About the Training Code

[Bingoang]

@enriccorona
Hi, enriccorona~
I really appreciate your great work!
I'm so interested in your SMPLicit, and will you release the training code?

Best wishes

[enriccorona]

Hi, sorry for the delay. I wasn't planning on releasing the training code since part of the data is private and I can't share it, but here you can download the code we used to train SMPLicit.
https://drive.google.com/file/d/1Z-urM0XKxHM-d3EbxAI8ugugu4eODpbm/view?usp=sharing

Since I wasn't planning on sharing it, I didn't spend much time cleaning it, but hopefully it's quite clear. You can run the train code on a GPU by running: "python train.py --dataset_mode bcnet_tshirts --model bcnet_sdf_dist2eachcluster_constrainshape --name SMPLicit_upperbody"

The dataloader's code is probably quite important: It loads BCNet's shirts vertices and faces (which are common since clothes are registered), and our annotations of different sleeve sizes. For each example, we pick a random sleeve length and build the mesh from those annotations. Next, we just sample a number of random points and calculate the unsigned distance. I left the uvmap renders in the folder to show how we represent occupancy, but they take almost 5 GB.

The model is quite simple though, it is based on an MLP where the input is the positional encoding of the point's position, and predicts unsigned distance. We do have a couple regularization hyperparameters to restrict the latent representations, but other than that everything is quite standard.

I'm sure you'll have to fix the paths to the BCNet dataset folder if you need to try anything. Let me know if you have any problems,

[Bingoang]

@enriccorona Wonderful!! Thank you so much for your patient reply and instruction, that's what we SMPLicit's fans want~ I will have a try!
Thanks again!

[albertrial]

Hi @enriccorona first of all, congratulations on the work done, really appreciated!

I am also trying to train SMPLicit with my own data. After several hours working on the dataloader's code, adapting my data to your model, I have been able to load my input (vertices, faces, occlusion UV maps, ...) and run the training code. However, I am finding that the model is not able to train at all (not even overfit on one or a few samples). This makes me think that something in my data is wrong, but I have analysed the sampled random points, the computed unsigned distance and the UV maps, and they all seem correct.

My question is, even if the BCNet dataset is private, do you know if I could obtain a single sample to see exactly what your input looks like? That would be very helpful.

Thank you very much!

[zoroinparis]

I'm really sorry to bother you. May I ask where can I get the dataset, or how should I deploy my own data for training?

[Lai-dongdong]

Hi, sorry for the delay. I wasn't planning on releasing the training code since part of the data is private and I can't share it, but here you can download the code we used to train SMPLicit. https://drive.google.com/file/d/1Z-urM0XKxHM-d3EbxAI8ugugu4eODpbm/view?usp=sharing

Since I wasn't planning on sharing it, I didn't spend much time cleaning it, but hopefully it's quite clear. You can run the train code on a GPU by running: "python train.py --dataset_mode bcnet_tshirts --model bcnet_sdf_dist2eachcluster_constrainshape --name SMPLicit_upperbody"

The dataloader's code is probably quite important: It loads BCNet's shirts vertices and faces (which are common since clothes are registered), and our annotations of different sleeve sizes. For each example, we pick a random sleeve length and build the mesh from those annotations. Next, we just sample a number of random points and calculate the unsigned distance. I left the uvmap renders in the folder to show how we represent occupancy, but they take almost 5 GB.

The model is quite simple though, it is based on an MLP where the input is the positional encoding of the point's position, and predicts unsigned distance. We do have a couple regularization hyperparameters to restrict the latent representations, but other than that everything is quite standard.

I'm sure you'll have to fix the paths to the BCNet dataset folder if you need to try anything. Let me know if you have any problems,

Hi, the link has expired. Could you please provide a new one?

[sunshineatnoon]

Hi, sorry for the delay. I wasn't planning on releasing the training code since part of the data is private and I can't share it, but here you can download the code we used to train SMPLicit. https://drive.google.com/file/d/1Z-urM0XKxHM-d3EbxAI8ugugu4eODpbm/view?usp=sharing
Since I wasn't planning on sharing it, I didn't spend much time cleaning it, but hopefully it's quite clear. You can run the train code on a GPU by running: "python train.py --dataset_mode bcnet_tshirts --model bcnet_sdf_dist2eachcluster_constrainshape --name SMPLicit_upperbody"
The dataloader's code is probably quite important: It loads BCNet's shirts vertices and faces (which are common since clothes are registered), and our annotations of different sleeve sizes. For each example, we pick a random sleeve length and build the mesh from those annotations. Next, we just sample a number of random points and calculate the unsigned distance. I left the uvmap renders in the folder to show how we represent occupancy, but they take almost 5 GB.
The model is quite simple though, it is based on an MLP where the input is the positional encoding of the point's position, and predicts unsigned distance. We do have a couple regularization hyperparameters to restrict the latent representations, but other than that everything is quite standard.
I'm sure you'll have to fix the paths to the BCNet dataset folder if you need to try anything. Let me know if you have any problems,

Hi, the link has expired. Could you please provide a new one?

Hi, Did you get the code?

[shash29-dev]

The link is expired. Please share again, if possible.