-
-
Save domluna/bf1d3061244e7d7d9a3da467a6614006 to your computer and use it in GitHub Desktop.
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
using ReverseDiffPrototype | |
const RDP = ReverseDiffPrototype | |
using Distributions | |
# 784 * 20 * 10 | |
W1 = randn(784, 20) | |
W2 = randn(20, 10) | |
X = randn(32, 784) | |
Y = rand(DiscreteUniform(1, 10), 32) | |
RDP.@forward sigmoid(x) = 1. ./ (1. + exp(-x)) | |
function softmax(x) | |
#= xx = x - maximum(x) =# | |
exped = exp(x) | |
return exped ./ sum(exped) | |
end | |
#= function ce_loss(x, y) =# | |
#= bs = size(x, 1) =# | |
#= inds = zip(1:bs, y) =# | |
#= return mean(map(i -> x[i[1], i[2]], inds)) =# | |
#= end =# | |
function nn_forward(w1, w2, x, y) | |
x2 = sigmoid(x * w1) | |
logits = sigmoid(x2 * w2) | |
return sum(softmax(logits)) | |
#= return sum(logits) =# | |
#= return ce_loss(softmax(logits), y) =# | |
end | |
function nn_backward(w1, w2, x, y) | |
∇w1 = RDP.gradient(w -> nn_forward(w, w2, x, y), w1) | |
∇w2 = RDP.gradient(w -> nn_forward(w1, w, x, y), w2) | |
return (∇w1, ∇w2) | |
end | |
nn_backward(W1, W2, X, Y) | |
gc() | |
@time nn_backward(W1, W2, X, Y) | |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment