NET_NAME = 'resnet101_v1d' # 'MobilenetV2'
# ---------------------------------------- System
ROOT_PATH = os.path.abspath('../../')
print(20 * "++--")
print(ROOT_PATH)
GPU_GROUP = "1,2,3"
NUM_GPU = len(GPU_GROUP.strip().split(','))
SHOW_TRAIN_INFO_INTE = 20
SMRY_ITER = 200
SAVE_WEIGHTS_INTE = 10000 * 2
SUMMARY_PATH = os.path.join(ROOT_PATH, 'output/summary')
TEST_SAVE_PATH = os.path.join(ROOT_PATH, 'tools/test_result')
pretrain_zoo = PretrainModelZoo()
PRETRAINED_CKPT = pretrain_zoo.pretrain_weight_path(NET_NAME, ROOT_PATH)
TRAINED_CKPT = os.path.join(ROOT_PATH, 'output/trained_weights')
EVALUATE_R_DIR = os.path.join(ROOT_PATH, 'output/evaluate_result_pickle/')
# ------------------------------------------ Train and test
RESTORE_FROM_RPN = False
FIXED_BLOCKS = 1 # allow 0~3
FREEZE_BLOCKS = [True, False, False, False, False] # for gluoncv backbone
USE_07_METRIC = True
EVAL_THRESHOLD = 0.5
ADD_BOX_IN_TENSORBOARD = True
MUTILPY_BIAS_GRADIENT = 2.0 # if None, will not multipy
GRADIENT_CLIPPING_BY_NORM = 10.0 # if None, will not clip
def eval_with_plac(self, img_dir, det_net, image_ext):
os.environ["CUDA_VISIBLE_DEVICES"] = self.args.gpu
# 1. preprocess img
img_plac = tf.placeholder(dtype=tf.uint8, shape=[None, None,
3]) # is RGB. not BGR
img_batch = tf.cast(img_plac, tf.float32)
pretrain_zoo = PretrainModelZoo()
if self.cfgs.NET_NAME in pretrain_zoo.pth_zoo or self.cfgs.NET_NAME in pretrain_zoo.mxnet_zoo:
img_batch = (img_batch / 255 - tf.constant(
self.cfgs.PIXEL_MEAN_)) / tf.constant(self.cfgs.PIXEL_STD)
else:
img_batch = img_batch - tf.constant(self.cfgs.PIXEL_MEAN)
img_batch = tf.expand_dims(img_batch, axis=0)
detection_boxes, detection_scores, detection_category = det_net.build_whole_detection_network(
input_img_batch=img_batch)
init_op = tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
restorer, restore_ckpt = det_net.get_restorer()
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
sess.run(init_op)
if not restorer is None:
restorer.restore(sess, restore_ckpt)
print('restore model')
all_boxes_r = []
imgs = os.listdir(img_dir)
pbar = tqdm(imgs)
for a_img_name in pbar:
# if '1546' not in a_img_name:
# continue
a_img_name = a_img_name.split(image_ext)[0]
raw_img = cv2.imread(
os.path.join(img_dir, a_img_name + image_ext))
raw_h, raw_w = raw_img.shape[0], raw_img.shape[1]
det_boxes_r_all, det_scores_r_all, det_category_r_all = [], [], []
img_short_side_len_list = self.cfgs.IMG_SHORT_SIDE_LEN if isinstance(
self.cfgs.IMG_SHORT_SIDE_LEN,
list) else [self.cfgs.IMG_SHORT_SIDE_LEN]
img_short_side_len_list = [
img_short_side_len_list[0]
] if not self.args.multi_scale else img_short_side_len_list
for short_size in img_short_side_len_list:
max_len = self.cfgs.IMG_MAX_LENGTH
if raw_h < raw_w:
new_h, new_w = short_size, min(
int(short_size * float(raw_w) / raw_h), max_len)
else:
new_h, new_w = min(
int(short_size * float(raw_h) / raw_w),
max_len), short_size
img_resize = cv2.resize(raw_img, (new_w, new_h))
resized_img, detected_boxes, detected_scores, detected_categories = \
sess.run(
[img_batch, detection_boxes, detection_scores, detection_category],
feed_dict={img_plac: img_resize[:, :, ::-1]}
)
if detected_boxes.shape[0] == 0:
continue
resized_h, resized_w = resized_img.shape[
1], resized_img.shape[2]
# detected_boxes = forward_convert(detected_boxes, False)
detected_boxes[:, 0::2] *= (raw_w / resized_w)
detected_boxes[:, 1::2] *= (raw_h / resized_h)
det_boxes_r_all.extend(detected_boxes)
det_scores_r_all.extend(detected_scores)
det_category_r_all.extend(detected_categories)
det_boxes_r_all = np.array(det_boxes_r_all)
det_scores_r_all = np.array(det_scores_r_all)
det_category_r_all = np.array(det_category_r_all)
box_res_rotate_ = []
label_res_rotate_ = []
score_res_rotate_ = []
if det_scores_r_all.shape[0] != 0:
for sub_class in range(1, self.cfgs.CLASS_NUM + 1):
index = np.where(det_category_r_all == sub_class)[0]
if len(index) == 0:
continue
tmp_boxes_r = det_boxes_r_all[index]
tmp_label_r = det_category_r_all[index]
tmp_score_r = det_scores_r_all[index]
if self.args.multi_scale:
tmp_boxes_r_ = backward_convert(tmp_boxes_r, False)
# try:
# inx = nms_rotate.nms_rotate_cpu(boxes=np.array(tmp_boxes_r_),
# scores=np.array(tmp_score_r),
# iou_threshold=self.cfgs.NMS_IOU_THRESHOLD,
# max_output_size=5000)
# except:
tmp_boxes_r_ = np.array(tmp_boxes_r_)
tmp = np.zeros([
tmp_boxes_r_.shape[0],
tmp_boxes_r_.shape[1] + 1
])
tmp[:, 0:-1] = tmp_boxes_r_
tmp[:, -1] = np.array(tmp_score_r)
# Note: the IoU of two same rectangles is 0, which is calculated by rotate_gpu_nms
jitter = np.zeros([
tmp_boxes_r_.shape[0],
tmp_boxes_r_.shape[1] + 1
])
jitter[:, 0] += np.random.rand(
tmp_boxes_r_.shape[0], ) / 1000
inx = rotate_gpu_nms(
np.array(tmp, np.float32) +
np.array(jitter, np.float32),
float(self.cfgs.NMS_IOU_THRESHOLD), 0)
else:
inx = np.arange(0, tmp_score_r.shape[0])
box_res_rotate_.extend(np.array(tmp_boxes_r)[inx])
score_res_rotate_.extend(np.array(tmp_score_r)[inx])
label_res_rotate_.extend(np.array(tmp_label_r)[inx])
if len(box_res_rotate_) == 0:
all_boxes_r.append(np.array([]))
continue
det_boxes_r_ = np.array(box_res_rotate_)
det_scores_r_ = np.array(score_res_rotate_)
det_category_r_ = np.array(label_res_rotate_)
if self.args.draw_imgs:
detected_indices = det_scores_r_ >= self.cfgs.VIS_SCORE
detected_scores = det_scores_r_[detected_indices]
detected_boxes = det_boxes_r_[detected_indices]
detected_categories = det_category_r_[detected_indices]
# detected_boxes = backward_convert(detected_boxes, False)
drawer = DrawBox(self.cfgs)
det_detections_r = drawer.draw_boxes_with_label_and_scores(
raw_img[:, :, ::-1],
boxes=detected_boxes,
labels=detected_categories,
scores=detected_scores,
method=2,
in_graph=True)
save_dir = os.path.join('test_hrsc', self.cfgs.VERSION,
'hrsc2016_img_vis')
tools.makedirs(save_dir)
cv2.imwrite(save_dir + '/{}.jpg'.format(a_img_name),
det_detections_r[:, :, ::-1])
det_boxes_r_ = backward_convert(det_boxes_r_, False)
x_c, y_c, w, h, theta = det_boxes_r_[:, 0], det_boxes_r_[:, 1], det_boxes_r_[:, 2], \
det_boxes_r_[:, 3], det_boxes_r_[:, 4]
boxes_r = np.transpose(np.stack([x_c, y_c, w, h, theta]))
dets_r = np.hstack((det_category_r_.reshape(-1, 1),
det_scores_r_.reshape(-1, 1), boxes_r))
all_boxes_r.append(dets_r)
pbar.set_description("Eval image %s" % a_img_name)
# fw1 = open(cfgs.VERSION + '_detections_r.pkl', 'wb')
# pickle.dump(all_boxes_r, fw1)
return all_boxes_r
def main(self):
with tf.Graph().as_default() as graph, tf.device('/cpu:0'):
num_gpu = len(cfgs.GPU_GROUP.strip().split(','))
global_step = slim.get_or_create_global_step()
lr = self.warmup_lr(cfgs.LR, global_step, cfgs.WARM_SETP, num_gpu)
tf.summary.scalar('lr', lr)
optimizer = tf.train.MomentumOptimizer(lr, momentum=cfgs.MOMENTUM)
refine_retinanet = build_whole_network.DetectionNetworkRefineRetinaNet(
cfgs=self.cfgs, is_training=True)
with tf.name_scope('get_batch'):
if cfgs.IMAGE_PYRAMID:
shortside_len_list = tf.constant(cfgs.IMG_SHORT_SIDE_LEN)
shortside_len = tf.random_shuffle(shortside_len_list)[0]
else:
shortside_len = cfgs.IMG_SHORT_SIDE_LEN
img_name_batch, img_batch, gtboxes_and_label_batch, num_objects_batch, img_h_batch, img_w_batch = \
self.reader.next_batch(dataset_name=cfgs.DATASET_NAME,
batch_size=cfgs.BATCH_SIZE * num_gpu,
shortside_len=shortside_len,
is_training=True)
# data processing
inputs_list = []
for i in range(num_gpu):
img = tf.expand_dims(img_batch[i], axis=0)
pretrain_zoo = PretrainModelZoo()
if self.cfgs.NET_NAME in pretrain_zoo.pth_zoo or self.cfgs.NET_NAME in pretrain_zoo.mxnet_zoo:
img = img / tf.constant([cfgs.PIXEL_STD])
gtboxes_and_label_r = tf.py_func(
backward_convert,
inp=[gtboxes_and_label_batch[i]],
Tout=tf.float32)
gtboxes_and_label_r = tf.reshape(gtboxes_and_label_r, [-1, 6])
gtboxes_and_label_h = get_horizen_minAreaRectangle(
gtboxes_and_label_batch[i])
gtboxes_and_label_h = tf.reshape(gtboxes_and_label_h, [-1, 5])
num_objects = num_objects_batch[i]
num_objects = tf.cast(tf.reshape(num_objects, [
-1,
]), tf.float32)
img_h = img_h_batch[i]
img_w = img_w_batch[i]
inputs_list.append([
img, gtboxes_and_label_h, gtboxes_and_label_r, num_objects,
img_h, img_w
])
tower_grads = []
biases_regularizer = tf.no_regularizer
weights_regularizer = tf.contrib.layers.l2_regularizer(
cfgs.WEIGHT_DECAY)
with tf.variable_scope(tf.get_variable_scope()):
for i in range(num_gpu):
with tf.device('/gpu:%d' % i):
with tf.name_scope('tower_%d' % i):
with slim.arg_scope(
[slim.model_variable, slim.variable],
device='/device:CPU:0'):
with slim.arg_scope(
[
slim.conv2d, slim.conv2d_in_plane,
slim.conv2d_transpose,
slim.separable_conv2d,
slim.fully_connected
],
weights_regularizer=weights_regularizer,
biases_regularizer=biases_regularizer,
biases_initializer=tf.
constant_initializer(0.0)):
gtboxes_and_label_h, gtboxes_and_label_r = tf.py_func(
self.get_gtboxes_and_label,
inp=[
inputs_list[i][1],
inputs_list[i][2],
inputs_list[i][3]
],
Tout=[tf.float32, tf.float32])
gtboxes_and_label_h = tf.reshape(
gtboxes_and_label_h, [-1, 5])
gtboxes_and_label_r = tf.reshape(
gtboxes_and_label_r, [-1, 6])
img = inputs_list[i][0]
img_shape = inputs_list[i][-2:]
img = tf.image.crop_to_bounding_box(
image=img,
offset_height=0,
offset_width=0,
target_height=tf.cast(
img_shape[0], tf.int32),
target_width=tf.cast(
img_shape[1], tf.int32))
outputs = refine_retinanet.build_whole_detection_network(
input_img_batch=img,
gtboxes_batch_h=gtboxes_and_label_h,
gtboxes_batch_r=gtboxes_and_label_r,
gpu_id=i)
gtboxes_in_img_h = self.drawer.draw_boxes_with_categories(
img_batch=img,
boxes=gtboxes_and_label_h[:, :-1],
labels=gtboxes_and_label_h[:, -1],
method=0)
gtboxes_in_img_r = self.drawer.draw_boxes_with_categories(
img_batch=img,
boxes=gtboxes_and_label_r[:, :-1],
labels=gtboxes_and_label_r[:, -1],
method=1)
tf.summary.image(
'Compare/gtboxes_h_gpu:%d' % i,
gtboxes_in_img_h)
tf.summary.image(
'Compare/gtboxes_r_gpu:%d' % i,
gtboxes_in_img_r)
if cfgs.ADD_BOX_IN_TENSORBOARD:
detections_in_img = self.drawer.draw_boxes_with_categories_and_scores(
img_batch=img,
boxes=outputs[0],
scores=outputs[1],
labels=outputs[2],
method=1)
tf.summary.image(
'Compare/final_detection_gpu:%d' %
i, detections_in_img)
loss_dict = outputs[-1]
total_loss_dict, total_losses = self.loss_dict(
loss_dict, num_gpu)
if i == num_gpu - 1:
regularization_losses = tf.get_collection(
tf.GraphKeys.REGULARIZATION_LOSSES)
# weight_decay_loss = tf.add_n(slim.losses.get_regularization_losses())
total_losses = total_losses + tf.add_n(
regularization_losses)
tf.get_variable_scope().reuse_variables()
grads = optimizer.compute_gradients(total_losses)
if cfgs.GRADIENT_CLIPPING_BY_NORM is not None:
grads = slim.learning.clip_gradient_norms(
grads, cfgs.GRADIENT_CLIPPING_BY_NORM)
tower_grads.append(grads)
self.log_printer(refine_retinanet, optimizer, global_step,
tower_grads, total_loss_dict, num_gpu, graph)
def worker(self, gpu_id, images, det_net, result_queue):
os.environ["CUDA_VISIBLE_DEVICES"] = str(gpu_id)
img_plac = tf.placeholder(dtype=tf.uint8, shape=[None, None, 3]) # is RGB. not BGR
img_batch = tf.cast(img_plac, tf.float32)
pretrain_zoo = PretrainModelZoo()
if self.cfgs.NET_NAME in pretrain_zoo.pth_zoo or self.cfgs.NET_NAME in pretrain_zoo.mxnet_zoo:
img_batch = (img_batch / 255 - tf.constant(self.cfgs.PIXEL_MEAN_)) / tf.constant(self.cfgs.PIXEL_STD)
else:
img_batch = img_batch - tf.constant(self.cfgs.PIXEL_MEAN)
img_batch = tf.expand_dims(img_batch, axis=0)
detection_boxes, detection_scores, detection_category = det_net.build_whole_detection_network(
input_img_batch=img_batch)
init_op = tf.group(
tf.global_variables_initializer(),
tf.local_variables_initializer()
)
restorer, restore_ckpt = det_net.get_restorer()
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
sess.run(init_op)
if not restorer is None:
restorer.restore(sess, restore_ckpt)
print('restore model %d ...' % gpu_id)
for img_path in images:
# if 'P0015' not in img_path:
# continue
img = cv2.imread(img_path)
# img = np.load(img_path.replace('images', 'npy').replace('.png', '.npy'))
box_res_rotate = []
label_res_rotate = []
score_res_rotate = []
imgH = img.shape[0]
imgW = img.shape[1]
img_short_side_len_list = self.cfgs.IMG_SHORT_SIDE_LEN if isinstance(self.cfgs.IMG_SHORT_SIDE_LEN, list) else [
self.cfgs.IMG_SHORT_SIDE_LEN]
img_short_side_len_list = [img_short_side_len_list[0]] if not self.args.multi_scale else img_short_side_len_list
if imgH < self.args.h_len:
temp = np.zeros([self.args.h_len, imgW, 3], np.float32)
temp[0:imgH, :, :] = img
img = temp
imgH = self.args.h_len
if imgW < self.args.w_len:
temp = np.zeros([imgH, self.args.w_len, 3], np.float32)
temp[:, 0:imgW, :] = img
img = temp
imgW = self.args.w_len
for hh in range(0, imgH, self.args.h_len - self.args.h_overlap):
if imgH - hh - 1 < self.args.h_len:
hh_ = imgH - self.args.h_len
else:
hh_ = hh
for ww in range(0, imgW, self.args.w_len - self.args.w_overlap):
if imgW - ww - 1 < self.args.w_len:
ww_ = imgW - self.args.w_len
else:
ww_ = ww
src_img = img[hh_:(hh_ + self.args.h_len), ww_:(ww_ + self.args.w_len), :]
for short_size in img_short_side_len_list:
max_len = self.cfgs.IMG_MAX_LENGTH
if self.args.h_len < self.args.w_len:
new_h, new_w = short_size, min(int(short_size * float(self.args.w_len) / self.args.h_len), max_len)
else:
new_h, new_w = min(int(short_size * float(self.args.h_len) / self.args.w_len), max_len), short_size
img_resize = cv2.resize(src_img, (new_w, new_h))
resized_img, det_boxes_r_, det_scores_r_, det_category_r_ = \
sess.run(
[img_batch, detection_boxes, detection_scores, detection_category],
feed_dict={img_plac: img_resize[:, :, ::-1]}
)
resized_h, resized_w = resized_img.shape[1], resized_img.shape[2]
src_h, src_w = src_img.shape[0], src_img.shape[1]
if len(det_boxes_r_) > 0:
det_boxes_r_ = forward_convert(det_boxes_r_, False)
det_boxes_r_[:, 0::2] *= (src_w / resized_w)
det_boxes_r_[:, 1::2] *= (src_h / resized_h)
for ii in range(len(det_boxes_r_)):
box_rotate = det_boxes_r_[ii]
box_rotate[0::2] = box_rotate[0::2] + ww_
box_rotate[1::2] = box_rotate[1::2] + hh_
box_res_rotate.append(box_rotate)
label_res_rotate.append(det_category_r_[ii])
score_res_rotate.append(det_scores_r_[ii])
if self.args.flip_img:
det_boxes_r_flip, det_scores_r_flip, det_category_r_flip = \
sess.run(
[detection_boxes, detection_scores, detection_category],
feed_dict={img_plac: cv2.flip(img_resize, flipCode=1)[:, :, ::-1]}
)
if len(det_boxes_r_flip) > 0:
det_boxes_r_flip = forward_convert(det_boxes_r_flip, False)
det_boxes_r_flip[:, 0::2] *= (src_w / resized_w)
det_boxes_r_flip[:, 1::2] *= (src_h / resized_h)
for ii in range(len(det_boxes_r_flip)):
box_rotate = det_boxes_r_flip[ii]
box_rotate[0::2] = (src_w - box_rotate[0::2]) + ww_
box_rotate[1::2] = box_rotate[1::2] + hh_
box_res_rotate.append(box_rotate)
label_res_rotate.append(det_category_r_flip[ii])
score_res_rotate.append(det_scores_r_flip[ii])
det_boxes_r_flip, det_scores_r_flip, det_category_r_flip = \
sess.run(
[detection_boxes, detection_scores, detection_category],
feed_dict={img_plac: cv2.flip(img_resize, flipCode=0)[:, :, ::-1]}
)
if len(det_boxes_r_flip) > 0:
det_boxes_r_flip = forward_convert(det_boxes_r_flip, False)
det_boxes_r_flip[:, 0::2] *= (src_w / resized_w)
det_boxes_r_flip[:, 1::2] *= (src_h / resized_h)
for ii in range(len(det_boxes_r_flip)):
box_rotate = det_boxes_r_flip[ii]
box_rotate[0::2] = box_rotate[0::2] + ww_
box_rotate[1::2] = (src_h - box_rotate[1::2]) + hh_
box_res_rotate.append(box_rotate)
label_res_rotate.append(det_category_r_flip[ii])
score_res_rotate.append(det_scores_r_flip[ii])
box_res_rotate = np.array(box_res_rotate)
label_res_rotate = np.array(label_res_rotate)
score_res_rotate = np.array(score_res_rotate)
box_res_rotate_ = []
label_res_rotate_ = []
score_res_rotate_ = []
threshold = {'roundabout': 0.1, 'tennis-court': 0.3, 'swimming-pool': 0.1, 'storage-tank': 0.2,
'soccer-ball-field': 0.3, 'small-vehicle': 0.2, 'ship': 0.2, 'plane': 0.3,
'large-vehicle': 0.1, 'helicopter': 0.2, 'harbor': 0.0001, 'ground-track-field': 0.3,
'bridge': 0.0001, 'basketball-court': 0.3, 'baseball-diamond': 0.3,
'container-crane': 0.05, 'airport': 0.1, 'helipad': 0.1}
for sub_class in range(1, self.cfgs.CLASS_NUM + 1):
index = np.where(label_res_rotate == sub_class)[0]
if len(index) == 0:
continue
tmp_boxes_r = box_res_rotate[index]
tmp_label_r = label_res_rotate[index]
tmp_score_r = score_res_rotate[index]
tmp_boxes_r_ = backward_convert(tmp_boxes_r, False)
# try:
# inx = nms_rotate.nms_rotate_cpu(boxes=np.array(tmp_boxes_r_),
# scores=np.array(tmp_score_r),
# iou_threshold=threshold[self.label_name_map[sub_class]],
# max_output_size=5000)
#
# except:
tmp_boxes_r_ = np.array(tmp_boxes_r_)
tmp = np.zeros([tmp_boxes_r_.shape[0], tmp_boxes_r_.shape[1] + 1])
tmp[:, 0:-1] = tmp_boxes_r_
tmp[:, -1] = np.array(tmp_score_r)
# Note: the IoU of two same rectangles is 0, which is calculated by rotate_gpu_nms
jitter = np.zeros([tmp_boxes_r_.shape[0], tmp_boxes_r_.shape[1] + 1])
jitter[:, 0] += np.random.rand(tmp_boxes_r_.shape[0], ) / 1000
inx = rotate_gpu_nms(np.array(tmp, np.float32) + np.array(jitter, np.float32),
float(threshold[self.label_name_map[sub_class]]), 0)
box_res_rotate_.extend(np.array(tmp_boxes_r)[inx])
score_res_rotate_.extend(np.array(tmp_score_r)[inx])
label_res_rotate_.extend(np.array(tmp_label_r)[inx])
result_dict = {'boxes': np.array(box_res_rotate_), 'scores': np.array(score_res_rotate_),
'labels': np.array(label_res_rotate_), 'image_id': img_path}
result_queue.put_nowait(result_dict)