def check_dL_dW1(x, y, W1, b1, W2, b2, delta=1e-7):
'''
Compute gradient of the weights W1 a using gradient check.
Input:
delta: a small number for gradient check, a float scalar.
Output:
dL_dW1: the approximated gradient of the loss L w.r.t. the weights W1
'''
h, p = W1.shape
dL_dW1 = np.asmatrix(np.zeros((h, p)))
for i in range(h):
for j in range(p):
d = np.asmatrix(np.zeros((h, p)))
d[i, j] = delta
z1, a1, z2, a2 = forward(x, W1 + d, b1, W2, b2)
L = sr.compute_L(a2, y)
z1, a1, z2, a2 = forward(x, W1, b1, W2, b2)
dL_dW1[i, j] = (L - sr.compute_L(a2, y)) / delta
return dL_dW1
def compute_acc(X, Y, W):
y_hat = sr.compute_a(sr.compute_z(X, W, 0))
y_hat_indicies = np.argmax(y_hat, axis=1)
y_indicies = np.argmax(Y, axis=1)
error = np.array([y_indicies == y_hat_indicies])
acc = error.sum() / X.shape[0]
return acc
X = np.load('mnist_train_images.npy')
Y = np.load('mnist_train_labels.npy')
# W = train(X, Y, batch_size=75, epsilon=0.07, n_epochs=2, alpha=.005)
W = train(X, Y)
""" Performance Evaluation with the Validation Set """
X_valid = np.load('mnist_validation_images.npy')
Y_valid = np.load('mnist_validation_labels.npy')
acc = compute_acc(X_valid, Y_valid, W) * 100
# unreg_loss = compute_L_unreg(X_valid, Y_valid, W)
unreg_loss = sr.compute_L(X_valid, Y_valid, W)
print("Validation Loss: %.2f" % unreg_loss)
print("Validation Accuracy: %.2f %%." % acc)
plt.plot(L)
plt.title("Loss over each iteration (mini-batch)")
plt.show()
