oop_and_ml/prj05.py

274 lines
No EOL
8.5 KiB
Python
Executable file

import numpy as np
import time
import pdb
from matplotlib import pyplot
# save theta to p5_params.npz that can be used by easynn
def save_theta(theta):
f1_W, f1_b, f2_W, f2_b = theta
np.savez_compressed("p5_params.npz", **{
"f1.weight": f1_W,
"f1.bias": f1_b,
"f2.weight": f2_W,
"f2.bias": f2_b
})
# initialize theta using uniform distribution [-bound, bound]
# return theta as (f1_W, f1_b, f2_W, f2_b)
def initialize_theta(bound):
f1_W = np.random.uniform(-bound, bound, (32, 784))
f1_b = np.random.uniform(-bound, bound, 32)
f2_W = np.random.uniform(-bound, bound, (10, 32))
f2_b = np.random.uniform(-bound, bound, 10)
return (f1_W, f1_b, f2_W, f2_b)
# forward:
# x = Flatten(images)
# g = Linear_f1(x)
# h = ReLU(g)
# z = Linear_f2(h)
# return (z, h, g, x)
def forward(images, theta):
# number of samples
N = images.shape[0]
# unpack theta into f1 and f2
f1_W, f1_b, f2_W, f2_b = theta
# x = Flatten(images)
x = images.astype(float).transpose(0,3,1,2).reshape((N, -1))
# g = Linear_f1(x)
g = np.zeros((N, f1_b.shape[0]))
for i in range(N):
g[i, :] = np.matmul(f1_W, x[i])+f1_b
# h = ReLU(g)
h = g*(g > 0)
# z = Linear_f2(h)
z = np.zeros((N, f2_b.shape[0]))
for i in range(N):
z[i, :] = np.matmul(f2_W, h[i])+f2_b
return (z, h, g, x)
# backprop:
# J = cross entropy between labels and softmax(z)
# return nabla_J
def backprop(labels, theta, z, h, g, x):
# number of samples
N = labels.shape[0]
# unpack theta into f1 and f2
f1_W, f1_b, f2_W, f2_b = theta
# nabla_J consists of partial J to partial f1_W, f1_b, f2_W, f2_b
p_f1_W = np.zeros(f1_W.shape)
p_f1_b = np.zeros(f1_b.shape)
p_f2_W = np.zeros(f2_W.shape)
p_f2_b = np.zeros(f2_b.shape)
for i in range(N):
# compute the contribution to nabla_J for sample i
# cross entropy and softmax
# compute partial J to partial z[i]
# scale by 1/N for averaging
expz = np.exp(z[i]-max(z[i]))
p_z = expz/sum(expz)/N
p_z[labels[i]] -= 1/N
# z = Linear_f2(h)
# compute partial J to partial h[i]
# accumulate partial J to partial f2_W, f2_b
p_h = np.dot(f2_W.T, p_z)
p_f2_W += np.outer(p_z, h[i])
p_f2_b += p_z
# h = ReLU(g)
# compute partial J to partial g[i]
p_g = p_h * (g[i] > 0)
# g = Linear_f1(x)
# accumulate partial J to partial f1_W, f1_b
p_f1_W += np.outer(p_g, x[i])
p_f1_b += p_g
return (p_f1_W, p_f1_b, p_f2_W, p_f2_b)
# apply SGD to update theta by nabla_J and the learning rate epsilon
# return updated theta
def update_theta(theta, nabla_J, epsilon):
# ToDo: modify code below as needed
#updated_theta = theta
#return updated_theta
f1_W, f1_b, f2_W, f2_b = theta
p_f1_W, p_f1_b, p_f2_W, p_f2_b = nabla_J
# update the weights and biases for the first layer (f1)
f1_W_updated = f1_W - epsilon * p_f1_W
f1_b_updated = f1_b - epsilon * p_f1_b
# update the weights and biases for the second layer (f2)
f2_W_updated = f2_W - epsilon * p_f2_W
f2_b_updated = f2_b - epsilon * p_f2_b
return (f1_W_updated, f1_b_updated, f2_W_updated, f2_b_updated)
def print_training_hyperparams_for_session(epsilon, batch_size, bound):
print("Starting training session with 10 epochs:")
print("")
print("Hyperparameters:")
print(f"epsilon: {epsilon}")
print(f"bound: {bound}")
print(f"batch_size: {batch_size}")
print("")
print("Results:")
def plot_epoch(epochs, accuracies, epsilon, batch_size, bound):
pyplot.figure(figsize=(10, 6))
pyplot.plot(epochs, accuracies, label=f"Epsilon: {epsilon}, Batch Size: {batch_size}, Bound: {bound}")
pyplot.xlabel('Epoch')
pyplot.ylabel('Accuracy')
pyplot.title('Training Accuracy over Epochs')
pyplot.legend()
pyplot.grid(True)
pyplot.show()
def plot_all_epochs(training_results):
pyplot.figure(figsize=(12, 8))
for epochs, accuracies, epsilon, batch_size, bound in training_results:
label = f"Epsilon: {epsilon}, Batch Size: {batch_size}, Bound: {bound}"
pyplot.plot(epochs, accuracies, label=label)
pyplot.xlabel('Epoch')
pyplot.ylabel('Accuracy')
pyplot.title('Training Accuracy over Epochs for Different Hyperparameters')
pyplot.legend()
pyplot.grid(True)
pyplot.show()
def plot_table(training_results):
# Setting up the data for the table
cell_text = []
columns = ['Epoch', 'Accuracy', 'Epsilon', 'Batch Size', 'Bound']
for result in training_results:
epochs, accuracies, epsilon, batch_size, bound = result
for epoch, accuracy in zip(epochs, accuracies):
cell_text.append([epoch, f"{accuracy:.3f}", epsilon, batch_size, bound])
# Determine the figure size needed for the table
figsize = (10, len(cell_text) * 0.2)
fig, ax = pyplot.subplots(figsize=figsize)
ax.axis('tight')
ax.axis('off')
# Create the table
table = ax.table(cellText=cell_text, colLabels=columns, loc='center', cellLoc='center')
# Adjust table scale
table.auto_set_font_size(False)
table.set_fontsize(8)
table.auto_set_column_width(col=list(range(len(columns))))
pyplot.show()
def start_training(epsilon, batch_size, bound, mnist_train):
# ToDo: set numpy random seed to the last 8 digits of your CWID
np.random.seed(20497299)
validation_images = mnist_train["images"][:1000]
validation_labels = mnist_train["labels"][:1000]
training_images = mnist_train["images"][1000:]
training_labels = mnist_train["labels"][1000:]
# hyperparameters
# we can experiment with these values to see if increasing or decreasing
# these values can influence our accuracy
# default values
#bound = 1 # initial weight range
#epsilon = 0.00001 # learning rate
#print_training_hyperparams_for_session(epsilon, batch_size, bound)
# start training
accuracies = []
epochs = []
start = time.time()
theta = initialize_theta(bound)
batches = training_images.shape[0]//batch_size
for epoch in range(10):
indices = np.arange(training_images.shape[0])
np.random.shuffle(indices)
for i in range(batches):
batch_images = training_images[indices[i*batch_size:(i+1)*batch_size]]
batch_labels = training_labels[indices[i*batch_size:(i+1)*batch_size]]
z, h, g, x = forward(batch_images, theta)
nabla_J = backprop(batch_labels, theta, z, h, g, x)
theta = update_theta(theta, nabla_J, epsilon)
# check accuracy using validation examples
z, _, _, _ = forward(validation_images, theta)
pred_labels = z.argmax(axis = 1)
accuracy = sum(pred_labels == validation_labels) / validation_images.shape[0]
accuracies.append(accuracy)
epochs.append(epoch)
#count = sum(pred_labels == validation_labels)
print("epoch %d, accuracy %.3f, time %.2f" % (
epoch, accuracy, time.time()-start))
#plot_epoch(epochs, accuracies, epsilon, batch_size, bound)
# save the weights to be submitted
save_theta(theta)
# return this data so we can plot it with matplotlib
return epochs, accuracies
def main():
training_results = []
mnist_train = np.load("mnist_train.npz")
training_results = []
# load training data once
mnist_train = np.load("mnist_train.npz")
# we can add to this list if we want to test combinations of hyperparameters
hyperparams = [
(0.00001, 1, 4), # default params
(0.00001, 0.1, 4),
(0.00001, 0.5, 4),
(0.00001, 0.7, 4),
(0.00001, 0.01, 4),
(0.00001, 0.01, 3),
(0.00001, 0.01, 2),
(0.000013, 0.012, 1),
(0.000013, 0.012002899999999983, 1),
(0.000013, 0.01200591999999996, 1),
]
for epsilon, bound, batch_size in hyperparams:
epochs, accuracies = start_training(epsilon, batch_size, bound, mnist_train)
training_results.append((epochs, accuracies, epsilon, batch_size, bound))
# uncomment if you would like to see plotted results
#plot_all_epochs(training_results)
# plot table
#plot_table(training_results[9])
if __name__ == '__main__':
main()