search

facebookresearch / PoseDiffusion

[ICCV 2023] PoseDiffusion: Solving Pose Estimation via Diffusion-aided Bundle Adjustment
Other
718 stars 42 forks source link

training code bug #20

Closed yuyu0927 closed 1 year ago

yuyu0927 commented 1 year ago

Hi Jianyuan,

When I run 'python train.py', there is a bug 'AttributeError: 'Vizstats' object has no attribute 'get status string''. I checked the codes, and I found both 'class VizStats(Stats)' in your code and 'class Stats(object)' in 'pytorch3d/implicitron/tools/stats.py' have no this attribute. I changed 'stats.get_status_string' into 'stats.print' and it works. But I am not sure if this is correct. 1698060039443

Kind regards, Qingyu

jytime commented 1 year ago

Hi Qingyu,

I think it should be there, as a function of pytorch3d stats. But it is just something to print the logs out, you can choose anything you want. Using stats.print is correct, because get_status_string is just a wrapper for stats.print.

I would be able to test this part after mid November.

yuyu0927 commented 1 year ago

Hi Jianyuan,

You're right, it should be due to the difference in versions of pytorch3d.

I have another question about training code. from my view, the hyperparameter 'len_train' is used to control the number of iteration in each epoch. Does it have any relationship with 'num_sequences', which is the number of sequences/videos? Should it become larger when the size of dataset is larger?

jytime commented 1 year ago

Hi @yuyu0927 len_train is the length of training dataloader. I think you can make it the same to various datasets. It should not be related to num_sequences, just determined by how many iterations you want for each epoch.