Hi, I am unable to replicate the results in the paper. the final results reported is always higher than the results we are getting using multiple tf versions and using the same seeds. For example, we ran motion_prediction --action discussion --visualize 0 and get 0.32 & 0.62 & 1.10 & 1.31 & 1.87 & 2.04 & 2.20 & 2.44 The results reported in the paper are: HMR 0.29 0.55 0.83 0.94 1.35 1.49 1.61 1.72

The validation loss never reduced beyond Best Validation Loss: 0.31701756060123426 in the first epoch.

Could you please let me know what tf.set_random_seed(112858) you set or if you are able to reproduce the results?

I I found the paper titled Investigating Pose Representations and Motion Contexts Modeling for 3D Motion Prediction, but your github title is Motion prediction with Hierarchical Motion Recurrent Network, and the project name is lie-group-motion-prediction. The paper explicitly uses Stiefel manifold parameterization of rotation instead of lie-group. Or maybe Stiefel manifold parameterization of rotation is the idea of lie-group？ Finally, is the code described in this paper consistent with this github code? Looking forward to your reply thank you very much

Hi, This is a great work, I would like to reproduct this project, however, it is hard to download not only by command"git" but also download directly, the documents are always incomplete after "downloading", and I try to use the proxy and it doesn't work, so could you please upload the project to the google cload or any other place for the download operation! Thanks again! Yours, Hu

Hello, I was wondering how to produce the results for 1000 frames output as in the video given on your site? Could you please upload the trained model? I tried to train the model by setting longterm to true and here is my output for walking: Video Thanks in advance.

When I set "datatype" as "xyz", I get follow issue, you may forget to set the "chain_idx"

Traceback (most recent call last): File "/home/cz5270/Lie-Group-Motion-Prediction/src/motion_prediction.py", line 316, in tf.app.run() File "/home/cz5270/anaconda3/envs/tf/lib/python3.7/site-packages/tensorflow/python/platform/app.py", line 40, in run _run(main=main, argv=argv, flags_parser=_parse_flags_tolerate_undef) File "/home/cz5270/anaconda3/envs/tf/lib/python3.7/site-packages/absl/app.py", line 300, in run _run_main(main, args) File "/home/cz5270/anaconda3/envs/tf/lib/python3.7/site-packages/absl/app.py", line 251, in _run_main sys.exit(main(argv)) File "/home/cz5270/Lie-Group-Motion-Prediction/src/motion_prediction.py", line 218, in main train() File "/home/cz5270/Lie-Group-Motion-Prediction/src/motion_prediction.py", line 44, in train prediction = models.seq2seq(x, dec_in, config, True) File "/home/cz5270/Lie-Group-Motion-Prediction/src/models.py", line 102, in seq2seq outputs, states = tf.contrib.legacy_seq2seq.rnn_decoder(dec_in, enc_state, dec_cell, loop_function=lf) File "/home/cz5270/anaconda3/envs/tf/lib/python3.7/site-packages/tensorflow/contrib/legacy_seq2seq/python/ops/seq2seq.py", line 152, in rnn_decoder output, state = cell(inp, state) File "/home/cz5270/Lie-Group-Motion-Prediction/src/models.py", line 184, in call output = xyz_resize(output, self.config) File "/home/cz5270/Lie-Group-Motion-Prediction/src/models.py", line 374, in xyz_resize index = config.chain_idx AttributeError: 'train_Config' object has no attribute 'chain_idx'

Hello, I tried to run the linearized loss by making changes to the loss function as on: Issue

The numbers I am getting on evaluation are very weird and high for the Mean Joint Error. I have attached the files for the reference. Could you provide the checkpoints for model weights to reproduce the result? Errors_walking.xlsx

For humans - The data is loaded from h3.6m for lie datatype whereas the logic of loading lie data exists in read_human function in read_data.py. Why is a different function created to read lie datatype ?

Hi, I realized in your code " loss_functions.py" there are "l2_loss" and ' linearizedlie_loss', I would like to ask do you have the code for the geodesic loss mentioned in your paper Investigating Pose Representations and Motion Contexts Modeling for 3D Motion Prediction.

After restoring your pretrained model and setting the random seed, I calculate the 3D coordinate error between y_t_xyz and y_p_xyz using the following L2-norm,

def xyz_error(gt, out):
    '''
    gt: B, 100, 66
    '''
    batch_size, seq_len, features = gt.shape
    t_3d = np.zeros(seq_len)

    gt = gt.reshape(batch_size, seq_len, -1, 3) # B, 100, 22, 3
    out = out.reshape(batch_size, seq_len, -1, 3) # B, 100, 22, 3
    for k in np.arange(0, seq_len):
        t_3d[k] = np.mean(np.linalg.norm(gt[:, k, :, :].reshape(-1, 3) - out[:, k, :, :].reshape(-1, 3), 2, axis=1))
    return t_3d

the results are larger than those in your paper, I want to know how do you calculate errors in 3D coordinate ? have you scaled the source data?