def _run_test(self,
flip_horizontal,
flip_vertical,
expected_output=None,
mock_random=None):
np.random.seed(1337)
num_samples = 2
orig_height = 5
orig_width = 8
channels = 3
if mock_random is None:
mock_random = [1 for _ in range(num_samples)]
mock_random = np.reshape(mock_random, [2, 1, 1, 1])
inp = np.random.random(
(num_samples, orig_height, orig_width, channels))
if expected_output is None:
expected_output = inp
if flip_horizontal:
expected_output = np.flip(expected_output, axis=1)
if flip_vertical:
expected_output = np.flip(expected_output, axis=2)
with test.mock.patch.object(random_ops,
'random_uniform',
return_value=mock_random):
with tf_test_util.use_gpu():
layer = image_preprocessing.RandomFlip(flip_horizontal,
flip_vertical)
actual_output = layer(inp, training=1)
self.assertAllClose(expected_output, actual_output)
def test_distribution(self, distribution):
if "CentralStorage" in type(distribution).__name__:
self.skipTest("Does not work with CentralStorageStrategy yet.")
# TODO(b/159738418): large image input causes OOM in ubuntu multi gpu.
np_images = np.random.random((32, 32, 32, 3)).astype(np.float32)
image_dataset = dataset_ops.Dataset.from_tensor_slices(
np_images).batch(16, drop_remainder=True)
with distribution.scope():
input_data = keras.Input(shape=(32, 32, 3), dtype=dtypes.float32)
image_preprocessor = keras.Sequential([
image_preprocessing.Resizing(height=256, width=256),
image_preprocessing.RandomCrop(height=224, width=224),
image_preprocessing.RandomTranslation(.1, .1),
image_preprocessing.RandomRotation(.2),
image_preprocessing.RandomFlip(),
image_preprocessing.RandomZoom(.2, .2)
])
preprocessed_image = image_preprocessor(input_data)
flatten_layer = keras.layers.Flatten(data_format="channels_last")
output = flatten_layer(preprocessed_image)
cls_layer = keras.layers.Dense(units=1, activation="sigmoid")
output = cls_layer(output)
model = keras.Model(inputs=input_data, outputs=output)
model.compile(loss="binary_crossentropy")
_ = model.predict(image_dataset)
def test_random_flip_inference(self):
with CustomObjectScope({'RandomFlip': image_preprocessing.RandomFlip}):
input_images = np.random.random((2, 5, 8, 3)).astype(np.float32)
expected_output = input_images
with tf_test_util.use_gpu():
layer = image_preprocessing.RandomFlip()
actual_output = layer(input_images, training=0)
self.assertAllClose(expected_output, actual_output)
def test_random_flip_default(self):
with CustomObjectScope({'RandomFlip': image_preprocessing.RandomFlip}):
input_images = np.random.random((2, 5, 8, 3)).astype(np.float32)
expected_output = np.flip(np.flip(input_images, axis=1), axis=2)
mock_random = [1, 1]
mock_random = np.reshape(mock_random, [2, 1, 1, 1])
with test.mock.patch.object(
random_ops, 'random_uniform', return_value=mock_random):
with self.cached_session(use_gpu=True):
layer = image_preprocessing.RandomFlip()
actual_output = layer(input_images, training=1)
self.assertAllClose(expected_output, actual_output)
def bm_layer_implementation(self, batch_size):
with ops.device_v2("/gpu:0"):
img = keras.Input(shape=(256, 256, 3), dtype=dtypes.float32)
preprocessor = keras.Sequential([
image_preprocessing.Resizing(224, 224),
image_preprocessing.RandomCrop(height=224, width=224),
image_preprocessing.RandomRotation(factor=(.2, .4)),
image_preprocessing.RandomFlip(mode="horizontal"),
image_preprocessing.RandomZoom(.2, .2)
])
_ = preprocessor(img)
num_repeats = 5
starts = []
ends = []
for _ in range(num_repeats):
ds = dataset_ops.Dataset.from_tensor_slices(
np.random.random((batch_size, 256, 256, 3)))
ds = ds.shuffle(batch_size * 100)
ds = ds.batch(batch_size)
ds = ds.prefetch(batch_size)
starts.append(time.time())
count = 0
# Benchmarked code begins here.
for i in ds:
_ = preprocessor(i)
count += 1
# Benchmarked code ends here.
ends.append(time.time())
avg_time = np.mean(np.array(ends) - np.array(starts)) / count
name = "image_preprocessing|batch_%s" % batch_size
baseline = self.run_dataset_implementation(batch_size)
extras = {
"dataset implementation baseline": baseline,
"delta seconds": (baseline - avg_time),
"delta percent": ((baseline - avg_time) / baseline) * 100
}
self.report_benchmark(iters=num_repeats,
wall_time=avg_time,
extras=extras,
name=name)
def test_distribution(self, distribution):
np_images = np.random.random((1000, 32, 32, 3)).astype(np.float32)
image_dataset = dataset_ops.Dataset.from_tensor_slices(
np_images).batch(32, drop_remainder=True)
with distribution.scope():
input_data = keras.Input(shape=(32, 32, 3), dtype=dtypes.float32)
image_preprocessor = keras.Sequential([
image_preprocessing.Resizing(height=256, width=256),
image_preprocessing.RandomCrop(height=224, width=224),
image_preprocessing.RandomTranslation(.1, .1),
image_preprocessing.RandomRotation(.2),
image_preprocessing.RandomFlip(),
image_preprocessing.RandomZoom(.2, .2)
])
preprocessed_image = image_preprocessor(input_data)
flatten_layer = keras.layers.Flatten(data_format="channels_last")
output = flatten_layer(preprocessed_image)
cls_layer = keras.layers.Dense(units=1, activation="sigmoid")
output = cls_layer(output)
model = keras.Model(inputs=input_data, outputs=preprocessed_image)
model.compile(loss="binary_crossentropy")
_ = model.predict(image_dataset)
def test_data_augment():
main_path = '../-test/download'
ds_train, ds_val = create_data_set(main_path)
# rt = image_preprocessing.RandomTranslation(height_factor=0.1, width_factor=0.1)
rt = image_preprocessing.RandomFlip()
# rt = image_preprocessing.RandomRotation(factor=0.1)
# ds_train.map(rt)
import matplotlib.pyplot as plt
plt.figure(figsize=(10, 10))
for images, labels in ds_train.take(1):
images = rt(images)
n = tf.shape(images)[0]
n = 9 if n > 9 else n
for i in range(n):
plt.subplot(3, 3, i + 1)
plt.imshow(images[i].numpy().astype("uint8"))
# plt.title(int(labels[i]))
plt.axis("off")
plt.show()
def test_config_with_custom_name(self): layer = image_preprocessing.RandomFlip(name='image_preproc') config = layer.get_config() layer_1 = image_preprocessing.RandomFlip.from_config(config) self.assertEqual(layer_1.name, layer.name)
import tensorflow as tf # 2.4
from tensorflow.keras import layers
from tensorflow.keras.models import Sequential
from tensorflow.python.keras.layers.preprocessing import image_preprocessing
from tensorflow.keras.applications import EfficientNetB1, Xception
# 数据增强
image_augmentation = Sequential(
[
image_preprocessing.RandomRotation(factor=0.1), # 随机旋转
image_preprocessing.RandomTranslation(height_factor=0.1,
width_factor=0.1), # 随机平移
image_preprocessing.RandomFlip(), # 随机翻转
image_preprocessing.RandomContrast(factor=0.1), # 随机改变对比度
image_preprocessing.RandomZoom(height_factor=0.1,
width_factor=0.1), # 随机缩放
# image_preprocessing.RandomHeight(factor=0.1), # 随机改变高度
# image_preprocessing.RandomWidth(factor=0.1), # 随机改变宽度
# image_preprocessing.RandomCrop(height, width), # 随机裁剪
# image_preprocessing.CenterCrop(height, width), # 中心裁剪
],
name="img_augmentation",
)
def get_time_suffix(suffix_len=15):
"""获取一个时间后缀"""
import time
