def main(args):
os.environ["CUDA_VISIBLE_DEVICES"] = config.INFERENCE_GPU_ID
# 覆盖参数
config.IMAGES_PER_GPU = 1
config.GPU_COUNT = 1
config.DETECTION_MIN_CONFIDENCE = 0.0 # 评估阶段保证有足够的边框输出
# 加载数据集
dataset = VocDataset(config.voc_path, class_mapping=config.CLASS_MAPPING)
dataset.prepare()
print("len:{}".format(len(dataset.get_image_info_list())))
test_image_info_list = [
info for info in dataset.get_image_info_list()
if info['type'] == args.data_set
]
print("len:{}".format(len(test_image_info_list)))
gen = Generator(test_image_info_list, config.IMAGE_INPUT_SHAPE,
config.MEAN_PIXEL)
# 加载模型
m = models.frcnn(config, stage='test')
if args.weight_path is not None:
m.load_weights(args.weight_path, by_name=True)
else:
m.load_weights(config.rcnn_weights, by_name=True)
# m.summary()
# 预测边框、得分、类别
s_time = time.time()
boxes, scores, class_ids, _, image_metas = m.predict_generator(
gen.gen_val(), steps=gen.size, use_multiprocessing=True)
print("预测 {} 张图像,耗时：{} 秒".format(len(test_image_info_list),
time.time() - s_time))
# 去除padding
image_metas = image_utils.batch_parse_image_meta(image_metas)
predict_scores = [np_utils.remove_pad(score)[:, 0] for score in scores]
predict_labels = [np_utils.remove_pad(label)[:, 0] for label in class_ids]
# 还原检测边框到
predict_boxes = [
image_utils.recover_detect_boxes(np_utils.remove_pad(box), window,
scale) for box, window, scale in
zip(boxes, image_metas['window'], image_metas['scale'])
]
# 以下是评估过程
annotations = eval_utils.get_annotations(test_image_info_list,
config.NUM_CLASSES)
detections = eval_utils.get_detections(predict_boxes, predict_scores,
predict_labels, config.NUM_CLASSES)
average_precisions = eval_utils.voc_eval(annotations,
detections,
iou_threshold=0.5,
use_07_metric=True)
print("ap:{}".format(average_precisions))
# 求mean ap 去除背景类
mAP = np.mean(np.array(list(average_precisions.values()))[1:])
print("mAP:{}".format(mAP))
print("整个评估过程耗时：{} 秒".format(time.time() - s_time))
def main(args):
os.environ["CUDA_VISIBLE_DEVICES"] = config.INFERENCE_GPU_ID
# 覆盖参数
config.IMAGES_PER_GPU = 1
config.GPU_COUNT = 1
# 加载数据
dataset = VocDataset(config.voc_path, class_mapping=config.CLASS_MAPPING)
dataset.prepare()
all_img_info = [
info for info in dataset.get_image_info_list()
if info['type'] == 'test'
] # 测试集
# 加载模型
m = models.frcnn(config, stage='test')
m.load_weights(args.weight_path, by_name=True)
m.summary()
# class map 转为 id map
id_mapping = class_map_to_id_map(config.CLASS_MAPPING)
def _show_inference(id, ax=None):
image = image_utils.load_image(all_img_info[id]['filepath'])
image, image_meta, _ = image_utils.resize_image_and_gt(
image, config.IMAGE_MAX_DIM, all_img_info[id]['boxes'])
boxes, scores, class_ids, _, image_metas = m.predict([
np.expand_dims(image, axis=0),
np.expand_dims(image_meta, axis=0)
])
boxes = np_utils.remove_pad(boxes[0])
scores = np_utils.remove_pad(scores[0])[:, 0]
class_ids = np_utils.remove_pad(class_ids[0])[:, 0]
visualize.display_instances(image,
boxes[:5],
class_ids[:5],
id_mapping,
scores=scores[:5],
ax=ax)
# print("boxes num:{}".format(boxes.shape[0]))
# 随机展示9张图像
image_ids = np.random.choice(len(all_img_info), 9, replace=False)
fig = plt.figure(figsize=(20, 20))
for idx, image_id in enumerate(image_ids):
ax = fig.add_subplot(3, 3, idx + 1)
_show_inference(image_id, ax)
fig.savefig('demo_images/inferece_examples.{}.png'.format(
np.random.randint(10)))
def main(args):
set_gpu_growth(config.GPU_COUNT)
dataset = VocDataset(config.voc_path, class_mapping=config.CLASS_MAPPING)
dataset.prepare()
train_img_info = [info for info in dataset.get_image_info_list() if info['type'] == 'trainval'] # 训练集
print("train_img_info:{}".format(len(train_img_info)))
test_img_info = [info for info in dataset.get_image_info_list() if info['type'] == 'test'] # 测试集
print("test_img_info:{}".format(len(test_img_info)))
m = models.frcnn(config, stage='train')
# 加载预训练模型
init_epochs = args.init_epochs
if args.init_epochs > 0:
m.load_weights('/tmp/frcnn-rcnn.{:03d}.h5'.format(init_epochs), by_name=True)
else:
m.load_weights(config.pretrained_weights, by_name=True)
m.summary()
#
if init_epochs < 40:
train(m, 'heads', 40, init_epochs, config.LEARNING_RATE, train_img_info, test_img_info)
if init_epochs < 120:
train(m, '3+', 120, max(init_epochs, 40), config.LEARNING_RATE, train_img_info, test_img_info)
if init_epochs < 160:
train(m, 'all', 160, max(init_epochs, 120), config.LEARNING_RATE / 10, train_img_info, test_img_info)
def main():
# 加载数据
dataset = VocDataset(config.voc_path, class_mapping=config.CLASS_MAPPING)
dataset.prepare()
all_img_info = [info for info in dataset.get_image_info_list() if info['type'] == 'trainval'] # 测试集
# 加载模型
m = models.frcnn(config, stage='test')
m.load_weights(config.rcnn_weights, by_name=True)
m.summary()
# class map 转为 id map
id_mapping = class_map_to_id_map(config.CLASS_MAPPING)
def _show_inference(id, ax=None):
image, image_meta, _ = load_image_gt(id,
all_img_info[id]['filepath'],
config.IMAGE_MAX_DIM,
all_img_info[id]['boxes'])
boxes, scores, class_ids, class_logits = m.predict(
[np.expand_dims(image, axis=0), np.expand_dims(image_meta, axis=0)])
boxes = np_utils.remove_pad(boxes[0])
scores = np_utils.remove_pad(scores[0])[:, 0]
class_ids = np_utils.remove_pad(class_ids[0])[:, 0]
visualize.display_instances(image, boxes[:5],
class_ids[:5],
id_mapping,
scores=scores[:5],
ax=ax)
# 随机展示9张图像
image_ids = np.random.choice(len(all_img_info), 9, replace=False)
fig = plt.figure(figsize=(20, 20))
for idx, image_id in enumerate(image_ids):
ax = fig.add_subplot(3, 3, idx + 1)
_show_inference(image_id, ax)
fig.savefig('demo_images/inferece_examples.{}.png'.format(np.random.randint(10)))
def main(args):
set_gpu_growth(config.GPU_COUNT)
dataset = VocDataset(config.voc_path, class_mapping=config.CLASS_MAPPING)
dataset.prepare()
train_img_info = [
info for info in dataset.get_image_info_list()
if info['type'] == 'trainval'
] # 训练集
print("train_img_info:{}".format(len(train_img_info)))
test_img_info = [
info for info in dataset.get_image_info_list()
if info['type'] == 'test'
] # 测试集
print("test_img_info:{}".format(len(test_img_info)))
m = models.frcnn(config, stage='train')
# 加载预训练模型
init_epochs = args.init_epochs
if args.init_epochs > 0:
m.load_weights('/tmp/frcnn-{}.{:03d}.h5'.format(
config.BASE_NET_NAME, init_epochs),
by_name=True)
else:
m.load_weights(config.pretrained_weights, by_name=True)
# 生成器
train_gen = Generator(train_img_info,
config.IMAGE_INPUT_SHAPE,
config.MEAN_PIXEL,
config.BATCH_SIZE,
config.MAX_GT_INSTANCES,
horizontal_flip=config.USE_HORIZONTAL_FLIP,
random_crop=config.USE_RANDOM_CROP)
# 生成器
val_gen = Generator(test_img_info, config.IMAGE_INPUT_SHAPE,
config.MEAN_PIXEL, config.BATCH_SIZE,
config.MAX_GT_INSTANCES)
# 训练conv3 及以上
models.set_trainable(config.TRAIN_LAYERS, m)
loss_names = [
"rpn_bbox_loss", "rpn_class_loss", "rcnn_bbox_loss", "rcnn_class_loss"
]
model_utils.compile(m, config.LEARNING_RATE, config.LEARNING_MOMENTUM,
config.GRADIENT_CLIP_NORM, config.WEIGHT_DECAY,
loss_names, config.LOSS_WEIGHTS)
m.summary()
# 增加个性化度量
metric_names = [
'gt_num', 'positive_anchor_num', 'negative_anchor_num',
'rpn_miss_gt_num', 'rpn_gt_min_max_iou', 'roi_num', 'positive_roi_num',
'negative_roi_num', 'rcnn_miss_gt_num', 'rcnn_miss_gt_num_as',
'gt_min_max_iou'
]
model_utils.add_metrics(m, metric_names, m.outputs[-11:])
# 训练
m.fit_generator(train_gen.gen(),
epochs=args.epochs,
steps_per_epoch=len(train_img_info) // config.BATCH_SIZE,
verbose=1,
initial_epoch=init_epochs,
validation_data=val_gen.gen(),
validation_steps=len(test_img_info) // config.BATCH_SIZE,
use_multiprocessing=True,
callbacks=get_call_back())
def main(args):
from tensorflow.python.keras import backend as K
import tensorflow as tf
cfg = tf.ConfigProto()
cfg.gpu_options.allow_growth = True
session = tf.Session(config=cfg)
K.set_session(session)
# 覆盖参数
config.IMAGES_PER_GPU = 1
config.GPU_COUNT = 1
config.DETECTION_MIN_CONFIDENCE = 0.0 # 评估阶段保证有足够的边框输出
# 加载数据集
dataset = VocDataset(config.voc_path, class_mapping=config.CLASS_MAPPING)
dataset.prepare()
print("len:{}".format(len(dataset.get_image_info_list())))
test_image_info_list = [
info for info in dataset.get_image_info_list()
if info['type'] == args.data_set
][:args.eval_num]
print("len:{}".format(len(test_image_info_list)))
gen = TestGenerator(test_image_info_list, config.IMAGE_INPUT_SHAPE,
config.MEAN_PIXEL)
# 加载模型
m = models.frcnn(config, stage='test')
if args.weight_path is not None:
m.load_weights(args.weight_path, by_name=True)
else:
m.load_weights(config.rcnn_weights, by_name=True)
# m.summary()
# 预测边框、得分、类别
s_time = time.time()
boxes, scores, class_ids, _, image_metas = m.predict_generator(
gen, steps=gen.size, verbose=1, workers=4, use_multiprocessing=False)
print("预测 {} 张图像,耗时：{} 秒".format(len(test_image_info_list),
time.time() - s_time))
# 去除padding
image_metas = image_utils.batch_parse_image_meta(image_metas)
predict_scores = [np_utils.remove_pad(score)[:, 0] for score in scores]
predict_labels = [np_utils.remove_pad(label)[:, 0] for label in class_ids]
# 还原检测边框到
predict_boxes = [
image_utils.recover_detect_boxes(np_utils.remove_pad(box), window,
scale) for box, window, scale in
zip(boxes, image_metas['window'], image_metas['scale'])
]
# 以下是评估过程
annotations = eval_utils.get_annotations(test_image_info_list,
config.NUM_CLASSES)
detections = eval_utils.get_detections(predict_boxes, predict_scores,
predict_labels, config.NUM_CLASSES)
average_precisions = eval_utils.voc_eval(annotations,
detections,
iou_threshold=0.5,
use_07_metric=True)
print("ap:{}".format(average_precisions))
# 求mean ap 去除背景类
mAP = np.mean(np.array(list(average_precisions.values()))[1:])
print("mAP:{}".format(mAP))
print("整个评估过程耗时：{} 秒".format(time.time() - s_time))