Open tekknolagi opened 12 months ago
Do we need to load pixels in [0,1]
instead of [0,255]
?
Can we use loss fn from https://github.com/google/jax/blob/main/examples/mnist_classifier_fromscratch.py ?
TODO: only store weight grads at top-level; drop interstitial variables to locals. also find out if there is a way to do the same thing with the weights (but they are currently used in backprop)
Use llvmlite?
https://llvmlite.readthedocs.io/en/latest/user-guide/ir/examples.html https://llvmlite.readthedocs.io/en/latest/user-guide/binding/examples.html
or MLIR Python bindings
Mention tinygrad & C renderer
Use https://github.com/silentbicycle/hopscotch (Tarjan strongly connected components) to parallelize parts of graphs
Use https://github.com/jart/matmul or https://github.com/google/highway or https://github.com/kfjahnke/zimt for gemm/matmul
Cast to
unsigned char
https://github.com/tekknolagi/tekknolagi.github.com/blob/7f3998d154f3d19a3b4c45a151237d256d851a31/_posts/2023-09-19-compiling-ml-models.md?plain=1#L1032