def __getitem__(self, index):
image = util.load_image(self.file_names[index])
label = util.load_label(self.file_names[index])
image, label = util.random_crop(image, label)
image, label = util.random_augmentation(image, label)
one_hot = util.one_hot(label, self.palette)
return np.float32(image) / 255.0, np.float32(one_hot)
def main():
_, palette = util.get_label_info(
os.path.join(config.data_dir, "class_dict.csv"))
model = nn.build_model(classes=len(palette))
model(tf.zeros((1, config.height, config.width, 3)))
file_names = [
file_name[:-4] for file_name in os.listdir(
os.path.join(config.data_dir, config.image_dir))
]
for file_name in tqdm.tqdm(file_names):
image = util.load_image(file_name)
label = util.load_label(file_name)
image, label = util.random_crop(image, label)
image = np.expand_dims(image, 0).astype('float32')
output = model.predict(image / 255.0)
output = np.array(output[0, :, :, :])
output = np.argmax(output, axis=-1)
output = util.colour_code_segmentation(output, palette)
output = np.uint8(output)
util.save_images([output, label],
os.path.join('results', f'{file_name}.jpg'),
titles=['Pred', 'Label'])
def __getitem__(self, index):
image = util.load_image(self.file_names[index])
boxes, label = util.load_label(self.file_names[index])
image, boxes = util.resize(image, boxes)
image, boxes = util.random_flip(image, boxes)
image = image[:, :, ::-1].astype(numpy.float32)
image = image / 255.0
y_true_1, y_true_2, y_true_3 = util.process_box(boxes, label)
return image, y_true_1, y_true_2, y_true_3
def get_boxes():
boxes = []
file_names = [
file_name[:-4] for file_name in os.listdir(
os.path.join(config.data_dir, config.label_dir))
]
for file_name in file_names:
for box in util.load_label(file_name)[0]:
boxes.append([box[2] - box[0], box[3] - box[1]])
return numpy.array(boxes)
def __getitem__(self, index):
file_name = self.file_names[index]
image = util.load_image(file_name)
label = util.load_label(file_name)
image, label = util.random_crop(image, label)
image, label = util.random_augmentation(image, label)
image = torch.from_numpy(image.transpose(2, 0, 1))
one_hot = util.one_hot_it(label, self.palette)
one_hot = torch.from_numpy(one_hot)
return image, one_hot
def __getitem__(self, index):
image = load_image(self.file_names[index])
boxes, label = load_label(self.file_names[index])
boxes = np.concatenate(
(boxes,
np.full(
shape=(boxes.shape[0], 1), fill_value=1., dtype=np.float32)),
axis=-1)
image, boxes = resize(image, boxes)
image = image.astype(np.float32)
image = image / 255.0
y_true_1, y_true_2, y_true_3 = process_box(boxes, label)
return image, y_true_1, y_true_2, y_true_3
def get_boxes():
boxes = []
f_names = [
f_name[:-4]
for f_name in os.listdir(join(config.data_dir, config.label_dir))
]
for f_name in f_names:
annotations = load_label(f_name)
for annotation in annotations[0]:
x_min = annotation[0]
y_min = annotation[1]
x_max = annotation[2]
y_max = annotation[3]
boxes.append([x_max - x_min, y_max - y_min])
return np.array(boxes)
def write_tf_record(queue, sentinel):
def byte_feature(value):
if not isinstance(value, bytes):
if not isinstance(value, list):
value = value.encode('utf-8')
else:
value = [val.encode('utf-8') for val in value]
if not isinstance(value, list):
value = [value]
return tf.train.Feature(bytes_list=tf.train.BytesList(value=value))
while True:
file_name = queue.get()
if file_name == sentinel:
break
in_image = util.load_image(file_name)[:, :, ::-1]
boxes, label = util.load_label(file_name)
in_image, boxes = util.resize(in_image, boxes)
y_true_1, y_true_2, y_true_3 = util.process_box(boxes, label)
in_image = in_image.astype('float32')
y_true_1 = y_true_1.astype('float32')
y_true_2 = y_true_2.astype('float32')
y_true_3 = y_true_3.astype('float32')
in_image = in_image.tobytes()
y_true_1 = y_true_1.tobytes()
y_true_2 = y_true_2.tobytes()
y_true_3 = y_true_3.tobytes()
features = tf.train.Features(
feature={
'in_image': byte_feature(in_image),
'y_true_1': byte_feature(y_true_1),
'y_true_2': byte_feature(y_true_2),
'y_true_3': byte_feature(y_true_3)
})
tf_example = tf.train.Example(features=features)
opt = tf.io.TFRecordOptions('GZIP')
with tf.io.TFRecordWriter(
os.path.join(config.data_dir, 'TF', file_name + ".tf"),
opt) as writer:
writer.write(tf_example.SerializeToString())
def get_boxes():
boxes = []
f_names = [
f_name[:-4]
for f_name in os.listdir(join(config.base_dir, config.label_dir))
]
for f_name in f_names:
annotations = load_label(f_name)
for annotation in annotations:
x_min = annotation[0]
y_min = annotation[1]
x_max = annotation[2]
y_max = annotation[3]
width = x_max - x_min
height = y_max - y_min
boxes.append([width, height])
return np.array(boxes)
def __getitem__(self, index):
image = load_image(self.f_names[index])
boxes, label = load_label(self.f_names[index])
boxes = np.concatenate(
(boxes,
np.full(
shape=(boxes.shape[0], 1), fill_value=1., dtype=np.float32)),
axis=-1)
image, boxes = random_horizontal_flip(image, boxes)
image, boxes = random_crop(image, boxes)
image, boxes = random_translate(image, boxes)
image, boxes = resize(image, boxes)
image = image.astype(np.float32)
image /= 255.
mean = [0.485, 0.456, 0.406]
std = [0.229, 0.224, 0.225]
image -= mean
image /= std
y_true_1, y_true_2, y_true_3 = process_box(boxes, label)
return image, y_true_1, y_true_2, y_true_3
def evaluate():
model = torch.load(os.path.join('weights', 'model.pt'),
device)['model'].float().eval()
half = device.type != 'cpu'
if half:
model.half()
model.eval()
img = torch.zeros((1, 3, config.image_size, config.image_size),
device=device)
_ = model(img.half() if half else img) if device.type != 'cpu' else None
file_names = [
file_name[:-4] for file_name in os.listdir(
os.path.join(config.data_dir, config.image_dir))
]
for file_name in file_names:
image = util.load_image(file_name)
label = util.load_label(file_name)
image, label = util.resize(image, label)
image = torch.from_numpy(np.expand_dims(image.transpose(2, 0, 1), 0))
image = image.to(device, non_blocking=True)
image = image.half() if half else image.float()
pred = model(image / 255.0)
pred = np.array(pred.cpu().detach().numpy()[0, :, :, :].transpose(
1, 2, 0))
pred = util.reverse_one_hot(pred)
pred = util.colour_code_segmentation(pred, palette)
pred = np.uint8(pred)
util.save_images([label, pred], ['TRUE', 'PRED'],
os.path.join('results', f'{file_name}.jpg'))
