def df(x,y,W,b):
p = part2.forward(x, W, b)
grad_w = np.matmul((p-y), x.T).T
grad_b = np.sum((p-y), 1).reshape((10,1))
return grad_w, grad_b
def part5(alpha_w, alpha_b, _max_iter, printing):
M = loadmat("mnist_all.mat")
results_val = []
results_train = []
results_test = []
x = []
train_set, train_set_sol, val_set, val_set_sol, test_set, test_set_sol = part4.generate_sets(
M, 5000)
W_init = np.random.rand(784, 10)
step = 20
for i in range(step, _max_iter, step):
print(i)
W, b = grad_descent(part3.f, part3.df, train_set, train_set_sol,
W_init, alpha_w, alpha_b, i, 0.9, True)
y_pred = part2.forward(val_set, W, b)
results_val += [
np.mean(np.argmax(y_pred, 0) == np.argmax(val_set_sol, 0))
]
y_pred = part2.forward(train_set, W, b)
results_train += [
np.mean(np.argmax(y_pred, 0) == np.argmax(train_set_sol, 0))
]
y_pred = part2.forward(test_set, W, b)
results_test += [
np.mean(np.argmax(y_pred, 0) == np.argmax(test_set_sol, 0))
]
print results_test
print results_train
print results_val
x += [i]
pickle.dump(W, open("part4W_shichen.p", "wb"))
plt.plot(x, results_val, 'y', label="validation set")
plt.plot(x, results_train, 'g', label="training set")
plt.plot(x, results_test, 'r', label="test set")
plt.title("Part 5 Learning Curve")
plt.ylabel("accuracy")
plt.xlabel("iterations")
plt.show()
def test2(database, solutions, W, b):
correct = 0
database = database.T
solutions = solutions.T
for i in range(len(database)):
guess = part2.forward(database[i], W, b)
if np.argmax(guess) == np.argmax(solutions[i]):
correct += 1
return correct / float(len(database))
def test(database, size, W, b):
test_set = []
correct = 0
for i in range(size):
rand_dgt = np.random.random_integers(0, 9)
test_set += [database["test" + str(rand_dgt)][i]]
guess = part2.forward(test_set[i], W, b)
if guess == rand_dgt:
correct += 1
return correct / size
def f(x, W, b, y):
p = part2.forward(x,W,b)
return -sum(y*log(p)) / x.shape[1]