"machine learning" + "fragment shader"
TensorFlow.js: Machine Learning for the Web and Beyond
This paper describes the design and development of the TensorFlow.js library, which was motivated by the importance of the JS community and web-based applications for ML
"In order to utilize the GPU, TensorFlow.js uses WebGL, a cross-platform web standard providing low-level 3D graphics APIs. Unlike OpenCL and CUDA, the WebGL API is based on OpenGL ES specification (Shreiner et al., 2013) which has no explicit support for GPGPU."
Fragment shaders run for each pixel independently and in parallel; TensorFlow.js takes advantage of this parallelization to accelerate ML computation.
On WebGL 2.0 devices, we use the gl.R32F texture type which allows us to avoid allocating memory for the green, blue, and alpha channels (shown as ‘G’, ‘B’, and ‘A’ respectively).
<aside> 💡 Take aways: how tensorflow utilities webgl; maybe webgl
</aside>
Few-shot Knowledge Transfer for Fine-grained Cartoon Face Generation
First, a basic translation model for the basic group (which consists of sufficient data) is trained. Then, given new samples of other groups, we extend the basic model by creating group-specific branches for each new group. Group-specific branches are updated directly to capture specific appearances for each group while the remaining group-shared parameters are updated indirectly to maintain the distribution of intermediate feature space. In this manner, our approach is capable to generate high-quality cartoon faces for various groups.
We design a multi-branch image translation network for fine-grained face generation. The main branch learns to translate images from the common group and maintain the distribution of the shared feature space while other branches learn specific characteristics for each rare group. In this manner, we can learn a few-shot image translation model effectively. The main contribution of the proposed work can be summarized as follows: