def generateData(batch_size, data=[]):
# print 'generateData...'
while True:
train_data = []
train_label = []
batch = 0
for i in (range(len(data))):
url = data[i]
batch += 1
# img = load_img(train_data_path + 'src/' + url)
_, img = load_img_normalization(im_bands,
(train_data_path + 'src/' + url),
data_type=im_type)
# Adapt dim_ordering automatically
img = img_to_array(img)
train_data.append(img)
# label = load_img(train_data_path + 'label/' + url, grayscale=True)
_, label = load_img_normalization(
1, (train_data_path + 'label/' + url))
label = img_to_array(label)
train_label.append(label)
if batch % batch_size == 0:
# print 'get enough bacth!\n'
train_data = np.array(train_data)
train_label = np.array(train_label)
# train_label = to_categorical(train_label, num_classes=n_label) # one_hot coding
train_label = train_label.reshape(
(batch_size, img_w * img_h, n_label))
yield (train_data, train_label)
train_data = []
train_label = []
batch = 0
def generateValidData(batch_size, data=[]):
# print 'generateValidData...'
while True:
valid_data = []
valid_label = []
batch = 0
for i in (range(len(data))):
url = data[i]
batch += 1
_, img = load_img_normalization(
config.im_bands, (config.train_data_path + '/src/' + url),
data_type=im_type)
# Adapt dim_ordering automatically
img = img_to_array(img)
valid_data.append(img)
_, label = load_img_normalization(
1, (config.train_data_path + '/label/' + url))
label = img_to_array(label)
valid_label.append(label)
if batch % batch_size == 0:
valid_data = np.array(valid_data)
valid_label = np.array(valid_label)
if config.nb_classes > 2:
valid_label = to_categorical(valid_label,
num_classes=config.nb_classes)
yield (valid_data, valid_label)
valid_data = []
valid_label = []
batch = 0
if not os.path.isdir(train_data_path):
print("train data does not exist in the path:\n {}".format(
train_data_path))
if FLAG_USING_NETWORK == 0:
model = binary_unet_jaccard_4orMore(input_bands, n_label)
elif FLAG_USING_NETWORK == 1:
model = binary_fcnnet_jaccard(n_label)
elif FLAG_USING_NETWORK == 2:
model = binary_segnet_jaccard(n_label)
print("Train by : {}".format(dict_network[FLAG_USING_NETWORK]))
train(model, model_save_path)
if FLAG_MAKE_TEST:
print("test ....................predict by trained model .....\n")
test_img_path = '../../data/test/sample1.png'
import sys
if not os.path.isfile(test_img_path):
print("no file: {}".format(test_img_path))
sys.exit(-1)
ret, input_img = load_img_normalization(test_img_path)
# model_save_path ='../../data/models/unet_buildings_onehot.h5'
new_model = load_model(model_save_path)
test_predict(input_img, new_model)
def val_data_generator(config, sample_url):
# print 'generate validating Data...'
norm_value = 255.0
w = config.img_w
h = config.img_h
label_list, img_list = [], []
for pic in sample_url:
_, t_img = load_img_normalization(
1, config.train_data_path + '/label/' + pic)
tp = np.unique(t_img)
if len(tp) < 2:
print("Only one value {} in {}".format(
tp, config.train_data_path + '/label/' + pic))
if tp[0] == 0:
print("no target value in {}".format(config.train_data_path +
'/label/' + pic))
continue
ret, s_img = load_img_bybandlist(
(config.train_data_path + '/src/' + pic),
bandlist=config.band_list)
if ret != 0:
continue
s_img = img_to_array(s_img)
img_list.append(s_img)
label_list.append(t_img)
assert len(label_list) == len(img_list)
train_data = []
train_label = []
batch = 0
while True:
if batch == 0:
train_data = []
train_label = []
# batch = 0
for i in (range(len(img_list))):
img = img_list[i]
label = label_list[i]
img = np.asarray(img).astype("float") / norm_value
label = np.asarray(label)
assert img.shape[0:2] == label.shape[0:2]
for i in range(img.shape[0] // h):
for j in range(img.shape[1] // w):
x = img[i * h:(i + 1) * h, (j * w):(j + 1) * w, :]
y = label[i * h:(i + 1) * h, (j * w):(j + 1) * w]
if config.label_nodata in np.unique(y):
continue
"""ignore pure background area"""
if len(np.unique(y)) < 2:
if (0 in np.unique(y)) and (np.random.random() < 0.75):
continue
x = img_to_array(x)
y = img_to_array(y)
train_data.append(x)
train_label.append(y)
batch += 1
if batch % config.batch_size == 0:
train_data = np.array(train_data)
train_label = np.array(train_label)
if config.nb_classes > 2:
train_label = to_categorical(
train_label, num_classes=config.nb_classes)
yield (train_data, train_label)
train_data = []
train_label = []
batch = 0
def train_data_generator(config, sample_url):
# print 'generateData...'
norm_value = 255.0
bits_num = 8
if '10' in config.im_type:
norm_value = 1024.0
bits_num = 16
elif '16' in config.im_type:
norm_value = 65535.0
bits_num = 16
else:
pass
label_list, img_list = [], []
for pic in sample_url:
_, t_img = load_img_normalization(
1, config.train_data_path + '/label/' + pic)
tp = np.unique(t_img)
if len(tp) < 2:
print("Only one value {} in {}".format(
tp, config.train_data_path + '/label/' + pic))
if tp[0] == 0:
print("no target value in {}".format(config.train_data_path +
'/label/' + pic))
continue
ret, s_img = load_img_bybandlist(
(config.train_data_path + '/src/' + pic),
bandlist=config.band_list)
if ret != 0:
continue
s_img = img_to_array(s_img)
s_img = np.asarray(s_img, np.uint16)
# plt.imshow(s_img[:,:,0])
# plt.show()
img_list.append(s_img)
label_list.append(t_img)
assert len(label_list) == len(img_list)
train_data = []
train_label = []
batch = 0
while True:
if batch == 0: # 防止训练集图像数量少于batch_size
train_data = []
train_label = []
for i in np.random.permutation(np.arange(len(img_list))):
try:
src_img = img_list[i]
except:
print("can not extract data from img")
try:
label_img = label_list[i]
except:
print("can not extract data from label")
random_size = random.randrange(config.img_w, config.img_w * 2 + 1,
config.img_w)
# random_size = config.img_w
img, label = random_crop(src_img, label_img, random_size,
random_size)
if config.label_nodata in np.unique(label):
continue
"""ignore pure background area"""
if len(np.unique(label)) < 2:
if (0 in np.unique(label)) and (np.random.random() < 0.75):
continue
if img.shape[1] != config.img_w or img.shape[0] != config.img_h:
# print("resize samples")
img = resample_data(img,
config.img_h,
config.img_w,
mode=Image.BILINEAR,
bits=bits_num)
label = resample_data(label,
config.img_h,
config.img_w,
mode=Image.NEAREST)
if config.augment:
img, label = data_augment(img, label, config.img_w,
config.img_h)
img = np.asarray(img).astype(np.float32) / norm_value
img = np.clip(img, 0.0, 1.0)
batch += 1
img = img_to_array(img)
label = img_to_array(label)
train_data.append(img)
train_label.append(label)
if batch % config.batch_size == 0:
train_data = np.array(train_data)
train_label = np.array(train_label)
# print("img shape:{}".format(train_data.shape))
# print("label shap:{}".format(train_label.shape))
if config.nb_classes > 2:
train_label = to_categorical(train_label,
num_classes=config.nb_classes)
yield (train_data, train_label)
train_data = []
train_label = []
batch = 0
