('tensorflow', 'preprocess')
def setUpClass(cls):
cls.img = np.random.random_sample([100,100,3]) * 255
cls.transforms = TRANSFORMS('onnxrt_qlinearops', 'preprocess')
def setUpClass(cls):
array = np.random.random_sample([100,100,3]) * 255
cls.img = mx.nd.array(array)
cls.transforms = TRANSFORMS('mxnet', 'preprocess')
def testLabelBalanceCOCORecord(self):
from PIL import Image
tf.compat.v1.disable_eager_execution()
random_array = np.random.random_sample([100, 100, 3]) * 255
random_array = random_array.astype(np.uint8)
im = Image.fromarray(random_array)
im.save('test.jpeg')
image = tf.compat.v1.gfile.FastGFile('test.jpeg', 'rb').read()
source_id = '000000397133.jpg'.encode('utf-8')
label = 'person'.encode('utf-8')
example1 = tf.train.Example(features=tf.train.Features(
feature={
'image/encoded':
tf.train.Feature(bytes_list=tf.train.BytesList(value=[image])),
'image/object/class/text':
tf.train.Feature(bytes_list=tf.train.BytesList(value=[label])),
'image/source_id':
tf.train.Feature(bytes_list=tf.train.BytesList(
value=[source_id])),
'image/object/bbox/xmin':
tf.train.Feature(float_list=tf.train.FloatList(value=[10])),
'image/object/bbox/ymin':
tf.train.Feature(float_list=tf.train.FloatList(value=[10])),
'image/object/bbox/xmax':
tf.train.Feature(float_list=tf.train.FloatList(value=[100])),
'image/object/bbox/ymax':
tf.train.Feature(float_list=tf.train.FloatList(value=[100])),
}))
example2 = tf.train.Example(features=tf.train.Features(
feature={
'image/encoded':
tf.train.Feature(bytes_list=tf.train.BytesList(value=[image])),
'image/object/class/text':
tf.train.Feature(bytes_list=tf.train.BytesList(value=[label])),
'image/source_id':
tf.train.Feature(bytes_list=tf.train.BytesList(
value=[source_id])),
'image/object/bbox/xmin':
tf.train.Feature(float_list=tf.train.FloatList(
value=[10, 20])),
'image/object/bbox/ymin':
tf.train.Feature(float_list=tf.train.FloatList(
value=[10, 20])),
'image/object/bbox/xmax':
tf.train.Feature(float_list=tf.train.FloatList(
value=[100, 200])),
'image/object/bbox/ymax':
tf.train.Feature(float_list=tf.train.FloatList(
value=[100, 200])),
}))
with tf.io.TFRecordWriter('test.record') as writer:
writer.write(example1.SerializeToString())
writer.write(example2.SerializeToString())
preprocesses = TRANSFORMS('tensorflow', 'preprocess')
preprocess = get_preprocess(preprocesses, {'ParseDecodeCoco': {}})
filters = FILTERS('tensorflow')
filter = filters['LabelBalanceCOCORecord'](2)
datasets = DATASETS('tensorflow')
dataset = datasets['COCORecord']('test.record', \
transform=preprocess, filter=filter)
dataloader = DATALOADERS['tensorflow'](dataset=dataset, batch_size=1)
for (inputs, labels) in dataloader:
self.assertEqual(inputs.shape, (1, 100, 100, 3))
self.assertEqual(labels[0].shape, (1, 2, 4))
dataset2 = create_dataset('tensorflow',
{'COCORecord': {
'root': 'test.record'
}}, {'ParseDecodeCoco': {}},
{'LabelBalance': {
'size': 2
}})
dataloader2 = DATALOADERS['tensorflow'](dataset=dataset2, batch_size=1)
for (inputs, labels) in dataloader2:
self.assertEqual(inputs.shape, (1, 100, 100, 3))
self.assertEqual(labels[0].shape, (1, 2, 4))
dataloader3 = create_dataloader('tensorflow', {'batch_size':1, 'dataset':{'COCORecord':{'root':'test.record'}},\
'filter':{'LabelBalance':{'size':2}}, 'transform':{'ParseDecodeCoco':{}}})
for (inputs, labels) in dataloader3:
self.assertEqual(inputs.shape, (1, 100, 100, 3))
self.assertEqual(labels[0].shape, (1, 2, 4))
os.remove('test.record')
os.remove('test.jpeg')
def setUpClass(cls):
cls.img = np.random.random_sample([10,10,3])*255
cls.mx_trans = TRANSFORMS('mxnet', 'preprocess')
cls.pt_trans = TRANSFORMS('pytorch', 'preprocess')
