def get_data(record, device):
data = dataset.val_dataset(record)
image, gt_boxes, ID = \
data['data'], data['boxes'], data['ID']
if config.eval_resize == False:
resized_img, scale = image, 1
else:
resized_img, scale = dataset.resize_img_by_short_and_max_size(
image, config.eval_image_short_size, config.eval_image_max_size)
original_height, original_width = image.shape[0:2]
height, width = resized_img.shape[0:2]
transposed_img = np.ascontiguousarray(resized_img.transpose(
2, 0, 1)[None, :, :, :],
dtype=np.float32)
im_info = np.array([height, width, scale, original_height, original_width],
dtype=np.float32)[None, :]
return transposed_img, gt_boxes, im_info, ID
def eval_all(args):
devs = args.devices.split(',')
misc.ensure_dir(config.eval_dir)
eval_file = open(os.path.join(config.eval_dir, 'results.txt'), 'a')
dataset_dict = dataset.val_dataset()
records = dataset_dict['records']
nr_records = len(records)
read_func = dataset_dict['read_func']
nr_devs = len(devs)
for epoch_num in range(args.start_epoch, args.end_epoch + 1):
model_file = osp.join(
config.output_dir, 'model_dump',
'epoch_{:d}'.format(epoch_num) + '.ckpt')
pbar = tqdm(total=nr_records)
all_results = []
if nr_devs == 1:
func, inputs = load_model(model_file, devs[0])
for record in records:
data_dict = read_func(record)
result_dict = inference(func, inputs, data_dict)
all_results.append(result_dict)
if args.show_image:
image = result_dict['data']
for db in result_dict['result_boxes']:
if db.score > config.test_vis_threshold:
db.draw(image)
'''if 'boxes' in result_dict.keys():
for db in result_dict['boxes']:
db.draw(image)'''
path_img_unique = '/home/erik/Documents/light_head_rcnn-master/output/predictions/'+str(uuid.uuid4())+'.jpg'
cv2.imwrite(path_img_unique, image)
cv2.imshow('image', image)
cv2.waitKey(0)
pbar.update(1)
else:
nr_image = math.ceil(nr_records / nr_devs)
result_queue = Queue(500)
procs = []
for i in range(nr_devs):
start = i * nr_image
end = min(start + nr_image, nr_records)
split_records = records[start:end]
proc = Process(target=worker, args=(
model_file, devs[i], split_records, read_func,
result_queue))
print('process:%d, start:%d, end:%d' % (i, start, end))
proc.start()
procs.append(proc)
for i in range(nr_records):
t = result_queue.get()
all_results.append(t)
pbar.update(1)
for p in procs:
p.join()
save_filename = save_result(all_results, config.eval_dir, model_file)
print('Save to %s finished, start evaulation!' % save_filename)
saved_stdout = sys.stdout
sys.stdout = eval_file
print("\nevaluation epoch {}".format(epoch_num))
'''if config.test_save_type == 'motherboard': #'coco':
from datasets_odgt.cocoval import cocoval
cocoval(save_filename)
else:
print("not implement")
embed()'''
sys.stdout = saved_stdout
eval_file.flush()
eval_file.close()
print("\n")
import tensorflow as tf
from tqdm import tqdm
from utils.py_utils import misc
if __name__ == '__main__':
parser = make_parser()
args = parser.parse_args()
args.devices = misc.parse_devices(args.devices)
if args.end_epoch == -1:
args.end_epoch = args.start_epoch
devs = args.devices.split(',')
misc.ensure_dir(config.eval_dir)
eval_file = open(os.path.join(config.eval_dir, 'results.txt'), 'a')
dataset_dict = dataset.val_dataset()
records = dataset_dict['records']
nr_records = len(records)
read_func = dataset_dict['read_func']
nr_devs = len(devs)
for epoch_num in range(args.start_epoch, args.end_epoch + 1):
model_file = osp.join(config.output_dir, 'model_dump',
'epoch_{:d}'.format(epoch_num) + '.ckpt')
pbar = tqdm(total=nr_records)
all_results = []
if nr_devs > 1:
with tf.variable_scope('', reuse=tf.AUTO_REUSE):
os.environ["CUDA_VISIBLE_DEVICES"] = devs[0]
tfconfig = tf.ConfigProto(allow_soft_placement=True)
def eval_all(args):
devs = args.devices.split(',')
misc.ensure_dir(config.eval_dir)
eval_file = open(os.path.join(config.eval_dir, 'results.txt'), 'a')
dataset_dict = dataset.val_dataset()
records = dataset_dict['records']
nr_records = len(records)
read_func = dataset_dict['read_func']
nr_devs = len(devs)
for epoch_num in range(args.start_epoch, args.end_epoch + 1):
model_file = osp.join(
config.output_dir, 'model_dump',
'epoch_{:d}'.format(epoch_num) + '.ckpt')
pbar = tqdm(total=nr_records)
all_results = []
if nr_devs == 1:
func, inputs = load_model(model_file, devs[0])
for record in records:
data_dict = read_func(record)
result_dict = inference(func, inputs, data_dict)
all_results.append(result_dict)
if args.show_image:
image = result_dict['data']
for db in result_dict['result_boxes']:
if db.score > config.test_vis_threshold:
db.draw(image)
if 'boxes' in result_dict.keys():
for db in result_dict['boxes']:
db.draw(image)
cv2.imwrite('/tmp/hehe.png', image)
# cv2.imshow('image', image)
# cv2.waitKey(0)
pbar.update(1)
else:
nr_image = math.ceil(nr_records / nr_devs)
result_queue = Queue(500)
procs = []
for i in range(nr_devs):
start = i * nr_image
end = min(start + nr_image, nr_records)
split_records = records[start:end]
proc = Process(target=worker, args=(
model_file, devs[i], split_records, read_func,
result_queue))
print('process:%d, start:%d, end:%d' % (i, start, end))
proc.start()
procs.append(proc)
for i in range(nr_records):
t = result_queue.get()
all_results.append(t)
pbar.update(1)
for p in procs:
p.join()
save_filename = save_result(all_results, config.eval_dir, model_file)
print('Save to %s finished, start evaulation!' % save_filename)
saved_stdout = sys.stdout
sys.stdout = eval_file
print("\nevaluation epoch {}".format(epoch_num))
if config.test_save_type == 'coco':
from datasets_odgt.cocoval import cocoval
cocoval(save_filename)
else:
print("not implement")
embed()
sys.stdout = saved_stdout
eval_file.flush()
eval_file.close()
print("\n")
def eval_all(args):
devs = args.devices.split(',')
misc.ensure_dir(config.eval_dir)
eval_file = open(os.path.join(config.eval_dir, 'results.txt'), 'a')
'''e.g. {
'record': files,
'nr_records': total_files,
'read_func': read_func,
}'''
dataset_dict = dataset.val_dataset()
records = dataset_dict['records']
nr_records = len(records)
read_func = dataset_dict['read_func'] # 数据处理
nr_devs = len(devs)
for epoch_num in range(args.start_epoch, args.end_epoch + 1):
model_file = osp.join(
config.output_dir, 'model_dump',
'epoch_{:d}.ckpt-{:d}'.format(epoch_num, args.step))
pbar = tqdm(total=nr_records)
all_results = []
if nr_devs == 1:
func, inputs = load_model(model_file, devs[0])
for record in records:
data_dict = read_func(record) # 返回一个包含测试文件信息的dict
result_dict = inference(func, inputs, data_dict)
all_results.append(result_dict)
# result_dict包含gt和检出框
# result_dict['boxes'] --> gt
# result_dict['result_boxes'] --> 检出框
if args.show_image:
image = result_dict['data']
# 咦,这里是把检出框画上去,db是一个对象,概率信息包含在db里面
for db in result_dict['result_boxes']:
if db.score > config.test_vis_threshold:
db.draw(image)
# 咦,这里是把gt画上去
if 'boxes' in result_dict.keys():
for db in result_dict['boxes']:
db.draw(image)
fname = os.path.basename(data_dict['image_path'])
cv2.imwrite(os.path.join('images', fname), image)
# cv2.imshow('image', image)
# cv2.waitKey(0)
# pbar.update(1)
else:
nr_image = math.ceil(nr_records / nr_devs)
result_queue = Queue(500)
procs = []
for i in range(nr_devs):
start = i * nr_image
end = min(start + nr_image, nr_records)
split_records = records[start:end]
proc = Process(target=worker, args=(
model_file, devs[i], split_records, read_func,
result_queue))
print('process:%d, start:%d, end:%d' % (i, start, end))
proc.start()
procs.append(proc)
for i in range(nr_records):
t = result_queue.get()
all_results.append(t)
pbar.update(1)
for p in procs:
p.join()
save_filename = save_result(all_results, config.eval_dir, model_file)
print('Save to %s finished, start evaulation!' % save_filename)
saved_stdout = sys.stdout
sys.stdout = eval_file
print("\nevaluation epoch {}".format(epoch_num))
if config.test_save_type == 'adas':
from datasets_odgt.cocoval import cocoval
cocoval(save_filename)
else:
print("not implement")
embed()
sys.stdout = saved_stdout
eval_file.flush()
eval_file.close()
print("\n")
def eval_all(args):
devs = args.devices.split(',')
misc.ensure_dir(config.eval_dir)
'''e.g. {
'record': files,
'nr_records': total_files,
'read_func': read_func,
}'''
dataset_dict = dataset.val_dataset()
records = dataset_dict['records']
nr_records = len(records)
read_func = dataset_dict['read_func'] # 数据处理
nr_devs = len(devs)
model_file = osp.join(
config.output_dir, 'model_dump',
'light-head.ckpt-{:d}'.format(args.step))
#pbar = tqdm(total=nr_records)
all_results = []
if nr_devs == 1:
print('restore from ckpt %s' % model_file)
func, inputs = load_model(model_file, devs[0])
num_records = len(records)
for i in range(num_records):
record = records[i]
'''data_dict, e.g. {
'data': xx,
'boxes': xx,
'image_id': xx,
'image_path': xx,
}
'''
sys.stdout.write('im_detect: {:d}/{:d}\r'.format(i, num_records))
data_dict = read_func(record) # 返回一个包含测试文件信息的dict
result_dict = inference(func, inputs, data_dict)
all_results.append(result_dict)
# result_dict包含gt和检出框
# result_dict['boxes'] --> gt
# result_dict['result_boxes'] --> 检出框
if args.show_image:
image = result_dict['data']
# 咦,这里是把检出框画上去,db是一个对象,概率信息包含在db里面
for db in result_dict['result_boxes']:
if db.score > config.test_vis_threshold:
# 见lib/detection_opr/box_utils/box.py
db.draw(image)
# 咦,这里是把gt画上去
if 'boxes' in result_dict.keys():
for db in result_dict['boxes']:
db.draw(image)
fname = os.path.basename(data_dict['image_path'])
cv2.imwrite(os.path.join('eval', fname), image)
# cv2.imshow('image', image)
# cv2.waitKey(0)
# pbar.update(1)
else:
nr_image = math.ceil(nr_records / nr_devs)
result_queue = Queue(500)
procs = []
for i in range(nr_devs):
start = i * nr_image
end = min(start + nr_image, nr_records)
split_records = records[start:end]
proc = Process(target=worker, args=(
model_file, devs[i], split_records, read_func,
result_queue))
print('process:%d, start:%d, end:%d' % (i, start, end))
proc.start()
procs.append(proc)
for i in range(nr_records):
t = result_queue.get()
all_results.append(t)
#pbar.update(1)
for p in procs:
p.join()
save_result_v2(all_results, config.eval_dir)
print('Save result finished, start evaulation!')
if config.test_save_type == 'adas':
from datasets_odgt.adasval import adasval
adasval(cfg)
else:
print("not implement")
embed()