test.py

"""
A program to evaluate the results of constructed neural networks
"""

import mnist_loader
# import libraries
import random
import numpy as np

# import networks
import network
import network_2

# import hyperparameters
from hyperparams import NUM_EPOCHS, LEARNING_RATE, LAYERS, REG_PARAM, MINI_BATCH_SIZE, EARLY_STOPPING


training_data, validation_data, test_data = mnist_loader.load_data_wrapper()
training_data, validation_data, test_data = list(training_data), list(validation_data), list(test_data)
# print(list(training_data)[5][0].shape) numpy array! 784 input layer

# Manipulating pseudo-random state
# So you can get the exact same results as I did!
random.seed(123581321)
np.random.seed(123581321)


"""
!!! General Note:
1. Hyperparameters are specified in hyperparams.py
2. Only the latest version of the network, complemented with additional features is uncommented
3. Re-write the hyperparameters to customise and explore
4. Further instructions in the respective sections
"""


"""
Neural network 1.0
SGD + backprop using sigmoid function
"""
print("Current: Network 1.0 with basic neural network structure.\n")
net = network.Network(LAYERS)
net.SGD(training_data, NUM_EPOCHS, MINI_BATCH_SIZE, LEARNING_RATE, validation_data = validation_data)
total_test = len(test_data)
results = 100.0 * net.evaluate(test_data)/total_test
print(f"Accuracy on test data: {results}%\n\n")


"""
Neural network 2.2
Similar to 1.0, with the addition of
a) L2 regularization
b) Cross-entropy function
c) Better initialization of weights
d) Early stopping

By default, only the accuracy on the validation set is displayed after each epoch.
Toggle the 4 display modes accordingly to your preference.

monitor_eval_cost: if set to True, gives the total cost incurred by the network on the evaluation data for that epoch
monitor_eval_acc : if set to True, gives the total number of correct classification out of the total size of validation data
monitor_trng_cost: if set to True, gives the total cost incurred by the network on the training data for that epoch
monitor)trng_acc : if set to True, gives the total number of correct classification out of the total size of the training data
"""
print(f"Current: Network 2.0 with L2 reg, cross-entropy cost, better initialisation of weights, early stopping of {EARLY_STOPPING} epochs.\n")
net = network_2.Network(LAYERS, cost=network_2.CrossEntropyCost)
net.SGD(training_data, NUM_EPOCHS, MINI_BATCH_SIZE,
        LEARNING_RATE, reg_param = REG_PARAM,
        eval_data = validation_data,
        monitor_eval_cost=False,
        monitor_eval_acc=True,
        monitor_trng_cost=False,
        monitor_trng_acc=False,
        early_stopping=EARLY_STOPPING)
total_test = len(test_data)
results = 100.0 * net.accuracy(test_data, convert=False)/total_test
print(f"Accuracy on test data: {results}%\n\n")