def main():
parser = argparse.ArgumentParser()
parser.add_argument("--yml_cfg_file", type=str, default="./cfgs/bsn.yml")
args = parser.parse_args()
cfg = load_yml(args.yml_cfg_file)
anet = cfg.anet
train_dataset = get_pgm_feature_dataset(cfg.tem.tem_csv_dir, cfg.pgm.proposal_csv_path, anet.json_path,
anet.video_info_new_csv_path, "training")
val_dataset = get_pgm_feature_dataset(cfg.tem.tem_csv_dir, cfg.pgm.proposal_csv_path, anet.json_path,
anet.video_info_new_csv_path, "validation")
queue = mp.Queue()
procs = []
mkdir_p(cfg.pgm.pgm_feature_path)
for i in range(len(train_dataset)):
queue.put(('training', i))
for i in range(len(val_dataset)):
queue.put(('validation', i))
for i in range(cfg.pgm.pgm_feature_workers):
proc = mp.Process(target=proc_cb, args=(queue, cfg, {"training": train_dataset, "validation": val_dataset}))
procs.append(proc)
proc.start()
t = 0
while not queue.empty():
log.log_info('{}s: Remain {} videos to be handled.'.format(t, queue.qsize()))
time.sleep(1)
t += 1
log.log_warn("All video processed.")
for proc in procs:
proc.terminate()
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--yml_cfg_file", default="./cfgs/bsn.yml")
args = parser.parse_args()
cfg = load_yml(args.yml_cfg_file)
trainer = PemTrainer(cfg, cfg.pem)
trainer.train()
def main():
# 解析参数
parser = argparse.ArgumentParser()
parser.add_argument("--yml_cfg_file", type=str)
args = parser.parse_args()
cfg = load_yml(args.yml_cfg_file)
# 加载模型
tem_model = Tem(cfg.input_features)
model_info_path = os.path.join(cfg.save_root, 'model_model_info.txt')
with open(model_info_path) as f:
model_path = f.readline().strip()
epoch = int(f.readline().strip())
state_dicts = torch.load(model_path)
tem_model.load_state_dict(state_dicts['models'][0])
# 构建数据集
anet = cfg.anet_dataset
val_dataset = ActivityNetDataset.get_ltw_feature_dataset(
anet.csv_path, anet.json_path, anet.video_info_new_csv_path,
anet.class_name_path, 'validation')
train_dataset = ActivityNetDataset.get_ltw_feature_dataset(
anet.csv_path, anet.json_path, anet.video_info_new_csv_path,
anet.class_name_path, 'training')
val_dataloader = DataLoader(val_dataset,
batch_size=cfg.batch_size,
shuffle=False,
num_workers=16)
train_dataloader = DataLoader(train_dataset,
batch_size=cfg.batch_size,
shuffle=False,
num_workers=16)
tem_model.cuda()
tem_model.eval()
log.log(log.INFO, "Start processing train dataset.")
results = []
for i, (features, _) in enumerate(val_dataloader):
features = features.cuda()
result = tem_model(features)
results.append(result.cpu().detach().numpy())
val_results = np.concatenate(results)
log.log(log.INFO, "Start processing validation dataset.")
results = []
for i, (features, _) in enumerate(train_dataloader):
features = features.cuda()
result = tem_model(features)
results.append(result.cpu().detach().numpy())
train_results = np.concatenate(results)
mkdir_p(os.path.split(cfg.tem_results_file)[0])
log.log(log.INFO, "Dumping to {}".format(cfg.tem_results_file))
with open(cfg.tem_results_file, 'wb') as f:
results = {"training": train_results, "validation": val_results}
pickle.dump(results, f)
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--yml_cfg_file", type=str, default="./cfgs/bsn.yml")
args = parser.parse_args()
cfg = load_yml(args.yml_cfg_file)
anet = cfg.anet
train_dataset = get_tem_dataset(cfg.tem.tem_csv_dir, anet.json_path,
anet.video_info_new_csv_path, "training")
val_dataset = get_tem_dataset(cfg.tem.tem_csv_dir, anet.json_path,
anet.video_info_new_csv_path, "validation")
mkdir_p(cfg.pgm.proposal_csv_path)
# prepare workers
queue = mp.Queue()
procs = []
# feed datas
for i in range(len(train_dataset)):
queue.put(('training', i))
for i in range(len(val_dataset)):
queue.put(('validation', i))
for i in range(cfg.pgm.proposal_workers):
proc = mp.Process(target=sub_proc,
args=(queue, cfg, {
"training": train_dataset,
"validation": val_dataset
}))
proc.start()
procs.append(proc)
t = 0
while not queue.empty():
remain = queue.qsize()
log.log_info("Time: {}s, remain {} videos to be handled.".format(
t, remain))
time.sleep(1)
t += 1
log.log_info("All videos handled.")
for proc in procs:
proc.terminate()
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--yml_cfg_file", default="./cfgs/bsn.yml")
args = parser.parse_args()
cfg = load_yml(args.yml_cfg_file)
anet = cfg.anet
val_dataset = get_post_processing_dataset(cfg.pem.pem_csv_dir,
anet.json_path,
anet.video_info_new_csv_path,
'validation')
queue = mp.Queue()
mp_dict = mp.Manager().dict()
# for i in range(100):
for i in range(len(val_dataset)):
queue.put(i)
procs = []
for i in range(cfg.post_processing.pp_workers):
proc = mp.Process(target=sub_proc,
args=(queue, mp_dict, val_dataset, cfg))
procs.append(proc)
proc.start()
for proc in procs:
proc.join()
output_dict = {
"version": "VERSION 1.3",
"results": dict(mp_dict),
"external_data": {}
}
log.log_info(len(mp_dict.keys()))
mkdir_p(os.path.split(cfg.eval.results_json)[0])
with open(cfg.eval.results_json, "w") as f:
json.dump(output_dict, f)
log.log_info("Dump results to {}.".format(cfg.eval.results_json))
plt.setp(plt.axes().get_yticklabels(), fontsize=fn_size)
mkdir_p(os.path.split(fig_save_path)[0])
plt.savefig(fig_save_path)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument(
"--gt_cfg_file",
default=
"/home/dechao_meng/datasets/activitynet/annotations/activity_net_1_3_new.json"
)
parser.add_argument(
"--yml_cfg_file",
default=
"/tmp/pycharm_project_580/action_proposals/trainer/bsn_trainer/cfgs/bsn.yml"
)
args = parser.parse_args()
cfg = load_yml(args.yml_cfg_file)
fig_save_path = cfg.eval.results_json[:-5] + time.asctime() + '.png'
uniform_average_nr_proposals_valid, uniform_average_recall_valid, uniform_recall_valid = run_evaluation(
args.gt_cfg_file,
cfg.eval.results_json,
max_avg_nr_proposals=100,
tiou_thresholds=np.linspace(0.5, 0.95, 10),
subset='validation')
plot_metric(uniform_average_nr_proposals_valid,
uniform_average_recall_valid, uniform_recall_valid,
fig_save_path)