ours_synthetic_ap_10k.py

import datetime
import logging
import argparse
import os
import random
import json
import numpy as np
import torch
import pprint
import shutil
from tensorboardX import SummaryWriter
from torch.utils.data import DataLoader,SubsetRandomSampler
from train_utils.utils import get_cosine_schedule_with_warmup
from models.hrnet import HighResolutionNet
from train_utils import transforms
from train_utils.dataset import CocoKeypoint
from train_utils.ssl_utils import ours_ap10k
from outer_tools.lib.config import cfg,update_config
from outer_tools.lib.utils.utils import create_logger
from torch.backends import cudnn as cudnn


def set_seed(args):
    random.seed(args.seed)
    np.random.seed(args.seed)
    torch.manual_seed(args.seed)
    torch.cuda.manual_seed_all(args.seed)


def main(cfg,args):
    args_dict = vars(args)
    logger.info('Args: %s', args_dict)

    if args.seed is not None:
        set_seed(args)
    #

    with open(args.keypoints_path, "r") as f:
        animal_kps_info = json.load(f)
    fixed_size = args.fixed_size
    heatmap_hw = (args.fixed_size[0] // 4, args.fixed_size[1] // 4)
    kps_weights = np.array(animal_kps_info["kps_weights"],
                           dtype=np.float32).reshape((args.num_joints,))

    data_transform = {
        "train": transforms.Compose([
            transforms.TransformMPL(args, mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
        ]),
        "val": transforms.Compose([
            transforms.AffineTransform(scale=None, rotation=None, fixed_size=fixed_size),
            transforms.KeypointToHeatMap(heatmap_hw=heatmap_hw, gaussian_sigma=2, keypoints_weights=kps_weights),
            transforms.ToTensor(),
            transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
        ])
    }

    data_root = args.data_root
    train_label_dataset = CocoKeypoint(root=data_root, dataset="ap_10k",mode="train",transform=data_transform["train"],
                                       fixed_size=args.fixed_size, data_type="keypoints")
    train_unlabel_dataset = CocoKeypoint(root=data_root, dataset="ap_10k",mode="train",transform=data_transform["train"],
                                         fixed_size=args.fixed_size, data_type="keypoints")

    label_img_id_path = "info/label_list/annotation_list_25.json"
    with open(label_img_id_path, 'r') as f:
        img_ids = json.load(f)

    train_label_dataset.valid_list = [ann for ann in train_label_dataset.valid_list if ann['image_id'] in img_ids]
    train_label_dataset.load_missing_anns(args.anns_info_path)
    train_label_dataset.get_kps_num(args)
    train_unlabel_dataset.get_kps_num(args)

    batch_size = args.batch_size
    nw = args.workers  # number of workers
    logger.info('Using %g dataloader workers' % nw)
    #
    base_weight_path = args.pretrained_model_path
    tea_weight_name = args.pretrained_weights_name
    stu_weight_name = 'pretrained_ori.pth'
    tea_pretrained_weights_path = os.path.join(base_weight_path,tea_weight_name)
    stu_pretrained_weights_path = os.path.join(base_weight_path,stu_weight_name)
    t_model = HighResolutionNet(num_joints=args.num_joints)
    s_model = HighResolutionNet(num_joints=args.num_joints)

    stu_checkpoint = torch.load(stu_pretrained_weights_path)
    tea_checkpoint = torch.load(tea_pretrained_weights_path)
    t_model.load_state_dict(tea_checkpoint)
    s_model.load_state_dict(stu_checkpoint,strict=False)

    logger.info(f"teacher model loaded from {tea_pretrained_weights_path}")
    logger.info(f"student model loaded from {stu_pretrained_weights_path}")

    t_model = torch.nn.DataParallel(t_model,device_ids=args.gpus).cuda()
    s_model = torch.nn.DataParallel(s_model,device_ids=args.gpus).cuda()
    t_model.train()
    s_model.train()

    train_label_loader = DataLoader(train_label_dataset,
                                    batch_size=batch_size,
                                    sampler=SubsetRandomSampler(range(len(train_label_dataset))),
                                    pin_memory=True,
                                    num_workers=nw,
                                    drop_last=False,
                                    collate_fn=train_label_dataset.collate_fn_mpl)
    train_unlabel_loader = DataLoader(train_unlabel_dataset,
                                      batch_size=batch_size * args.mu,
                                      sampler=SubsetRandomSampler(range(len(train_unlabel_dataset))),
                                      pin_memory=True,
                                      num_workers=nw,
                                      drop_last=False,
                                      collate_fn=train_unlabel_dataset.collate_fn_mpl)

    t_params = [p for p in t_model.parameters() if p.requires_grad]
    s_params = [p for p in s_model.parameters() if p.requires_grad]

    t_optimizer = torch.optim.AdamW(t_params,lr=args.teacher_lr,weight_decay=args.weight_decay)
    s_optimizer = torch.optim.AdamW(s_params,lr=args.student_lr,weight_decay=args.weight_decay)

    t_scaler = torch.cuda.amp.GradScaler() if args.amp else None
    s_scaler = torch.cuda.amp.GradScaler() if args.amp else None

    t_scheduler = get_cosine_schedule_with_warmup(t_optimizer,
                                                  args.warmup_steps,
                                                  args.total_steps)
    s_scheduler = get_cosine_schedule_with_warmup(s_optimizer,
                                                  args.warmup_steps,
                                                  args.total_steps,
                                                  args.student_wait_steps)

    if args.resume:
        if os.path.isfile(args.resume):
            logger.info("=> loading checkpoint {}".format(args.resume))
            checkpoint = torch.load(args.resume, map_location='cpu')
            args.start_step = checkpoint['step'] + 1
            t_model.module.load_state_dict(checkpoint['teacher_model'])
            s_model.module.load_state_dict(checkpoint['student_model'])
            t_optimizer.load_state_dict(checkpoint['teacher_optimizer'])
            s_optimizer.load_state_dict(checkpoint['student_optimizer'])
            t_scheduler.load_state_dict(checkpoint['teacher_scheduler'])
            s_scheduler.load_state_dict(checkpoint['student_scheduler'])
            if args.amp:
                t_scaler.load_state_dict(checkpoint['teacher_scaler'])
                s_scaler.load_state_dict(checkpoint['student_scaler'])
            logger.info("=> loaded checkpoint {} (epoch {})".format(args.resume,checkpoint['step']))
        else:
            logger.info("=> no checkpoint found at {}".format(args.resume))

    writer_dir = os.path.join(args.output_dir,"summary")
    if not os.path.exists(writer_dir):
        os.mkdir(writer_dir)
    writer_dict = {
        'writer': SummaryWriter(log_dir=writer_dir),
        'train_global_steps': 0,
        'valid_global_steps': 0,
    }

    ours_ap10k(cfg,args,train_label_loader,train_unlabel_loader,t_model, s_model,t_optimizer,s_optimizer,
               t_scheduler, s_scheduler,t_scaler,s_scaler,writer_dict)

    writer_dict['writer'].close()

    logger.info("Best OKS:{} at Epoch{}".format(
        args.best_oks,args.best_oks_epoch
    ))
    logger.info("Best PCK:{} at Epoch{}".format(
        args.best_pck, args.best_pck_epoch
    ))

    old_name = args.output_dir
    new_name = f"./experiment/{args.name}_{args.info}_{os.path.basename(old_name)}"
    os.rename(old_name,new_name)

    return


if __name__ == "__main__":
    logger = logging.getLogger(__name__)
    parser = argparse.ArgumentParser(
        description=__doc__)
    parser.add_argument('--name', default="ours_for_25_5_imgs_settings", type=str, help='experiment name')
    parser.add_argument('--info', default="", type=str, help='experiment info')
    parser.add_argument('--gpus', default=[0,1], help='device')
    parser.add_argument('--data-root', default='../dataset/ap_10k', type=str, help='data path')
    parser.add_argument('--pretrained-model-path', default='./pretrained_weights',
                        type=str, help='pretrained weights base path')
    parser.add_argument('--pretrained-weights-name', default='25_5_imgs_SL_hrnet_pretrained.pth',
                        type=str, help='pretrained weights name')

    parser.add_argument('--anns-info-path', default='info/25_5_imgs_keypoints_anns_info.json',
                        type=str, help='missing anns info path')
    parser.add_argument('--output-dir', default='./experiment',type=str, help='output dir depends on the time')
    parser.add_argument('--resume', default=None, type=str, help='path to resume file')

    parser.add_argument('--total-steps', default=120000, type=int, help='number of total steps to run')
    parser.add_argument('--eval-step', default=5, type=int, help='number of eval steps to run')
    parser.add_argument('--start-step', default=0, type=int,help='manual epoch number (useful on restarts)')
    parser.add_argument('--warmup-steps', default=900, type=int, help='warmup steps')
    parser.add_argument('--student-wait-steps', default=0, type=int, help='warmup steps')

    parser.add_argument('--uda-steps', default=60000, type=int, help='warmup steps of lambda-u')
    parser.add_argument('--down-step', default=9000,type=int,help='warmup steps of conditional PL')

    parser.add_argument('--feedback-steps-start', default=3000, type=float, help='start steps of feedback')
    parser.add_argument('--feedback-steps-complete', default=6000, type=float, help='warmup steps of feedback')
    parser.add_argument('--feedback-weight', default=2, type=float, help='feedback scalar')

    parser.add_argument('--teacher_lr', default=1e-5, type=float,
                        help='initial learning rate, 1e-5 is the default value for training')
    parser.add_argument('--student_lr', default=1e-3, type=float,
                        help='initial learning rate, 1e-3 is the default value for training')

    parser.add_argument('--wd', '--weight-decay', default=1e-4, type=float,metavar='W',
                        help='weight decay (default: 1e-4)',dest='weight_decay')
    parser.add_argument('--grad-clip', default=1e9, type=float, help='gradient norm clipping')
    parser.add_argument('--workers', default=8, type=int, help='number of workers for DataLoader')
    parser.add_argument('--batch-size', default=4, type=int, help='batch size of label data')
    parser.add_argument('--mu', default=1, type=int, help='batch size factor of unlabel data ')
    parser.add_argument('--seed', default=2, type=int, help='seed for initializing training')
    parser.add_argument('--keypoints-path', default="./info/ap_10k_keypoints_format.json", type=str,
                        help='keypoints_format.json path')
    parser.add_argument('--fixed-size', default=[256, 256], nargs='+', type=int, help='input size')
    parser.add_argument('--num-joints', default=17, type=int, help='num_joints')
    parser.add_argument('--best-oks', default=0, type=float,help='best OKS performance during training')
    parser.add_argument('--best-oks-epoch', default=0, type=int,help='best OKS performance Epoch during training')
    parser.add_argument('--best-pck', default=0, type=float,help='best PCK performance during training')
    parser.add_argument('--best-pck-epoch', default=0, type=int,help='best PCK performance Epoch during training')

    parser.add_argument("--amp",default=True,action="store_true",
                        help="Use torch.cuda.amp for mixed precision training")
    parser.add_argument('--cfg',default='./outer_tools/experiments/ap10k/hrnet/w32_256x192_adam_lr1e-3_ap10k.yaml',
                        help='experiment configure file name',type=str)
    args = parser.parse_args()

    now = datetime.datetime.now()
    now_time = now.strftime("%Y-%m_%d_%H-%M-%S")
    args.output_dir = os.path.join(args.output_dir,now_time)
    output_dir = args.output_dir
    if not os.path.exists(output_dir):
        os.makedirs(output_dir)

    info_output_dir = os.path.join(output_dir,'info')
    if not os.path.exists(info_output_dir):
        os.mkdir(info_output_dir)
    results_output_dir = os.path.join(output_dir,'results')
    if not os.path.exists(results_output_dir):
        os.mkdir(results_output_dir)
    save_weights_output_dir = os.path.join(output_dir,'save_weights')
    if not os.path.exists(save_weights_output_dir):
        os.mkdir(save_weights_output_dir)

    source_file_path = 'ours_synthetic_ap_10k.py'
    target_file_path = os.path.join(output_dir, 'ours_synthetic_ap_10k.py')
    shutil.copy(source_file_path, target_file_path)
    print(f"file saved to: {target_file_path}")

    source_file_path = 'train_utils/ssl_utils.py'
    target_file_path = os.path.join(output_dir, 'ssl_utils.py')
    shutil.copy(source_file_path, target_file_path)
    print(f"file saved to: {target_file_path}")

    gpu_list = args.gpus
    str_list = [str(num) for num in gpu_list]
    gpus = ",".join(str_list)
    os.environ["CUDA_VISIBLE_DEVICE"] = gpus

    update_config(cfg, args)
    logger, final_output_dir = create_logger(
        cfg, args.cfg, 'test')

    logger.info(pprint.pformat(args))
    logger.info(cfg)

    # cudnn related setting
    cudnn.benchmark = cfg.CUDNN.BENCHMARK
    torch.backends.cudnn.deterministic = cfg.CUDNN.DETERMINISTIC
    torch.backends.cudnn.enabled = cfg.CUDNN.ENABLED

    main(cfg,args)