def test_object_detection_data_shape(self, config):
"""Check the tensor shape of classification dataset"""
db = DBLoader()
db.load_built_in(config.name)
train_example = db.train.get_sample()
test_example = db.train.get_sample()
img_info = config.info["image"]
expected_img_shape = \
img_info["height"], img_info["width"], img_info["channel"]
box_info = config.info["boxes"]
nClass = box_info["classes"]
nBox = max(
box_info["train"]["max boxes of single data"],
box_info["test"]["max boxes of single data"],
)
expected_box_shape = nBox, 4 + nClass
for example in [train_example, test_example]:
image = example["image"].numpy()
boxes = example["boxes"].numpy()
for img_shp, expected in zip(image.shape, expected_img_shape):
assert (expected is None) or img_shp == expected
assert boxes.shape == expected_box_shape
def _check_acc_on_imagenet():
loader = DBLoader()
loader.load_built_in("imagenet", parser=ReshapeImageNet())
imagenet = loader.test.to_tfdataset(1, 1)
resnet50 = get_resnet()
resnet50.compile(metrics="acc")
resnet50.evaluate(x=imagenet)
def test_examine_image(self, tmp_path):
imagenet = DBLoader()
imagenet.load_built_in("imagenet", parser=RemapImagenet())
dataset = imagenet.test.to_tfdataset(batch=1, epoch=1)
dataset = dataset.take(20)
for idx, data in enumerate(dataset):
image, label = data
img = image[0, ..., ::-1].numpy()
cv2.imwrite(str(tmp_path.joinpath("{:02d}.bmp".format(idx))), img)
def load_simpson(tmpdir):
parser = SimpsonFMT()
loader = DBLoader()
loader.load(str(tmpdir), parser=parser)
train = loader.train.to_tfdataset(epoch=1, batch=1)
for example in train:
print(example)
break
def load_mvtectad():
parser = MVTectFMT()
loader = DBLoader()
loader.load(DST_DIR, parser=parser)
train = loader.train.to_tfdataset(1, 1)
for example in train:
print(example)
break
def test_loadable(self, config):
"""Can load all built-in datasets"""
if config.name == "imagenet":
pytest.skip(msg="count ImageNet takes too much time")
db = DBLoader()
assert db.info == ""
assert db.train is None
assert db.test is None
db.load_built_in(config.name)
assert db.train.count == config.train["count"]
assert db.test.count == config.test["count"]
assert db.info != ""
def test_inference_speed(self, model):
if model == "resnet":
model = resnet = tf.keras.applications.ResNet50V2(include_top=True,
input_shape=(256,
256,
3),
weights=None,
classes=1000)
elif model == "resnext":
model = ResNeXt50(input_shape=(256, 256, 3),
include_fc=True,
n_fc=1000,
load_pretrained=True)
print(model.count_params())
imagenet = DBLoader()
imagenet.load_built_in("imagenet", parser=RemapImagenet())
dataset = imagenet.test.to_tfdataset(batch=1, epoch=1)
dataset = dataset.take(1000)
start = time.time()
model.predict(dataset)
print(time.time() - start)
def test_classification_data_shape(self, config):
"""Check the tensor shape of classification dataset"""
db = DBLoader()
db.load_built_in(config.name)
train_example = db.train.get_sample()
test_example = db.train.get_sample()
info = config.info["image"]
if config.name == "imagenet":
expected_shape = (256, 256, 3)
else:
expected_shape = info["height"], info["width"], info["channel"]
for example in [train_example, test_example]:
image = example["image"].numpy()
label = example["label"].numpy()
assert image.shape == expected_shape
if config.name == "imagenet":
assert label.shape == (1000, )
else:
assert label.shape == ()
def manually_examine_seg(
self,
nImage: int,
out_dir: pathlib.Path,
resize_shape=None,
):
parser = CoCoSeg(one_hot_class=False, resize_shape=resize_shape)
db = DBLoader()
db.load_built_in("coco-seg", parser=parser)
ds = db.train.to_tfdataset(epoch=1, batch=1)
for cnt, example in enumerate(ds):
image = example["image"].numpy()[0, ...].astype(np.uint8)
image = image[..., ::-1]
image = np.array(image)
mask = example["segment_mask"].numpy()[0, ..., 0]
img = self.get_masked_image(image, mask)
img.save(str(out_dir.joinpath("train_{}.bmp".format(cnt))))
if cnt == nImage:
break
ds = db.test.to_tfdataset(epoch=1, batch=1)
for cnt, example in enumerate(ds):
image = example["image"].numpy()[0, ...].astype(np.uint8)
image = image[..., ::-1]
image = np.array(image)
mask = example["segment_mask"].numpy()[0, ..., 0]
img = self.get_masked_image(image, mask)
img.save(str(out_dir.joinpath("test_{}.bmp".format(cnt))))
if cnt == nImage:
break
def manually_examine_bbox(
self,
nImage: int,
out_dir: pathlib.Path,
resize_shape=None,
):
parser = CoCoObjDet(one_hot_class=False, resize_shape=resize_shape)
db = DBLoader()
db.load_built_in("coco-objdet", parser=parser)
ds = db.train.to_tfdataset(epoch=1, batch=1)
for cnt, example in enumerate(ds):
image = example["image"].numpy()[0, ...].astype(np.uint8)
image = image[..., ::-1]
image = np.array(image)
boxes = example["boxes"].numpy()[0, ...]
self.draw_boxes(image, boxes)
cv2.imwrite(str(out_dir.joinpath("train_{}.bmp".format(cnt))),
image)
if cnt == nImage:
break
ds = db.test.to_tfdataset(epoch=1, batch=1)
for cnt, example in enumerate(ds):
image = example["image"].numpy()[0, ...].astype(np.uint8)
image = image[..., ::-1]
image = np.array(image)
boxes = example["boxes"].numpy()[0, ...]
self.draw_boxes(image, boxes)
cv2.imwrite(str(out_dir.joinpath("test_{}.bmp".format(cnt))),
image)
if cnt == nImage:
break
image, (224, 224),
preserve_aspect_ratio=False,
antialias=False
)
label = example.pop("label")
return image, label
if __name__ == "__main__":
model = ResNeXt50(
input_shape=(224, 224, 3),
include_fc=True, n_fc=1000,
load_pretrained=False
)
imagenet = DBLoader()
imagenet.load_built_in("imagenet", parser=None)
optimizer_params = {
"learning_rate": 0.0025, "momentum": 0.9
}
hyper = {
"SGD": str(optimizer_params),
"lr_decay": 0.1,
"loss": keras.losses.CategoricalCrossentropy(from_logits=True),
"batch": 32,
}
logging.info("hyper info: {}".format(hyper))
trainer = KerasBaseTrainer(
model=model,