From 6d4283c99d31d2594a3a669f78653e033f8f8a93 Mon Sep 17 00:00:00 2001
From: castrong <strongcreativesystems@gmail.com>
Date: Mon, 29 Jun 2020 18:42:42 -0700
Subject: [PATCH 1/2] added testing_utils.jl file, added make_random_network
 function

---
 src/utils/testing_utils.jl | 34 ++++++++++++++++++++++++++++++++++
 1 file changed, 34 insertions(+)
 create mode 100644 src/utils/testing_utils.jl

diff --git a/src/utils/testing_utils.jl b/src/utils/testing_utils.jl
new file mode 100644
index 00000000..660fe2bf
--- /dev/null
+++ b/src/utils/testing_utils.jl
@@ -0,0 +1,34 @@
+using Random
+"""
+    make_random_network(layer_sizes::Vector{Int, 1}, [min_weight = -1.0], [max_weight = 1.0], [min_bias = -1.0], [max_bias = 1.0], [rng = 1.0])
+    read_layer(output_dim::Int, f::IOStream, [act = ReLU()])
+
+Generate a network with random weights and bias. The first layer is treated as the input.
+The values for the weights and bias will be uniformly drawn from the range between min_weight
+and max_weight and min_bias and max_bias respectively. The last layer will have an ID()
+activation function and the rest will have ReLU() activation functions. Allow a random number
+generator(rng) to be passed in. This allows for seeded random network generation.
+"""
+function make_random_network(layer_sizes::Vector{Int, 1}, min_weight = -1.0, max_weight = 1.0, min_bias = -1.0, max_bias = 1.0, rng=MersenneTwister(0))
+    # Create each layer based on the layer_size
+    layers = []
+    for index in 1:(length(layer_sizes)-1)
+        # Use Id activation for the last layer - otherwise use ReLU activation
+        if index == length(layer_sizes)
+            cur_activation = NeuralVerification.Id()
+        else
+            cur_activation = NeuralVerification.ReLU()
+        end
+
+        # Dimension: num_out x num_in
+        cur_weights = min_weight + (max_weight - min_weight) * rand(rng, Float64, (layer_sizes(index+1), layer_sizes(index)))
+        cur_weights = reshape(cur_weights, (layer_sizes(index+1), layer_sizes(index))) # for edge case where 1 dimension is equal to 1 this keeps it from being a 1-d vector
+
+        # Dimension: num_out x 1
+        cur_bias = min_weight + (max_weight - min_weight) * rand(rng, Float64, (layer_sizes(index+1)))
+        push!(layers, Layer(cur_weights, cur_bias, cur_activation))
+    end
+
+    return Network(layers)
+
+end

From bacf128749e6e278d55f8f51964205d1fbb07bdb Mon Sep 17 00:00:00 2001
From: castrong <strongcreativesystems@gmail.com>
Date: Mon, 29 Jun 2020 18:52:53 -0700
Subject: [PATCH 2/2] fixed bugs

---
 src/utils/testing_utils.jl | 21 ++++++++++++---------
 1 file changed, 12 insertions(+), 9 deletions(-)

diff --git a/src/utils/testing_utils.jl b/src/utils/testing_utils.jl
index 660fe2bf..44ca1d94 100644
--- a/src/utils/testing_utils.jl
+++ b/src/utils/testing_utils.jl
@@ -1,6 +1,8 @@
 using Random
+using NeuralVerification;
+
 """
-    make_random_network(layer_sizes::Vector{Int, 1}, [min_weight = -1.0], [max_weight = 1.0], [min_bias = -1.0], [max_bias = 1.0], [rng = 1.0])
+    make_random_network(layer_sizes::Vector{Int}, [min_weight = -1.0], [max_weight = 1.0], [min_bias = -1.0], [max_bias = 1.0], [rng = 1.0])
     read_layer(output_dim::Int, f::IOStream, [act = ReLU()])
 
 Generate a network with random weights and bias. The first layer is treated as the input.
@@ -9,26 +11,27 @@ and max_weight and min_bias and max_bias respectively. The last layer will have
 activation function and the rest will have ReLU() activation functions. Allow a random number
 generator(rng) to be passed in. This allows for seeded random network generation.
 """
-function make_random_network(layer_sizes::Vector{Int, 1}, min_weight = -1.0, max_weight = 1.0, min_bias = -1.0, max_bias = 1.0, rng=MersenneTwister(0))
+function make_random_network(layer_sizes::Vector{Int}, min_weight = -1.0, max_weight = 1.0, min_bias = -1.0, max_bias = 1.0, rng=MersenneTwister(0))
     # Create each layer based on the layer_size
     layers = []
     for index in 1:(length(layer_sizes)-1)
+        cur_size = layer_sizes[index]
+        next_size = layer_sizes[index+1]
         # Use Id activation for the last layer - otherwise use ReLU activation
-        if index == length(layer_sizes)
+        if index == (length(layer_sizes)-1)
             cur_activation = NeuralVerification.Id()
         else
             cur_activation = NeuralVerification.ReLU()
         end
 
         # Dimension: num_out x num_in
-        cur_weights = min_weight + (max_weight - min_weight) * rand(rng, Float64, (layer_sizes(index+1), layer_sizes(index)))
-        cur_weights = reshape(cur_weights, (layer_sizes(index+1), layer_sizes(index))) # for edge case where 1 dimension is equal to 1 this keeps it from being a 1-d vector
+        cur_weights = min_weight .+ (max_weight - min_weight) * rand(rng, Float64, (next_size, cur_size))
+        cur_weights = reshape(cur_weights, (next_size, cur_size)) # for edge case where 1 dimension is equal to 1 this keeps it from being a 1-d vector
 
         # Dimension: num_out x 1
-        cur_bias = min_weight + (max_weight - min_weight) * rand(rng, Float64, (layer_sizes(index+1)))
-        push!(layers, Layer(cur_weights, cur_bias, cur_activation))
+        cur_bias = min_bias .+ (max_bias - min_bias) * rand(rng, Float64, (next_size))
+        push!(layers, NeuralVerification.Layer(cur_weights, cur_bias, cur_activation))
     end
-
+    
     return Network(layers)
-
 end