Hi! 
I'm building a C++ ML library. I built it because I don't like how python is the default language for ML. Currently it allows you to train neural nets on CPU with very fast training & inference speeds.
Planning to expand to CUDA and ROCm soon! 
#Wolf - Fast native C++ ML Library
stray cobalt
modern hedge
But Python isn't actually used for training or inference.
stray cobalt
It's intended to be way faster than using a c++ pytorch wrapper
modern hedge
How would you achieve this?
modern hedge
How are you comparing this?
stray cobalt
Using pytorch nn.sequential with the same config
modern hedge
Is your implementation doing the same thing as pytorch? Does it produce equivalent or better models?
stray cobalt
the math done is the same, standard stochastic gradient descent
There is randomness involved with the initialization of the weight vector
I could try benchmarking them more when I have the time
stray cobalt
What kind of results were you able to achieve?
Training a
Linear(784, 128),
ReLU(),
Linear(128, 10)
model, which is 100k parameters on 50k training set data on mnist took 1 second per epoch on my AMD Ryzen 5 7535HS