baseline的骨骼检测流程记录

0.背景

---

2018年coco挑战赛亚军，pipeline非常简洁，VGG16+额外几层反卷积分，输出关节点热力图。主要了解pipeline构建过程，学习一下代码怎么写。https://github.com/microsoft/human-pose-estimation.pytorch

1.相关知识

---

1.1 关键点数量

17个关键点，具体如下。所以网络层最终输出  B* 17* Heatmap_H* Heatmap_W

    '''
    "keypoints": {
        0: "nose",
        1: "left_eye",
        2: "right_eye",
        3: "left_ear",
        4: "right_ear",
        5: "left_shoulder",
        6: "right_shoulder",
        7: "left_elbow",
        8: "right_elbow",
        9: "left_wrist",
        10: "right_wrist",
        11: "left_hip",
        12: "right_hip",
        13: "left_knee",
        14: "right_knee",
        15: "left_ankle",
        16: "right_ankle"
    },
	"skeleton": [
        [16,14],[14,12],[17,15],[15,13],[12,13],[6,12],[7,13], [6,7],[6,8],
        [7,9],[8,10],[9,11],[2,3],[1,2],[1,3],[2,4],[3,5],[4,6],[5,7]]
    '''

1.2  网络结构

        self.features = nn.Sequential(*features)
        #原先是VGG16，224-->7.1/32，更换为MobilenetV2，  7*7*1280

        self.deconv_layers = self._make_deconv_layer(
            extra.NUM_DECONV_LAYERS,
            extra.NUM_DECONV_FILTERS,
            extra.NUM_DECONV_KERNELS,
        )   #1/32-->1/4   channal=256

        self.final_layer = nn.Conv2d(
            in_channels=extra.NUM_DECONV_FILTERS[-1],
            out_channels=cfg.MODEL.NUM_JOINTS,
            kernel_size=extra.FINAL_CONV_KERNEL,
            stride=1,
            padding=1 if extra.FINAL_CONV_KERNEL == 3 else 0
        )  #1/4,channal=256-->17.    默认输入3*256*192，输出17*64*48

部分初始化参数

    def init_weights(self, pretrained=''):
        if os.path.isfile(pretrained):
            logger.info('=> init deconv weights from normal distribution')
            for name, m in self.deconv_layers.named_modules():
                if isinstance(m, nn.ConvTranspose2d):
                    logger.info('=> init {}.weight as normal(0, 0.001)'.format(name))
                    logger.info('=> init {}.bias as 0'.format(name))
                    nn.init.normal_(m.weight, std=0.001)
                    if self.deconv_with_bias:
                        nn.init.constant_(m.bias, 0)
                elif isinstance(m, nn.BatchNorm2d):
                    logger.info('=> init {}.weight as 1'.format(name))
                    logger.info('=> init {}.bias as 0'.format(name))
                    nn.init.constant_(m.weight, 1)
                    nn.init.constant_(m.bias, 0)
            logger.info('=> init final conv weights from normal distribution')
            for m in self.final_layer.modules():
                if isinstance(m, nn.Conv2d):
                    # nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
                    logger.info('=> init {}.weight as normal(0, 0.001)'.format(name))
                    logger.info('=> init {}.bias as 0'.format(name))
                    nn.init.normal_(m.weight, std=0.001)
                    nn.init.constant_(m.bias, 0)

            # pretrained_state_dict = torch.load(pretrained)
            logger.info('=> loading pretrained model {}'.format(pretrained))
            # self.load_state_dict(pretrained_state_dict, strict=False)
            checkpoint = torch.load(pretrained)
            if isinstance(checkpoint, OrderedDict):
                state_dict = checkpoint
            elif isinstance(checkpoint, dict) and 'state_dict' in checkpoint:
                state_dict_old = checkpoint['state_dict']
                state_dict = OrderedDict()
                # delete 'module.' because it is saved from DataParallel module
                for key in state_dict_old.keys():
                    if key.startswith('module.'):
                        # state_dict[key[7:]] = state_dict[key]
                        # state_dict.pop(key)，原先权重可能是在多卡下存储的，或多一个 
                        # modele.。解决办法两种，一种保存模型时，model.module.dict()
                        # 一种 model =     \    
                        #  torch.nn.DataParallel(model,device_ids=gpus).cuda()
                        # 再加载  model.load
                        model.load_state_dict(checkpoint)  # mudule 类自带的函数,strict=False               
                        state_dict[key[7:]] = state_dict_old[key]
                    else:
                        state_dict[key] = state_dict_old[key]
            else:
                raise RuntimeError(
                    'No state_dict found in checkpoint file {}'.format(pretrained))
            self.load_state_dict(state_dict, strict=False)  #mudule 类自带的函数
        else:
            logger.error('=> imagenet pretrained model dose not exist')
            logger.error('=> please download it first')
            raise ValueError('imagenet pretrained model does not exist')

1.3 文件夹布局

重要的几个

--core
    --__init__.py
    --config.py  #用easydict储存配置信息
    --evaluate.py #计算PCK0.5，得出各关节点在一个Batch的ACC
    --function.py  ##定义 train（）和 val（）
    --inference.py ##get_max_preds（），输入 b*64*32， 输出b*17*2（峰值坐标,有的为0），b*17*1（score）
    --loss.py ##定义loss类，forward中实现 L2loss，输出(1,)
--dataset
    --__init__.py
    --coco.py ##自定义数据类coco，继承jioint类,加载GT在自身中。还包含evalution函数计算AP，最重要
    --Jioint.py ##这里定义 __getitem__
--model
    --pose.py ##定义网络
--utils.py
    --transform.py ##定义 flip、仿射变换等数据增强
    --vis.py ##展示图片
    --utils.py ##定义logger，optimizer等操作

2.数据加载

2.1 coco类介绍

info、image、license  共享     annotion滑头最多

images数组和annotations数组的元素数量是不相等的，annotion数量多于图片，每个图片里的每一个对象有一个自己的id，且有对应image对应的image_id.      catergories里只有一个人类。

annotion：

2.2 自定义coco数据类

self.image_set_index = self._load_image_set_index()  #[1122,1212,12121,...]  int array，存有train or val所有的图片id

self.db = self._get_db()  #[{},{}],keypoint  x,y,vision,  center，一个{}就是一个人的信息，很全面


    def __getitem__(self, idx):
        db_rec = copy.deepcopy(self.db[idx]) #{}

        image_file = db_rec['image'] #'xx/.jpg'
        filename = db_rec['filename'] if 'filename' in db_rec else ''
        imgnum = db_rec['imgnum'] if 'imgnum' in db_rec else ''  #0

        if self.data_format == 'zip':
            from utils import zipreader
            data_numpy = zipreader.imread(
                image_file, cv2.IMREAD_COLOR | cv2.IMREAD_IGNORE_ORIENTATION)
        else:
            data_numpy = cv2.imread(
                image_file, cv2.IMREAD_COLOR | cv2.IMREAD_IGNORE_ORIENTATION)

        if data_numpy is None:
            logger.error('=> fail to read {}'.format(image_file))
            raise ValueError('Fail to read {}'.format(image_file))

        joints = db_rec['joints_3d'] #[x,y,0] 
        joints_vis = db_rec['joints_3d_vis'] #[0,0,0] or [1,1,0]

        c = db_rec['center'] #[111,222]
        s = db_rec['scale'] #[1.1,2.3],people's h,w/200
        score = db_rec['score'] if 'score' in db_rec else 1 #1
        r = 0

        if self.is_train:
            sf = self.scale_factor #0.3
            rf = self.rotation_factor #40
            s = s * np.clip(np.random.randn()*sf + 1, 1 - sf, 1 + sf)
            r = np.clip(np.random.randn()*rf, -rf*2, rf*2) \
                if random.random() <= 0.6 else 0

            if self.flip and random.random() <= 0.5:
                data_numpy = data_numpy[:, ::-1, :]  #hwc,w
                joints, joints_vis = fliplr_joints(
                    joints, joints_vis, data_numpy.shape[1], self.flip_pairs)  #17*3,17*3
                c[0] = data_numpy.shape[1] - c[0] - 1

        trans = get_affine_transform(c, s, r, self.image_size) ##return a matrix,2*3
        input = cv2.warpAffine(
            data_numpy,
            trans,
            (int(self.image_size[0]), int(self.image_size[1])),
            flags=cv2.INTER_LINEAR)  #仿射变换

        if self.transform:
            input = self.transform(input)  ##normaize

        for i in range(self.num_joints):
            if joints_vis[i, 0] > 0.0:
                joints[i, 0:2] = affine_transform(joints[i, 0:2], trans)

        target, target_weight = self.generate_target(joints, joints_vis) 
        # targrt = 17*heatmap， target_weight = 17*1，关键点在边缘或者没有的 ->0

        target = torch.from_numpy(target)
        target_weight = torch.from_numpy(target_weight)

        meta = {
            'image': image_file,
            'filename': filename,
            'imgnum': imgnum,
            'joints': joints, #(x,y,0) or (0,0,0)
            'joints_vis': joints_vis, #(1,1,0) or (0,0,0)
            'center': c,
            'scale': s,
            'rotation': r,
            'score': score
        }

        return input, target, target_weight, meta

3.train流程

    model = eval('models.'+'pose_mobilenetv2'+'.get_pose_net2')(
        config, is_train=False
    )   # 调用get_pose_net，参数在后面，返回model,output  B*() *16

    checkpoint  = torch.load('0.633-model_best.pth.tar')
    
    #model.load_state_dict(checkpoint['state_dict'])  #mudule 类自带的函数,strict=False
    #writer = SummaryWriter(log_dir='tensorboard_logs')

    gpus = [int(i) for i in config.GPUS.split(',')]  # [0]
    model = torch.nn.DataParallel(model, device_ids=gpus).cuda()
    model.load_state_dict(checkpoint)  # mudule 类自带的函数,strict=False

    # define loss function (criterion) and optimizer
    criterion = JointsMSELoss(
        use_target_weight=config.LOSS.USE_TARGET_WEIGHT
    ).cuda()  #  call(output, target, target_weight)

    optimizer = get_optimizer(config, model) #LR=0.001

    lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer,15,0.000001,-1)
    # Data loading code
    normalize = transforms.Normalize(mean=[0.485, 0.456, 0.406],
                                     std=[0.229, 0.224, 0.225])
    train_dataset = eval('dataset.'+config.DATASET.DATASET)(
        config,
        config.DATASET.ROOT,
        config.DATASET.TRAIN_SET,
        True,
        transforms.Compose([
            transforms.ToTensor(),
            normalize,
        ])
    )  # COCO,include img_id[].gt[]

    train_loader = torch.utils.data.DataLoader(
        train_dataset,
        batch_size=config.TRAIN.BATCH_SIZE*len(gpus),
        shuffle=config.TRAIN.SHUFFLE,
        num_workers=config.WORKERS,
        pin_memory=True
    )

    best_perf = 0.0
    best_model = False
    for epoch in range(150, 220):   #config.TRAIN.BEGIN_EPOCH, config.TRAIN.END_EPOCH
        lr_scheduler.step()

        # train for one epoch
        train(config, train_loader, model, criterion, optimizer, epoch,
              final_output_dir, tb_log_dir)


        # evaluate on validation set，return AP
        perf_indicator = validate(config, valid_loader, valid_dataset, model,
                                  criterion, final_output_dir, tb_log_dir)
            
        if perf_indicator > best_perf:
            best_perf = perf_indicator
            best_model = True
        else:
            best_model = False

        logger.info('=> saving checkpoint to {}'.format(final_output_dir))
        save_checkpoint({
            'epoch': epoch + 1,
            'model': get_model_name(config),
            'state_dict': model.state_dict(),
            'perf': perf_indicator,
            'optimizer': optimizer.state_dict(),
        }, best_model, final_output_dir)

    final_model_state_file = os.path.join(final_output_dir,
                                          'final_state.pkl')
    logger.info('saving final model state to {}'.format(
        final_model_state_file))
    torch.save(model.module.state_dict(), final_model_state_file)

3.1 train函数

def train(config, train_loader, model, criterion, optimizer, epoch,
          output_dir, tb_log_dir, writer_dict=None):
    batch_time = AverageMeter()
    data_time = AverageMeter()
    losses = AverageMeter()
    acc = AverageMeter()

    # switch to train mode
    model.train()

    end = time.time()
    for i, (input, target, target_weight, meta) in enumerate(train_loader):
        # measure data loading time
        data_time.update(time.time() - end)

        # compute output
        output = model(input)  #17*h*w
        target = target.cuda(non_blocking=True)  #17*4*h,0-1
        target_weight = target_weight.cuda(non_blocking=True) #17*1,0 or 1

        loss = criterion(output, target, target_weight) # 1*1

        # compute gradient and do update step
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

        # measure accuracy and record loss
        losses.update(loss.item(), input.size(0))

        _, avg_acc, cnt, pred = accuracy(output.detach().cpu().numpy(),
                                         target.detach().cpu().numpy())  #[18,],scalr(与gt比较计算pck0.5),cnt=16,b*17*2(top点坐标，有部分为0,因为峰值低于0)
        acc.update(avg_acc, cnt)

        # measure elapsed time
        batch_time.update(time.time() - end)
        end = time.time()

        if i % config.PRINT_FREQ == 0:
            msg = 'Epoch: [{0}][{1}/{2}]\t' \
                  'Time {batch_time.val:.3f}s ({batch_time.avg:.3f}s)\t' \
                  'Speed {speed:.1f} samples/s\t' \
                  'Data {data_time.val:.3f}s ({data_time.avg:.3f}s)\t' \
                  'Loss {loss.val:.5f} ({loss.avg:.5f})\t' \
                  'Accuracy {acc.val:.3f} ({acc.avg:.3f})'.format(
                      epoch, i, len(train_loader), batch_time=batch_time,
                      speed=input.size(0)/batch_time.val,
                      data_time=data_time, loss=losses, acc=acc)
            logger.info(msg)
            '''
            writer = writer_dict['writer']
            global_steps = writer_dict['train_global_steps']
            writer.add_scalar('train_loss', losses.val, global_steps)
            writer.add_scalar('train_acc', acc.val, global_steps)
            writer_dict['train_global_steps'] = global_steps + 1
            '''
            prefix = '{}_{}'.format(os.path.join(output_dir, 'train'), i)
            save_debug_images(config, input, meta, target, pred*4, output,
                              prefix)

3.2 计算 pck0.5（按关节点种类来的）

def calc_dists(preds, target, normalize): #b*17*2,
    preds = preds.astype(np.float32)
    target = target.astype(np.float32)
    dists = np.zeros((preds.shape[1], preds.shape[0]))
    for n in range(preds.shape[0]):
        for c in range(preds.shape[1]):
            if target[n, c, 0] > 1 and target[n, c, 1] > 1: #ingore 边缘关键点
                normed_preds = preds[n, c, :] / normalize[n] #/6.4，/4.8
                normed_targets = target[n, c, :] / normalize[n]
                dists[c, n] = np.linalg.norm(normed_preds - normed_targets) #better is 0,>0，L2距离
            else:
                dists[c, n] = -1
    return dists #17*b


def dist_acc(dists, thr=0.5): # 1*b
    ''' Return percentage below threshold while ignoring values with a -1 '''
    dist_cal = np.not_equal(dists, -1)
    num_dist_cal = dist_cal.sum()
    if num_dist_cal > 0:
        return np.less(dists[dist_cal], thr).sum() * 1.0 / num_dist_cal  #距离小于0.5，pck0.5
    else:
        return -1


def accuracy(output, target, hm_type='gaussian', thr=0.5):
    '''
    Calculate accuracy according to PCK,
    but uses ground truth heatmap rather than x,y locations
    First value to be returned is average accuracy across 'idxs',
    followed by individual accuracies
    '''
    idx = list(range(output.shape[1])) #17
    norm = 1.0
    if hm_type == 'gaussian':
        pred, _ = get_max_preds(output) #b*17*2,max坐标，部分为0
        target, _ = get_max_preds(target)
        h = output.shape[2]
        w = output.shape[3]
        norm = np.ones((pred.shape[0], 2)) * np.array([h, w]) / 10 #6.4，4.8
    dists = calc_dists(pred, target, norm) #17*b

    acc = np.zeros((len(idx) + 1))
    avg_acc = 0
    cnt = 0

    for i in range(len(idx)):
        acc[i + 1] = dist_acc(dists[idx[i]])
        if acc[i + 1] >= 0:
            avg_acc = avg_acc + acc[i + 1]
            cnt += 1

    avg_acc = avg_acc / cnt if cnt != 0 else 0
    if cnt != 0:
        acc[0] = avg_acc
    return acc, avg_acc, cnt, pred

4 总结

可视化代码：

def save_batch_image_with_joints(batch_image, batch_joints, batch_joints_vis,
                                 file_name, nrow=8, padding=2):
    '''
    batch_image: [batch_size, channel, height, width]
    batch_joints: [batch_size, num_joints, 3],
    batch_joints_vis: [batch_size, num_joints, 1],
    }
    一行8个
    '''
    grid = torchvision.utils.make_grid(batch_image, nrow, padding, True)
    ndarr = grid.mul(255).clamp(0, 255).byte().permute(1, 2, 0).cpu().numpy() #？均值方差不用动？
    ndarr = ndarr.copy()

    nmaps = batch_image.size(0)
    xmaps = min(nrow, nmaps)
    ymaps = int(math.ceil(float(nmaps) / xmaps))
    height = int(batch_image.size(2) + padding)
    width = int(batch_image.size(3) + padding)
    k = 0
    for y in range(ymaps):
        for x in range(xmaps):
            if k >= nmaps:
                break
            joints = batch_joints[k]
            joints_vis = batch_joints_vis[k]

            for joint, joint_vis in zip(joints, joints_vis):
                joint[0] = x * width + padding + joint[0]
                joint[1] = y * height + padding + joint[1]
                if joint_vis[0]:
                    cv2.circle(ndarr, (int(joint[0]), int(joint[1])), 2, [255, 0, 0], 2)
            k = k + 1
    cv2.imwrite(file_name, ndarr)

收获

1.使用  logger打印信息

2.部分加载模型权重

3.自定义一个复杂数据类

4.使用tensorboardX查看训练过程

5.model得到输出后loss在GPU上计算，计算ACC转化到 cpu（）numpy（）上进行后续操作