def main():
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "1"
args = parse_args()
args.out = args.save_path + 'result_test.pkl'
if args.out is not None and not args.out.endswith(('.pkl', '.pickle')):
raise ValueError('The output file must be a pkl file.')
cfg = mmcv.Config.fromfile(args.config)
# set cudnn_benchmark
if cfg.get('cudnn_benchmark', False):
torch.backends.cudnn.benchmark = True
cfg.model.pretrained = None
cfg.data.test.test_mode = True
# get dataset
dataset = obj_from_dict(cfg.data.test, datasets, dict(test_mode=True))
# build model
assert args.gpus == 1
model = build_detector(cfg.model, train_cfg=None, test_cfg=cfg.test_cfg)
load_checkpoint(model, args.checkpoint)
model = MMDataParallel(model, device_ids=[0])
# model_flow = FlowNetCHead()
# load_checkpoint(model_flow, args.checkpointflow)
data_loader = build_dataloader(dataset,
imgs_per_gpu=1,
workers_per_gpu=cfg.data.workers_per_gpu,
num_gpus=1,
dist=False,
shuffle=False)
if args.load_result:
outputs = mmcv.load(args.out)
else:
# test
outputs = single_test(model,
data_loader,
cfg,
args.show,
save_path=args.save_path)
if args.out:
if not args.load_result:
print('writing results to {}'.format(args.out))
mmcv.dump(outputs, args.out)
eval_types = args.eval
if eval_types:
print('Starting evaluate {}'.format(' and '.join(eval_types)))
if not isinstance(outputs[0], dict):
result_file = args.out + '.json'
results2json_videoseg(dataset, outputs, result_file)
ytvos_eval(result_file, eval_types, dataset.ytvos)
else:
NotImplemented
def main():
args = parse_args()
assert args.out or args.eval or args.format_only or args.show, \
('Please specify at least one operation (save/eval/format/show the '
'results) with the argument "--out", "--eval", "--format_only" '
'or "--show"')
if args.eval and args.format_only:
raise ValueError('--eval and --format_only cannot be both specified')
if args.out is not None and not args.out.endswith(('.pkl', '.pickle')):
raise ValueError('The output file must be a pkl file.')
cfg = mmcv.Config.fromfile(args.config)
# set cudnn_benchmark
if cfg.get('cudnn_benchmark', False):
torch.backends.cudnn.benchmark = True
cfg.model.pretrained = None
cfg.data.test.test_mode = True
vid_idx_list = args.vid_idx
cfg.data.test.vid_idx_list = vid_idx_list
# init distributed env first, since logger depends on the dist info.
if args.launcher == 'none':
distributed = False
else:
distributed = True
init_dist(args.launcher, **cfg.dist_params)
dataset = obj_from_dict(cfg.data.test, datasets, dict(test_mode=True))
assert args.gpus == 1
model = build_detector(
cfg.model, train_cfg=None, test_cfg=cfg.test_cfg)
load_checkpoint(model, args.checkpoint)
model = MMDataParallel(model, device_ids=[0])
data_loader = build_dataloader(
dataset,
imgs_per_gpu=1,
workers_per_gpu=cfg.data.workers_per_gpu,
num_gpus=1,
dist=False,
shuffle=False)
if args.load_result:
outputs = mmcv.load(args.out)
else:
outputs = single_test(model, data_loader, args.show, save_path=args.save_path)
if args.out:
if not args.load_result:
print('writing results to {}'.format(args.out))
mmcv.dump(outputs, args.out)
eval_types = args.eval
if eval_types:
print('Starting evaluate {}'.format(' and '.join(eval_types)))
if not isinstance(outputs[0], dict):
result_file = args.out + '.json'
dataset = obj_from_dict(cfg.data.test_video, datasets, dict(test_mode=True, vid_idx_list = vid_idx_list))
results2json_videoseg(dataset, outputs, result_file)
ytvos_eval(result_file, eval_types, dataset.ytvos, vid_idx_list = vid_idx_list)
else:
NotImplemented
def main():
import os
os.environ['CUDA_LAUNCH_BLOCKING'] = "1"
os.environ["CUDA_VISIBLE_DEVICES"] = "1"
args = parse_args()
args.out = args.save_path + 'result_test.pkl'
if args.out is not None and not args.out.endswith(('.pkl', '.pickle')):
raise ValueError('The output file must be a pkl file.')
cfg = mmcv.Config.fromfile(args.config)
# set cudnn_benchmark
if cfg.get('cudnn_benchmark', False):
torch.backends.cudnn.benchmark = True
cfg.model.pretrained = None
cfg.data.test.test_mode = True
# get dataset
dataset = obj_from_dict(cfg.data.test, datasets, dict(test_mode=True))
# build model
assert args.gpus == 1
model = build_detector(
cfg.model, train_cfg=None, test_cfg=cfg.test_cfg)
load_checkpoint(model, args.checkpoint)
# model.load_flow()
resume = "/home/ubuntu/code/fengda/MaskTrackRCNN/pretrained_models/epoch_12.pth"
load_checkpoint(model, resume)
model = MMDataParallel(model, device_ids=[0])
# device = torch.device("cuda")
# model.to(device)
# from mmdet.models.decision_net.ddpg import ActorNet as Net
# rl_model = Net(512, 4, checkpoint="/home/ubuntu/code/fengda/MaskTrackRCNN/tools/output_ddpg/2020-03-11 21:12:05/epoch0--0.09926817708735841/policy_net.pkl")
# rl_model = rl_model.cuda()
rl_model = None
# args.load_result = "/home/ubuntu/datasets/YT-VIS/results/test-flow-trained-29-all/result_test.pkl"
Test_video = False
if not Test_video:
if args.load_result:
outputs = mmcv.load(args.out)
else:
data_loader = build_dataloader(
dataset,
imgs_per_gpu=1,
workers_per_gpu=cfg.data.workers_per_gpu,
num_gpus=1,
dist=False,
shuffle=False)
outputs,stats = single_test(model, data_loader, args, rl_model=rl_model, show=args.show, save_path=args.save_path)
if args.out:
if not args.load_result:
print('writing results to {}'.format(args.out))
mmcv.dump(outputs, args.out)
eval_types = args.eval
if eval_types:
print('Starting evaluate {}'.format(' and '.join(eval_types)))
if not isinstance(outputs[0], dict):
result_file = args.out + '.json'
results2json_videoseg(dataset, outputs, result_file)
ytvos_eval(result_file, eval_types, dataset.ytvos, show=True)
else:
NotImplemented
# # test
else:
import pandas as pd
from mmdet.models.decision_net.utils import modify_cfg, get_dataloader
from mmdet.datasets.utils import get_dataset
# val_videos = list(pd.read_csv('train.csv').video_name)
val_videos = list(pd.read_csv('val.csv').video_name)
# val_videos = ['0065b171f9', '01c76f0a82', '4083cfbe15']
from tqdm import tqdm
results = {}
for video_name in tqdm(val_videos):
# get data loader of the selected video
cfg_test = modify_cfg(cfg, video_name)
ann_file = cfg_test.ann_file
# self.dataset = obj_from_dict(self.cfg_test, datasets, dict(test_mode=True))
try:
dataset = obj_from_dict(cfg_test, datasets, dict(test_mode=True))
print('video name: {}.\t len of dataset:{}.'.format(video_name, len(dataset)))
data_loader = build_dataloader(dataset, imgs_per_gpu=1, workers_per_gpu=0, num_gpus=1, dist=False, shuffle=False)
outputs, stats = single_test(model, data_loader, args, args.show, save_path=os.path.join(args.save_path, video_name))
# print(np.mean(stats))
except:
print(traceback.print_exc())
continue
# break
# continue
if args.out:
if not args.load_result:
print('writing results to {}'.format(args.out))
mmcv.dump(outputs, args.out)
eval_types = args.eval
if eval_types:
print('Starting evaluate {}'.format(' and '.join(eval_types)))
if not isinstance(outputs[0], dict):
result_file = args.out + '.json'
results2json_videoseg(dataset, outputs, result_file)
stats = ytvos_eval(result_file, eval_types, dataset.ytvos, show=True)
results[video_name] = stats[0]
else:
NotImplemented
'''save dict into json file'''
import json
with open(os.path.join(args.save_path, 'test-flow-trained-29.json'), 'w') as json_file:
json.dump(results, json_file, ensure_ascii=False)
def test(model, rl_model, data_loader, args, epoch):
print("================= Testing.=================")
model.eval()
results = []
scales = []
stats = []
dataset = data_loader.dataset
prog_bar = mmcv.ProgressBar(len(dataset))
num = 0
device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")
last_full = dict()
scale_facotrs = [1, 1 / 2, 1 / 3, 1 / 4]
for i, data in enumerate(data_loader):
is_first = data['img_meta'][0].data[0][0]['is_first']
if is_first:
# Full resolution if it is the first frame of the video.
scale_facotr = 1
num = 1
def state_reset():
"""Reset Features for states."""
feat_self = get_self_feat(model, data['img'][0])
feat_diff = torch.zeros_like(feat_self)
feat_FAR = torch.tensor([0.]).cuda()
feat_history = torch.zeros([10]).cuda()
return [feat_diff, feat_self, feat_FAR, feat_history]
state = state_reset()
else:
action = rl_model(state).argmax()
scale_facotr = scale_facotrs[action]
if scale_facotr == 1:
with torch.no_grad():
# data = get_low_data(deepcopy(data), 1/4, size_divisor=32, is_train=False)
result = model(return_loss=False, rescale=True, **data)
last_full['img'] = data['img'][0]
last_full['feat_map_last'] = result[2]
last_full['feat_self_last'] = get_self_feat(model, data['img'][0])
else:
# resize data
low_data = get_low_data(deepcopy(data), scale_facotr, size_divisor=32, is_train=False)
with torch.no_grad():
result = model(return_loss=False, rescale=True, key_frame=last_full, **low_data)
def state_step(s, a):
feat_self = get_self_feat(model, data['img'][0])
feat_diff = get_diff_feat(last_full['feat_self_last'], feat_self)
if a == 1:
feat_FAR = torch.tensor([0.]).cuda()
else:
feat_FAR = s[2] + 0.05
if feat_FAR > 1:
feat_FAR = torch.tensor([1.]).cuda()
feat_history = torch.zeros([10]).cuda()
return [feat_diff, feat_self, feat_FAR, feat_history]
state = state_step(state, scale_facotr)
scales.append(scale_facotr)
result = result[0:2]
results.append(result)
batch_size = data['img'][0].size(0)
for _ in range(batch_size):
prog_bar.update()
eval_types = ['segm']
result_file = 'results.json'
results2json_videoseg(dataset, results, result_file)
stat = ytvos_eval(result_file, eval_types, dataset.ytvos, show=False)
ec = [2.486862653249294, 0.8707167229351835, 0.7147069263013582, 0.6858342799104453, 0.6757044771429019]
print("Count for 1 : ", scales.count(1))
print("Count for 1/2 : ", scales.count(1/2))
print("Count for 1/3 : ", scales.count(1/3))
print("Count for 1/4 : ", scales.count(1/4))
e = scales.count(1)*ec[1] + scales.count(1/2)*ec[2] + scales.count(1/3)*ec[3] + scales.count(1/4)*ec[4]
print("Energy Consumption:{}".format(e))
# Save model
if args.save_model:
# now = time.strftime("%Y-%m-%d %X",time.localtime())
dir = os.path.join(args.output_folder, "epoch%d-EC%.3f-MAP%.3f" %(epoch, e, stat[0]))
if not os.path.exists(dir):
os.makedirs(dir)
agent.save_model(dir)
return results, stats
