Merge pull request #101 from castrong/make_random_network_util

Add a function to generate random networks

src/utils/testing_utils.jl

@@ -0,0 +1,37 @@
+using Random
+using NeuralVerification;
+
+"""
+    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.
+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}, 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)-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, (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_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