EfficientNet on tinygrad + WebGL
This is a WebGL port of EfficientNet using tinygrad. WebGL isn't yet in master, but we have a PR up for it.
Try it here!
To run the model on WebGL, we added a new tinygrad backend. The beauty in tinygrad is that you can do this in about ~100 lines of code, depending on the backend. We used moderngl to run the kernels in a Python environment, and modified extra/export_model.py to support WebGL. It exports all the kernels and a small JS runtime in a single net.js file.
With the WebGL backend the kernels are fragment shaders, and the tensor buffers are 2D textures. Indexing is based on gl_FragCoord. The get the texture x coordinate (window space), we have to modulo the index with the texture width, and to get the y, we divide with texture width. The window-space x,y are then translated to uv space by dividing by texture width and height. Smaller buffers are Nx1, but if the max supported texture dimension is exceeded alongside the x dimension, we create NxM textures.
Each "kernel" invocation actually is just rendering to a framebuffer texture, and that framebuffer texture can be used in subsequent kernels. To get the output of the model we can use glReadPixels.
MIT