test_weights.py

#!/usr/bin/env python3
import numpy as np
import math

def sigmoid(x):
  return 1 / (1 + math.exp(-x))

X = np.load("X.npy")
Y = np.load("Y.npy")

hidden_weights_1 = np.load("hidden_weights_1.npy")
hidden_bias_1 = np.load("hidden_bias_1.npy")
hidden_weights_2 = np.load("hidden_weights_2.npy")
hidden_bias_2 = np.load("hidden_bias_2.npy")
output_weights = np.load("output_weights.npy")
output_bias = np.load("output_bias.npy")

print(X.shape)
print(hidden_weights_1.shape)
print(hidden_bias_1.shape)
print(hidden_weights_2.shape)
print(hidden_bias_2.shape)
print(output_weights.shape)
print(output_bias.shape)

def from_onehot(x):
  return max((cl,i) for (i,cl) in enumerate(x))[1]

#
# Test that we can replicate the results of the trained model 
# from the saved weights
#
def predict(x):
    hidden_out_1 = x @ hidden_weights_1 + hidden_bias_1
    hidden_out_1 = hidden_out_1 * (hidden_out_1 > 0)    # relu

    hidden_out_2 = hidden_out_1 @ hidden_weights_2 + hidden_bias_2
    hidden_out_2 = hidden_out_2 * (hidden_out_2 > 0)    # relu

    output = hidden_out_2 @ output_weights + output_bias
    #print(output)
    return from_onehot(output)

# Count correct predictions on training set
correct = 0
for x,y in zip(X,Y):
  if from_onehot(y) == predict(x): correct += 1
print("Accuracy: %.4f" % (correct / len(Y)))