Why doesn't my smplx.pkl file work?

[Ginjing-Yuan]

When I put my photo trained pkl file in DINAR, it reports:
RuntimeError: The size of tensor a (134) must match the size of tensor b (165) at non-singleton dimension 1.
If I change my pkl to random one of your pkls, the training will start. I suppose our pkl files may generate from different model. Which one are you using?

[Ginjing-Yuan]

When I put my photo trained pkl file in DINAR, it reports: RuntimeError: The size of tensor a (134) must match the size of tensor b (165) at non-singleton dimension 1. If I change my pkl to random one of your pkls, the training will start. I suppose our pkl files may generate from different model. Which one are you using?

I found the 'body_pose' part of pkl you provided had more parameters.

[david-svitov]

Hi! First please make shure that you use SMPL-X model for producing your .pkl with the same parameters as in dataloader:

DINAR/utils/smplx_model.py

Line 20 in 7368958

gender2model = {k: smplx_flame.SMPLX(

If it doesn't help please provide some additional information. Like the line on witch you get this error. And if it's possible the example of your .pkl

[Ginjing-Yuan]

Hi David! I try to follow your parameters, but it doesn't work. These are all preprocessed data I used, and I also add a pkl file produced by me from your dataset. Here is the google drive link.
https://drive.google.com/file/d/1MDfl049Ii3NlANjzuSW6psRTTFRh3BwS/view?usp=drive_link

[wf1024966]

Hi David! I try to follow your parameters, but it doesn't work. These are all preprocessed data I used, and I also add a pkl file produced by me from your dataset. Here is the google drive link. https://drive.google.com/file/d/1MDfl049Ii3NlANjzuSW6psRTTFRh3BwS/view?usp=drive_link

I have the same problem.

[Ginjing-Yuan]

Moreover, when I use pickle check the pkl file, there is a parameter which is 'ltrb': (28, 65, 1046, 1083) misses too.

[wf1024966]

When I put my photo trained pkl file in DINAR, it reports: RuntimeError: The size of tensor a (134) must match the size of tensor b (165) at non-singleton dimension 1. If I change my pkl to random one of your pkls, the training will start. I suppose our pkl files may generate from different model. Which one are you using?

I found the 'body_pose' part of pkl you provided had more parameters.

你好，这个问题我已经有解决方式了，fit_single_frame.py中把result['body_pose'] = pose_embedding.detach().cpu().numpy()改成 result['body_pose'] = (vposer.decode(pose_embedding).get('pose_body').detach().cpu().numpy()).reshape((1,63))，就可以得到正确的结果。 dict中 如果没有'ltrb': (28, 65, 1046, 1083)，代码也能正常运行。

[Ginjing-Yuan]

When I put my photo trained pkl file in DINAR, it reports: RuntimeError: The size of tensor a (134) must match the size of tensor b (165) at non-singleton dimension 1. If I change my pkl to random one of your pkls, the training will start. I suppose our pkl files may generate from different model. Which one are you using?

I found the 'body_pose' part of pkl you provided had more parameters.

你好，这个问题我已经有解决方式了，fit_single_frame.py中把result['body_pose'] = pose_embedding.detach().cpu().numpy()改成 result['body_pose'] = (vposer.decode(pose_embedding).get('pose_body').detach().cpu().numpy()).reshape((1,63))，就可以得到正确的结果。 dict中 如果没有'ltrb': (28, 65, 1046, 1083)，代码也能正常运行。

你好，我用修改过的代码得到了如下报错：

Traceback (most recent call last):
File "smplifyx/main.py", line 273, in
main(**args)
File "smplifyx/main.py", line 246, in main
fit_single_frame(img, keypoints[[person_id]],
File "C:\Users\91802\anaconda3\envs\sm382\smplify-x-master\smplifyx\fit_single_frame.py", line 473, in fit_single_frame
result['body_pose'] = vposer.decode(pose_embedding.get('pose_body').detach().cpu().numpy()).reshape((1,63))
AttributeError: 'Tensor' object has no attribute 'get' 请问你是怎么解决的？

[Dragonkingpan]

When I put my photo trained pkl file in DINAR, it reports: RuntimeError: The size of tensor a (134) must match the size of tensor b (165) at non-singleton dimension 1. If I change my pkl to random one of your pkls, the training will start. I suppose our pkl files may generate from different model. Which one are you using?

I found the 'body_pose' part of pkl you provided had more parameters.

你好，这个问题我已经有解决方式了，fit_single_frame.py中把result['body_pose'] = pose_embedding.detach().cpu().numpy()改成 result['body_pose'] = (vposer.decode(pose_embedding).get('pose_body').detach().cpu().numpy()).reshape((1,63))，就可以得到正确的结果。 dict中 如果没有'ltrb': (28, 65, 1046, 1083)，代码也能正常运行。

我也试图按你的办法这样进行修改，

 vp = vposer.decode(pose_embedding)
 result['body_pose'] = (vp.get('pose_body').detach().cpu().numpy()).reshape((1, 63))

但是vposer解码出来的vp没有pose_body这个东西，直接报错了。
您是否能把您用的smplx的代码给我一份，谢谢！

[wf1024966]

When I put my photo trained pkl file in DINAR, it reports: RuntimeError: The size of tensor a (134) must match the size of tensor b (165) at non-singleton dimension 1. If I change my pkl to random one of your pkls, the training will start. I suppose our pkl files may generate from different model. Which one are you using?

I found the 'body_pose' part of pkl you provided had more parameters.

你好，这个问题我已经有解决方式了，fit_single_frame.py中把result['body_pose'] = pose_embedding.detach().cpu().numpy()改成 result['body_pose'] = (vposer.decode(pose_embedding).get('pose_body').detach().cpu().numpy()).reshape((1,63))，就可以得到正确的结果。 dict中 如果没有'ltrb': (28, 65, 1046, 1083)，代码也能正常运行。

我也试图按你的办法这样进行修改，

 vp = vposer.decode(pose_embedding)
 result['body_pose'] = (vp.get('pose_body').detach().cpu().numpy()).reshape((1, 63))

但是vposer解码出来的vp没有pose_body这个东西，直接报错了。 您是否能把您用的smplx的代码给我一份，谢谢！

你能把dict发给我看一下吗？看下字典里是哪些元素

[wf1024966]

result

When I put my photo trained pkl file in DINAR, it reports: RuntimeError: The size of tensor a (134) must match the size of tensor b (165) at non-singleton dimension 1. If I change my pkl to random one of your pkls, the training will start. I suppose our pkl files may generate from different model. Which one are you using?

I found the 'body_pose' part of pkl you provided had more parameters.

你好，这个问题我已经有解决方式了，fit_single_frame.py中把result['body_pose'] = pose_embedding.detach().cpu().numpy()改成 result['body_pose'] = (vposer.decode(pose_embedding).get('pose_body').detach().cpu().numpy()).reshape((1,63))，就可以得到正确的结果。 dict中 如果没有'ltrb': (28, 65, 1046, 1083)，代码也能正常运行。

你好，我用修改过的代码得到了如下报错：

Traceback (most recent call last): File "smplifyx/main.py", line 273, in main(**args) File "smplifyx/main.py", line 246, in main fit_single_frame(img, keypoints[[person_id]], File "C:\Users\91802\anaconda3\envs\sm382\smplify-x-master\smplifyx\fit_single_frame.py", line 473, in fit_single_frame result['body_pose'] = vposer.decode(pose_embedding.get('pose_body').detach().cpu().numpy()).reshape((1,63)) AttributeError: 'Tensor' object has no attribute 'get' 请问你是怎么解决的？

你好，你这里代码写错了，pose_embedding是tensor，vposer.decode(pose_embedding)得到字典，然后才能用get方式得到这个字典里的key 'pose_body' 里的值。

[Ginjing-Yuan]

result

When I put my photo trained pkl file in DINAR, it reports: RuntimeError: The size of tensor a (134) must match the size of tensor b (165) at non-singleton dimension 1. If I change my pkl to random one of your pkls, the training will start. I suppose our pkl files may generate from different model. Which one are you using?

I found the 'body_pose' part of pkl you provided had more parameters.

你好，这个问题我已经有解决方式了，fit_single_frame.py中把result['body_pose'] = pose_embedding.detach().cpu().numpy()改成 result['body_pose'] = (vposer.decode(pose_embedding).get('pose_body').detach().cpu().numpy()).reshape((1,63))，就可以得到正确的结果。 dict中 如果没有'ltrb': (28, 65, 1046, 1083)，代码也能正常运行。

你好，我用修改过的代码得到了如下报错：
Traceback (most recent call last): File "smplifyx/main.py", line 273, in main(**args) File "smplifyx/main.py", line 246, in main fit_single_frame(img, keypoints[[person_id]], File "C:\Users\91802\anaconda3\envs\sm382\smplify-x-master\smplifyx\fit_single_frame.py", line 473, in fit_single_frame result['body_pose'] = vposer.decode(pose_embedding.get('pose_body').detach().cpu().numpy()).reshape((1,63)) AttributeError: 'Tensor' object has no attribute 'get' 请问你是怎么解决的？

你好，你这里代码写错了，pose_embedding是tensor，vposer.decode(pose_embedding)得到字典，然后才能用get方式得到这个字典里的key 'pose_body' 里的值。

我copy您的代码还是报错
result['body_pose'] = (vposer.decode(pose_embedding).get('pose_body').detach().cpu().numpy()).reshape((1,63))
AttributeError: 'Tensor' object has no attribute 'get' 。

当我在fit_single_frame前print了vposer.decode(pose_embedding)，
print(vposer.decode(pose_embedding))
result['body_pose'] = (vposer.decode(pose_embedding).get('pose_body').detach().cpu().numpy()).reshape((1,63))
报错前打出了：
tensor([[[[ 9.9986e-01, -1.6692e-02, -2.7833e-03, 1.6616e-02, 9.9954e-01,
-2.5422e-02, 3.2063e-03, 2.5372e-02, 9.9967e-01],
[ 9.9874e-01, -4.8467e-02, 1.2789e-02, 4.7347e-02, 9.9593e-01,
7.6745e-02, -1.6457e-02, -7.6043e-02, 9.9697e-01],
[ 9.9995e-01, -4.4295e-04, 1.0162e-02, 1.8311e-03, 9.9057e-01,
-1.3701e-01, -1.0005e-02, 1.3702e-01, 9.9052e-01],
[ 9.9778e-01, -8.8067e-03, -6.5952e-02, -1.0669e-02, 9.5720e-01,
-2.8923e-01, 6.5676e-02, 2.8929e-01, 9.5499e-01],
[ 9.9907e-01, 3.2578e-02, 2.8256e-02, -1.9618e-02, 9.2683e-01,
-3.7496e-01, -3.8403e-02, 3.7405e-01, 9.2661e-01],
[ 9.9926e-01, -2.4563e-02, -2.9638e-02, 2.7915e-02, 9.9255e-01,
1.1857e-01, 2.6505e-02, -1.1931e-01, 9.9250e-01],
[ 9.2808e-01, -2.0060e-01, 3.1374e-01, 7.6028e-02, 9.2683e-01,
3.6770e-01, -3.6455e-01, -3.1740e-01, 8.7542e-01],
[ 9.6471e-01, 3.9284e-02, -2.6036e-01, 4.6853e-02, 9.4742e-01,
3.1656e-01, 2.5910e-01, -3.1759e-01, 9.1214e-01],
[ 9.9987e-01, -1.5678e-02, -4.7349e-03, 1.4951e-02, 9.9181e-01,
-1.2686e-01, 6.6850e-03, 1.2677e-01, 9.9191e-01],
[ 9.9952e-01, 5.0317e-03, 3.0614e-02, -6.0931e-03, 9.9938e-01,
3.4678e-02, -3.0421e-02, -3.4848e-02, 9.9893e-01],
[ 9.9984e-01, -1.3444e-02, 1.1786e-02, 1.3680e-02, 9.9970e-01,
-2.0155e-02, -1.1512e-02, 2.0313e-02, 9.9973e-01],
[ 9.9595e-01, 3.4300e-02, 8.3116e-02, -4.4194e-02, 9.9176e-01,
1.2028e-01, -7.8305e-02, -1.2347e-01, 9.8925e-01],
[ 9.9953e-01, -9.9169e-03, 2.8893e-02, 9.0059e-03, 9.9946e-01,
3.1489e-02, -2.9190e-02, -3.1214e-02, 9.9909e-01],
[ 9.9939e-01, -3.2153e-02, 1.3901e-02, 3.1169e-02, 9.9733e-01,
6.5993e-02, -1.5986e-02, -6.5519e-02, 9.9772e-01],
[ 9.9956e-01, 1.7088e-02, 2.4266e-02, -1.7326e-02, 9.9980e-01,
9.6616e-03, -2.4096e-02, -1.0078e-02, 9.9966e-01],
[ 6.9919e-01, 6.9857e-01, -1.5208e-01, -7.0285e-01, 7.1059e-01,
3.2621e-02, 1.3086e-01, 8.4085e-02, 9.8783e-01],
[ 6.4331e-01, -7.3618e-01, 2.1023e-01, 7.5950e-01, 6.4826e-01,
-5.4001e-02, -9.6527e-02, 1.9441e-01, 9.7616e-01],
[ 9.5359e-01, 3.5420e-02, -2.9903e-01, -6.6384e-02, 9.9335e-01,
-9.4028e-02, 2.9371e-01, 1.0952e-01, 9.4960e-01],
[ 9.1868e-01, 6.3024e-02, 3.8994e-01, -5.3356e-02, 9.9794e-01,
-3.5588e-02, -3.9138e-01, 1.1888e-02, 9.2015e-01],
[ 9.9327e-01, -8.3776e-02, -7.9967e-02, 1.0367e-01, 9.5095e-01,
2.9146e-01, 5.1627e-02, -2.9779e-01, 9.5323e-01],
[ 9.9609e-01, 8.8226e-02, 5.4535e-03, -8.6616e-02, 9.6188e-01,
2.5939e-01, 1.7640e-02, -2.5885e-01, 9.6576e-01]]]],
device='cuda:0', grad_fn=)
rientation: 0%| | 0/1 [00:58<?, ?it/s]
Traceback (most recent call last):
File "smplifyx/main.py", line 273, in
main(**args)
File "smplifyx/main.py", line 246, in main
fit_single_frame(img, keypoints[[person_id]],
File "C:\Users\91802\anaconda3\envs\sm382\smplify-x-master\smplifyx\fit_single_frame.py", line 475, in fit_single_frame
result['body_pose'] = (result['body_pose'].get('pose_body').detach().cpu().numpy()).reshape((1,63))
AttributeError: 'Tensor' object has no attribute 'get'

[wf1024966]

result

When I put my photo trained pkl file in DINAR, it reports: RuntimeError: The size of tensor a (134) must match the size of tensor b (165) at non-singleton dimension 1. If I change my pkl to random one of your pkls, the training will start. I suppose our pkl files may generate from different model. Which one are you using?

I found the 'body_pose' part of pkl you provided had more parameters.

你好，这个问题我已经有解决方式了，fit_single_frame.py中把result['body_pose'] = pose_embedding.detach().cpu().numpy()改成 result['body_pose'] = (vposer.decode(pose_embedding).get('pose_body').detach().cpu().numpy()).reshape((1,63))，就可以得到正确的结果。 dict中 如果没有'ltrb': (28, 65, 1046, 1083)，代码也能正常运行。

你好，我用修改过的代码得到了如下报错：
Traceback (most recent call last): File "smplifyx/main.py", line 273, in main(**args) File "smplifyx/main.py", line 246, in main fit_single_frame(img, keypoints[[person_id]], File "C:\Users\91802\anaconda3\envs\sm382\smplify-x-master\smplifyx\fit_single_frame.py", line 473, in fit_single_frame result['body_pose'] = vposer.decode(pose_embedding.get('pose_body').detach().cpu().numpy()).reshape((1,63)) AttributeError: 'Tensor' object has no attribute 'get' 请问你是怎么解决的？

你好，你这里代码写错了，pose_embedding是tensor，vposer.decode(pose_embedding)得到字典，然后才能用get方式得到这个字典里的key 'pose_body' 里的值。

我copy您的代码还是报错 result['body_pose'] = (vposer.decode(pose_embedding).get('pose_body').detach().cpu().numpy()).reshape((1,63)) AttributeError: 'Tensor' object has no attribute 'get' 。

当我在fit_single_frame前print了vposer.decode(pose_embedding)， print(vposer.decode(pose_embedding)) result['body_pose'] = (vposer.decode(pose_embedding).get('pose_body').detach().cpu().numpy()).reshape((1,63)) 报错前打出了： tensor([[[[ 9.9986e-01, -1.6692e-02, -2.7833e-03, 1.6616e-02, 9.9954e-01, -2.5422e-02, 3.2063e-03, 2.5372e-02, 9.9967e-01], [ 9.9874e-01, -4.8467e-02, 1.2789e-02, 4.7347e-02, 9.9593e-01, 7.6745e-02, -1.6457e-02, -7.6043e-02, 9.9697e-01], [ 9.9995e-01, -4.4295e-04, 1.0162e-02, 1.8311e-03, 9.9057e-01, -1.3701e-01, -1.0005e-02, 1.3702e-01, 9.9052e-01], [ 9.9778e-01, -8.8067e-03, -6.5952e-02, -1.0669e-02, 9.5720e-01, -2.8923e-01, 6.5676e-02, 2.8929e-01, 9.5499e-01], [ 9.9907e-01, 3.2578e-02, 2.8256e-02, -1.9618e-02, 9.2683e-01, -3.7496e-01, -3.8403e-02, 3.7405e-01, 9.2661e-01], [ 9.9926e-01, -2.4563e-02, -2.9638e-02, 2.7915e-02, 9.9255e-01, 1.1857e-01, 2.6505e-02, -1.1931e-01, 9.9250e-01], [ 9.2808e-01, -2.0060e-01, 3.1374e-01, 7.6028e-02, 9.2683e-01, 3.6770e-01, -3.6455e-01, -3.1740e-01, 8.7542e-01], [ 9.6471e-01, 3.9284e-02, -2.6036e-01, 4.6853e-02, 9.4742e-01, 3.1656e-01, 2.5910e-01, -3.1759e-01, 9.1214e-01], [ 9.9987e-01, -1.5678e-02, -4.7349e-03, 1.4951e-02, 9.9181e-01, -1.2686e-01, 6.6850e-03, 1.2677e-01, 9.9191e-01], [ 9.9952e-01, 5.0317e-03, 3.0614e-02, -6.0931e-03, 9.9938e-01, 3.4678e-02, -3.0421e-02, -3.4848e-02, 9.9893e-01], [ 9.9984e-01, -1.3444e-02, 1.1786e-02, 1.3680e-02, 9.9970e-01, -2.0155e-02, -1.1512e-02, 2.0313e-02, 9.9973e-01], [ 9.9595e-01, 3.4300e-02, 8.3116e-02, -4.4194e-02, 9.9176e-01, 1.2028e-01, -7.8305e-02, -1.2347e-01, 9.8925e-01], [ 9.9953e-01, -9.9169e-03, 2.8893e-02, 9.0059e-03, 9.9946e-01, 3.1489e-02, -2.9190e-02, -3.1214e-02, 9.9909e-01], [ 9.9939e-01, -3.2153e-02, 1.3901e-02, 3.1169e-02, 9.9733e-01, 6.5993e-02, -1.5986e-02, -6.5519e-02, 9.9772e-01], [ 9.9956e-01, 1.7088e-02, 2.4266e-02, -1.7326e-02, 9.9980e-01, 9.6616e-03, -2.4096e-02, -1.0078e-02, 9.9966e-01], [ 6.9919e-01, 6.9857e-01, -1.5208e-01, -7.0285e-01, 7.1059e-01, 3.2621e-02, 1.3086e-01, 8.4085e-02, 9.8783e-01], [ 6.4331e-01, -7.3618e-01, 2.1023e-01, 7.5950e-01, 6.4826e-01, -5.4001e-02, -9.6527e-02, 1.9441e-01, 9.7616e-01], [ 9.5359e-01, 3.5420e-02, -2.9903e-01, -6.6384e-02, 9.9335e-01, -9.4028e-02, 2.9371e-01, 1.0952e-01, 9.4960e-01], [ 9.1868e-01, 6.3024e-02, 3.8994e-01, -5.3356e-02, 9.9794e-01, -3.5588e-02, -3.9138e-01, 1.1888e-02, 9.2015e-01], [ 9.9327e-01, -8.3776e-02, -7.9967e-02, 1.0367e-01, 9.5095e-01, 2.9146e-01, 5.1627e-02, -2.9779e-01, 9.5323e-01], [ 9.9609e-01, 8.8226e-02, 5.4535e-03, -8.6616e-02, 9.6188e-01, 2.5939e-01, 1.7640e-02, -2.5885e-01, 9.6576e-01]]]], device='cuda:0', grad_fn=) rientation: 0%| | 0/1 [00:58<?, ?it/s] Traceback (most recent call last): File "smplifyx/main.py", line 273, in main(**args) File "smplifyx/main.py", line 246, in main fit_single_frame(img, keypoints[[person_id]], File "C:\Users\91802\anaconda3\envs\sm382\smplify-x-master\smplifyx\fit_single_frame.py", line 475, in fit_single_frame result['body_pose'] = (result['body_pose'].get('pose_body').detach().cpu().numpy()).reshape((1,63)) AttributeError: 'Tensor' object has no attribute 'get'

step 1

human_body_prior工程不要用cvpr19分支去安装编译，git checkout到master分支去安装编译

step 2 - change

old version

from human_body_prior.tools.model_loader import load_vposer

new version

from human_body_prior.tools.model_loader import load_model
from human_body_prior.models.vposer_model import VPoser

step 3 - change

old version

vposer, _ = load_vposer(vposer_ckpt, vp_model='snapshot')

new version

vposer, _ = load_model(vposer_ckpt, model_code=VPoser, remove_words_in_model_weights='vp_model.', disable_grad=True).

step 4

官网去下载新的vposer的ckpt文件，替换原来的pt文件，放到相应文件夹下。

[mayank64ce]

Hi @wf1024966 I tried to follow your steps but it gives the following error:

Traceback (most recent call last):
  File "smplifyx/main.py", line 274, in <module>
    main(**args)
  File "smplifyx/main.py", line 247, in main
    fit_single_frame(img, keypoints[[person_id]],
  File "/home/ma982513/Projects/smplify-x/smplifyx/fit_single_frame.py", line 274, in fit_single_frame
    init_t = fitting.guess_init(body_model, gt_joints, edge_indices,
  File "/home/ma982513/.conda/envs/smplifyx/lib/python3.8/site-packages/torch/autograd/grad_mode.py", line 26, in decorate_context
    return func(*args, **kwargs)
  File "/home/ma982513/Projects/smplify-x/smplifyx/fitting.py", line 74, in guess_init
    body_pose = vposer.decode(
TypeError: decode() got an unexpected keyword argument 'output_type'

[mayank64ce]

@Dragonkingpan @wf1024966 @Ginjing-Yuan anyone of you getting strange output from smplifyx like this : vchoutas/smplify-x#210

[david-svitov]

I see there was quite a long discussion here. I took a new machine, ran the code on it, and wrote down all the steps I took to fix it in this document:
https://docs.google.com/document/d/1LTua0D7XDpVJ9Jrk7s1WSjXIYIHBN89txCUXl1wCsvk/edit?usp=sharing