buptxyb666 / adaptivepose Goto Github PK

@buptxyb666

1. z['ap'] 虽然输出了，但是没有参与最终的计算， 这个ap更像是网络自己学的一个中间级的过程，这个ap没有监督 能学到像论文的示意么。
2. Resample2D的作用是啥，在flownet2中搜到了，但还是不太清楚， 是将不同位置的特征进行融合么。 self.gradient_mul是为了控制ap回传的梯度范围么，这个是经验值吗。
3. 我之前在centernet中加入oks 替换掉了原来的RegWeightedL1Loss_coco()， 你这是额外加了oks的loss，这没有重复学这个回归信息么， 如果为了加快收敛可不可以先RegWeightedL1Loss_coco()，再oks。我一直很好奇如果是自己的数据集，oks中的sigma一般怎么估计呀。 还有oks 的应该是以绝对位置作为计算吧， 我看代码里好像是相对中心点的偏移。
   有时间的解答一下， 谢谢！

Thanks for your work, I have a question,Do you mean "torch12 " is "pythoch1.12" ?Looking forward to your reply and answer

Hi, wheather estimate the three-dimensional coordinate information point of each person?

您好，请问通过bash main_hrnet48_coco640.sh，一开始训练就报错FileNotFoundError: [Errno 2] No such file or directory: '/home/aa/AdaptivePose-master/src/lib/../../exp/multi_pose_wodet/hrnet_48_coco640/model_last.pth'是什么原因，这不应该是在训练一轮后生成的么？训练的时候已经吧main_hrnet48_coco640.sh中的测试部分注释掉了。

Would author transplant code to the frame of mmpose ? So many start-of-the-art methods are published on mmpose~

当运行bash main_dla34_coco512.sh的时候，发生错误：test.py: error: unrecognized arguments: multi_pose_wodet
（我把训练部分删除了，只保留了test.py部分）

We found about 0.3-0.5 AP noise during training process. We think the user can decay the learning rate of the DCN in regression head to 1/10 for stabilizing the training.

你好，按source prepare_enve.sh进行安装的。运行完后，显示如下图

执行main_dla34_coco512.sh,报NameError: name 'resample2d_cuda' is not defined.

请问这是什么问题啊

当时使用torch==1.2.0和torchvision==0.4.0版本进行环境配置时，运行训练代码报错提示：CUDA error: CUBLAS_STATUS_EXECUTION_FAILED，猜测是因为CUDA与torch的版本对应不上；

但当我安装与CUDA11.6相对应的torch版本后，编译安装resample2d_cuda，又出现报错提示：undefined symbol: _ZN6caffe26detail37_typeMetaDataInstance_preallocated_32E

请问该如何解决？

Run the demo using command with model pose_hrnet_w48_384x288.pth:

python demo.py multi_pose_wodet \
    --gpus 0 \
    --exp_id adp_hrnet_48 \
    --dataset coco_hp_wodet \
    --not_reg_offset \
    --not_reg_hp_offset \
    --K 20 \
    --not_hm_hp \
    --arch hrnet_48 \
    --input_h 1920 \
    --input_w 1080 \
    --keep_res \
    --debug 1 \
    --demo ../data/video1/video1.mp4 \
    --vis_thresh 0.1

and get results with bad pose estimation.

Input and output videos can be found here: BaiduDisk with code: hogg

Questions:

• How to improve the performance of pose estimation ?
• How to use model hrnet_48_crowdpose640.pth ?
  • I tried to add PRETRAINED: 'hrnet_48_crowdpose640.pth' in hrnet48.yaml, but it did not work.

hi, can you provide the pretrain_models of CrowdPose data, I train the model by pose_hrnet_w32_384x288.pth and the outcome is not great ,thank you!

Can you publish a py to onnx

你好 ， 我运行demo.py的时候，Coco的数据格式，调试的时候adapt_pts 为空，然后报错，请问这个值是干嘛的？ 怎么运行demo呢 谢谢！

由于本人远程实验室的机器，且没有root权限，所以无法使用您准备脚本配置环境，故均按照你的脚本改了改，现在在最后测试的时候出现这样的问题

python test.py multi_pose_wodet --exp_id coco512 --dataset coco_hp_wodet --resume --not_reg_offset --not_reg_hp_offset --K 20 --not_hm_hp --arch dla_34 --input_res 512 --keep_res
Keep resolution testing.
training chunk_sizes: [32]
The output will be saved to /home/renyouheng/AI_Code/AdaptivePose/src/lib/../../exp/multi_pose_wodet/coco512
heads {'hm': 1, 'hps': 34}
Namespace(K=20, aggr_weight=0.0, agnostic_ex=False, arch='dla_34', aug_ddd=0.5, aug_rot=0, batch_size=32, cat_spec_wh=False, center_thresh=0.1, chunk_sizes=[32], data_dir='/home/renyouheng/AI_Code/AdaptivePose/src/lib/../../data', dataset='coco_hp_wodet', debug=0, debug_dir='/home/renyouheng/AI_Code/AdaptivePose/src/lib/../../exp/multi_pose_wodet/coco512/debug', debugger_theme='white', demo='/home/users/yabo.xiao/CNet_v2/img', dense_hp=False, dense_wh=False, dep_weight=1, dim_weight=1, down_ratio=4, eval_oracle_dep=False, eval_oracle_hm=False, eval_oracle_hmhp=False, eval_oracle_hp_offset=False, eval_oracle_kps=False, eval_oracle_offset=False, eval_oracle_wh=False, exp_dir='/home/renyouheng/AI_Code/AdaptivePose/src/lib/../../exp/multi_pose_wodet', exp_id='coco512', fix_res=False, flip=0.5, flip_idx=[[1, 2], [3, 4], [5, 6], [7, 8], [9, 10], [11, 12], [13, 14], [15, 16]], flip_test=False, giou=False, giou_weight=1.0, gpus=[0], gpus_str='0', head_conv=256, heads={'hm': 1, 'hps': 34}, hide_data_time=False, hm_hp=False, hm_hp_weight=1, hm_weight=1, hp_weight=1, input_h=512, input_res=512, input_w=512, keep_res=True, kitti_split='3dop', load_model='/home/renyouheng/AI_Code/AdaptivePose/src/lib/../../exp/multi_pose_wodet/coco512/model_last.pth', lr=0.000125, lr_step=[90, 120], master_batch_size=32, mean=array([[[0.40789655, 0.44719303, 0.47026116]]], dtype=float32), metric='loss', mse_loss=False, nms=False, no_color_aug=False, norm_wh=False, not_cuda_benchmark=False, not_hm_hp=True, not_prefetch_test=False, not_rand_crop=False, not_reg_bbox=False, not_reg_hp_offset=True, not_reg_offset=True, num_classes=1, num_epochs=140, num_iters=-1, num_stacks=1, num_workers=4, off_weight=1, output_h=128, output_path='/opt/tiger/adaptivepose', output_res=128, output_w=128, pad=31, peak_thresh=0.2, print_iter=0, rect_mask=False, reg_bbox=True, reg_hp_offset=False, reg_loss='l1', reg_offset=False, resume=True, root_dir='/home/renyouheng/AI_Code/AdaptivePose/src/lib/../..', rot_weight=1, rotate=0, save_all=False, save_dir='/home/renyouheng/AI_Code/AdaptivePose/src/lib/../../exp/multi_pose_wodet/coco512', scale=0.4, scores_thresh=0.1, seed=317, shift=0.1, std=array([[[0.2886383 , 0.27408165, 0.27809834]]], dtype=float32), task='multi_pose_wodet', test=False, test_scales=[1.0], trainval=False, val_intervals=1, vis_thresh=0.3, wh_weight=0.1)
==> initializing coco 2017 val data.
loading annotations into memory...
Done (t=0.35s)
creating index...
index created!
Loaded val 5000 samples
Creating model...
backbone_Params: 19.72M
Traceback (most recent call last):
File "test.py", line 134, in
prefetch_test(opt)
File "test.py", line 58, in prefetch_test
detector = Detector(opt)
File "/home/renyouheng/AI_Code/AdaptivePose/src/lib/detectors/multi_pose_wodet.py", line 26, in init
super(MultiPoseDetector_wodet, self).init(opt)
File "/home/renyouheng/AI_Code/AdaptivePose/src/lib/detectors/base_detector.py", line 24, in init
self.model = create_model(opt.arch, opt.heads, opt.head_conv, is_train=False)
File "/home/renyouheng/AI_Code/AdaptivePose/src/lib/models/model.py", line 30, in create_model
model = get_model(num_layers=num_layers, heads=heads, head_conv=head_conv)
File "/home/renyouheng/AI_Code/AdaptivePose/src/lib/models/networks/pose_dla_dcn_v2.py", line 562, in get_pose_net
head_conv=head_conv)
File "/home/renyouheng/AI_Code/AdaptivePose/src/lib/models/networks/pose_dla_dcn_v2.py", line 528, in init
self.sampler = Feat_sampler(head_conv)
TypeError: init() missing 1 required positional argument: 'hps_channel'

您好！我很好奇src/lib/models/predictor.py 133行代码 为何不是返回 offset，而是offset1?

不知道demo怎么运行起来