Added train by epoch for Trainer and added support for texts

lucidrains / meshgpt-pytorch

Implementation of MeshGPT, SOTA Mesh generation using Attention, in Pytorch

MIT License

744 stars 59 forks source link

Added train by epoch for Trainer and added support for texts #12

Open MarcusLoppe opened 10 months ago

MarcusLoppe commented 10 months ago

Trainer

Added train function that will train with epochs instead of steps
- Added option to display graph (maybe remove since it requires matlib?)
Added checkpoint_every_epoch to init so checkpoints can be saved every X.

Data.py

Modified custom_collate so it wont pad if the data isn't a tensor

MeshAutoencoder

Was crashing if texts was in the args, added dummy parameterto prevent this.

Setup.py

The setup file was missing a comma

lucidrains commented 10 months ago

this is awesome Marcus! will take a look at it tomorrow morning and get it merged!

adeerAI commented 10 months ago

Awesome work, thanks for including your suggestions to the main, this allows better understanding on the user's side.

MarcusLoppe commented 10 months ago

Btw since I don't really think grad_accum_every is very useful I removed it from the train function, what is your option?

I forgot and left grad_accum_every in the loss function, so if it wont be used in the train function it should be removed from:

self.accelerator.backward(loss / self.grad_accum_every)

lucidrains commented 10 months ago

Btw since I don't really think grad_accum_every is very useful I removed it from the train function, what is your option?

I forgot and left grad_accum_every in the loss function, so if it wont be used in the train function it should be removed from:

self.accelerator.backward(loss / self.grad_accum_every)

i'm sure researchers will want to stretch to the next level if this approach pans out (multiple meshes, scenes etc)

probably good to keep it for the gpu poor

MarcusLoppe commented 10 months ago

Another thing :) I'm not very experienced in using github forks but it seems like the pull request added later commits then when I made the request.

I made bit of a error and replaced entire meshgpt_pytorch.py since there was some weird stash thing and I wanted to ensure it was up to date. I reverted but it seems like that stash thing messed it up bit, please double check if this is the case

lucidrains commented 10 months ago

hey sorry, the caching stuff took up my entire morning

will revisit this PR tomorrow!

Ginjing-Yuan commented 4 months ago

Hi! Thanks for your amazing job! May I ask why in MeshGPT_dem.ipynb has a AttributeError: module 'meshgpt_pytorch.mesh_render' has no attribute 'combind_mesh_with_rows'