Training with artificial images is becoming increasingly important to address the lack of real data sets in various niche areas. Yet, many today’s approaches write 2D/3D simulations from scratch.
To improve this situation and make better use of existing pipelines, we’ve been working towards an integration between Blender, an open-source real-time physics enabled animation software, and PyTorch. Today we announce blendtorch, an open-source Python library that seamlessly integrates distributed Blender renderings into PyTorch data pipelines at 60FPS (640x480 RGBA).
@ptrblck and @tom thanks for your feedback. We have just released v0.2 which brings an even better PyTorch integration, plus OpenAI.gym support that allows us to use Blender as modelling, simulation and interactive manipulation framework to train your RL agents through the well known OpenAI gym interface. Below is an animated gif showing a recreation of the cartpole demo running in Blender.
I am trying to create a chatbot using Blender and Pytorch. I’ve got the blend file in which I’ve created the poses (the rain rig). The Rain rig responds to the audio input based on the (.dat) file. And honestly, I’m still clueless about how it gonna create poses based on output from the conversational module in real-time.
Although I came across blendtorch- a Python framework to seamlessly integrate Blender into PyTorch datasets for deep learning from artificial visual data
But I’m still wondering how it will create poses on the go ie. based on conversational input in real-time. Can you please guide me a little on how shall I proceed?