From 1177fd2aee414e46f0880807a74c79cbee131592 Mon Sep 17 00:00:00 2001
From: Ryan Tjoa <ryan.j.tjoa@gmail.com>
Date: Fri, 22 Dec 2023 22:58:51 -0800
Subject: [PATCH] remove mnist example

---
 examples/mnist/Project.toml |  4 ---
 examples/mnist/mnist.jl     | 56 -------------------------------------
 2 files changed, 60 deletions(-)
 delete mode 100644 examples/mnist/Project.toml
 delete mode 100644 examples/mnist/mnist.jl

diff --git a/examples/mnist/Project.toml b/examples/mnist/Project.toml
deleted file mode 100644
index 17ae51c6..00000000
--- a/examples/mnist/Project.toml
+++ /dev/null
@@ -1,4 +0,0 @@
-[deps]
-CUDD = "345a2cc7-28d8-58b2-abdf-cff77ea7d7f1"
-Dice = "94eeadfc-1ba6-40b7-9ed8-dee71557d525"
-MLDatasets = "eb30cadb-4394-5ae3-aed4-317e484a6458"
diff --git a/examples/mnist/mnist.jl b/examples/mnist/mnist.jl
deleted file mode 100644
index c68b2cf1..00000000
--- a/examples/mnist/mnist.jl
+++ /dev/null
@@ -1,56 +0,0 @@
-# Proof-of-concept of training a network to classify MNIST digits
-# It's slow - we should use a 3rd party ML framework instead
-
-# Eventual goal would be to do MNIST-sum, as in
-# https://arxiv.org/pdf/1805.10872.pdf
-
-using MLDatasets: MNIST
-using Revise
-using Dice
-using Random
-
-map_sigmoid(x::ADNode) = ad_map(Dice.sigmoid, Dice.deriv_sigmoid, x)
-
-function discrete_weights(::Type{DistUInt{W}}, weights::Vector{<:ADNode}) where W
-    res = DistUInt{W}(length(weights) - 1)
-    acc = last(weights)
-    for i in reverse(1:length(weights) - 1)
-        acc += weights[i]
-        res = @dice_ite if flip(weights[i] / acc)
-            DistUInt{W}(i - 1)
-        else
-            res
-        end
-    end
-    res
-end
-
-HIDDEN_LAYER_SIZE = 100
-TRAIN_ROWS = 20
-
-Random.seed!(0)
-θ1_init = rand(28 * 28, HIDDEN_LAYER_SIZE)
-θ2_init = rand(HIDDEN_LAYER_SIZE, 10)
-
-clear_vars!()
-θ1 = new_var!("θ1", θ1_init)
-θ2 = new_var!("θ2", θ2_init)
-
-x = reshape(MNIST(:train).features, (28 * 28, 60000)) |> transpose
-x = x[1:TRAIN_ROWS, :]
-y = MNIST(:train).targets[1:TRAIN_ROWS]
-
-dw = map_sigmoid(map_sigmoid(x * θ1) * θ2)
-dw = [dw[i] for i in CartesianIndices((TRAIN_ROWS, 10))]
-predictions = [discrete_weights(DistUInt{4}, dw[i, :]) for i in 1:TRAIN_ROWS]
-correct = [prob_equals(pred, DistUInt{4}(label)) for (pred, label) in zip(predictions, y)]
-
-pr_corrects = [dist[true] for dist in pr(correct...)]
-expected_accuracy = sum(pr_corrects) / length(pr_corrects)
-println("Expected accuracy: $(expected_accuracy)")
-
-train_vars!(correct; epochs=1000, learning_rate=0.003)
-
-pr_corrects = [dist[true] for dist in pr(correct...)]
-expected_accuracy = sum(pr_corrects) / length(pr_corrects)
-println("Trained expected accuracy: $(expected_accuracy)")