def callback(epoch, paramvec, vals):
print 'epoch {}: {}'.format(
epoch, prediction_error(paramvec, test_im, test_angle))
update_training_progress(fig, ax, line, vals)
plot_images_and_angles(test_im, predict(test_im, paramvec),
prediction_fig)
if save_figs: prediction_fig.figure.savefig('orientations.png')
def callback(epoch, paramvec, vals):
print 'epoch {}: {}'.format(epoch, prediction_error(paramvec, test_im, test_angle))
update_training_progress(fig, ax, line, vals)
plot_images_and_angles(test_im, predict(test_im, paramvec), prediction_fig)
if save_figs: prediction_fig.figure.savefig('orientations.png')
fig.draw_artist(line)
if save_figs: fig.savefig('training.png')
plt.pause(1e-6)
if __name__ == "__main__":
npr.seed(0)
plt.ion()
# load training data and plot some examples
(images,
angles), (test_im,
test_angle) = load_training_data('data/labeled_images.pkl',
augmentation=19,
hold_out=80)
data_fig = plot_images_and_angles(images[:80], angles[:80])
if save_figs: data_fig.figure.savefig('data.png')
imsize = images.shape[1]
hdims = [50, 50]
# l2_reg = empirical_l2_reg(images, hdims)
l2_reg = 0.
paramvec, unflatten = flatten(init_gmlp(hdims, imsize, 1))
predict, loss, prediction_error = make_regression(l2_reg, unflatten)
# make a figure for training progress
fig, ax = plt.subplots()
line, = ax.plot([])
plt.draw()
def update_training_progress(fig, ax, line, vals):
line.set_data(range(1, len(vals)+1), vals)
ax.set_ylim(min(vals), np.percentile(vals, 99))
ax.set_xlim(1, len(vals)+1)
fig.draw_artist(line)
if save_figs: fig.savefig('training.png')
plt.pause(1e-6)
if __name__ == "__main__":
npr.seed(0)
plt.ion()
# load training data and plot some examples
(images, angles), (test_im, test_angle) = load_training_data(
'data/labeled_images.pkl', augmentation=19, hold_out=80)
data_fig = plot_images_and_angles(images[:80], angles[:80])
if save_figs: data_fig.figure.savefig('data.png')
imsize = images.shape[1]
hdims = [50, 50]
# l2_reg = empirical_l2_reg(images, hdims)
l2_reg = 0.
paramvec, unflatten = flatten(init_gmlp(hdims, imsize, 1))
predict, loss, prediction_error = make_regression(l2_reg, unflatten)
# make a figure for training progress
fig, ax = plt.subplots()
line, = ax.plot([])
plt.draw()