BShapeGen

A Simple Toolset for Creating a Model and Predicting 3D Polygon Vertex Data using PyTorch and Maya

Table of Contents

• Key Features
• Requirements
• Setup
• Step by Step
• Video Demo
• Results
• Next Steps

Key Features

• Initialize Test Data Scene
• Export Training Data
• Build Model
• Predict and Import Data

Requirements

• Anaconda Installed
• Maya +2020 Installed

Setup

1. Clone or Download Source and Unzip to Local Machine

2. Open a Terminal or Command Prompt and Navigate to the top bshagegen directory
   

   Windows ie. c:/dev/src/bshapegen

   Mac ie. ~\dev\src\bshapegen

3. Windows - in command prompt run:

   .\cmd\init_env_conda.bat
   

   MacOs M1 - in terminal run:

   source ./sh/init_env_conda_M1.sh
   

4. Confirm PyTorch is Installed by running the following from the same shell
   

   python -c "import torch;print(torch.__version__)"
   

   Successful return should look like this (versions might differ)

   2.1.0
   

Step by Step

1. Clone Repo Locally
2. Launch Maya
3. Run this in a python tab in the script editor (change bsg_py_path to your local machine location)

import os
import sys
#
bsg_py_path = '/Users/zoshua/src/bshapegen/py'
sys.path.append(bsg_py_path)
#
from bshapegen.maya import bsg_ui
#
bsg_ui = bsg_ui.BSG_UI();
bsg_ui.show(dockable=True);

4. Dock the BlendShapeGen dialog if needed
5. Update the Work Dir location if needed
6. Click the Build Test Scene button
7. Click the Export Training Data button
8. Click the Build Model button
9. Click the Export > Predict > Import Data button

Video Demo

Color Guide:

• yellow = training input neutral data (model build)
• orange = training output pose data (model build)
• cyan = neutral input data (model prediction)
• blue = reference pose data (used to measure and visualize model accuracy)
• magenta = pose output data (model prediction)

Sample Build and Predict Results

train samples	neurons	epochs	learning rate	loss	validation	predict samples	predict err sum
100	512	150	0.001	0.007996	0.007446	5	3.16174
100	1024	150	0.001	0.002554	0.002974	5	2.92702
100	1024	300	0.001	0.000129	0.000123	5	2.98027

Next Steps

• Explore alternate model configurations
• Explore higher resolution mesh data