Function main

ann_class2/sgd.py:24–178

()

def main():
    Xtrain, Xtest, Ytrain, Ytest = get_normalized_data()
    print("Performing logistic regression...")

    N, D = Xtrain.shape
    Ytrain_ind = y2indicator(Ytrain)
    Ytest_ind = y2indicator(Ytest)

    # 1. full
    W = np.random.randn(D, 10) / np.sqrt(D)
    W0 = W.copy() # save for later
    b = np.zeros(10)
    test_losses_full = []
    lr = 0.9
    reg = 0.
    t0 = datetime.now()
    last_dt = 0
    intervals = []
    for i in range(50):
        p_y = forward(Xtrain, W, b)

        gW = gradW(Ytrain_ind, p_y, Xtrain) / N
        gb = gradb(Ytrain_ind, p_y) / N

        W += lr*(gW - reg*W)
        b += lr*(gb - reg*b)

        p_y_test = forward(Xtest, W, b)
        test_loss = cost(p_y_test, Ytest_ind)
        dt = (datetime.now() - t0).total_seconds()

        # save these
        dt2 = dt - last_dt
        last_dt = dt
        intervals.append(dt2)

        test_losses_full.append([dt, test_loss])
        if (i + 1) % 10 == 0:
            print("Cost at iteration %d: %.6f" % (i + 1, test_loss))
    p_y = forward(Xtest, W, b)
    print("Final error rate:", error_rate(p_y, Ytest))
    print("Elapsted time for full GD:", datetime.now() - t0)

    # save the max time so we don't surpass it in subsequent iterations
    max_dt = dt
    avg_interval_dt = np.mean(intervals)


    # 2. stochastic
    W = W0.copy()
    b = np.zeros(10)
    test_losses_sgd = []
    lr = 0.001
    reg = 0.

    t0 = datetime.now()
    last_dt_calculated_loss = 0
    done = False