def plot():
with open('results.pkl') as f:
all_learning_curves, all_val_loss, all_test_loss,\
all_fakedata, all_fakedatasize, all_L2 = zip(*pickle.load(f))
import matplotlib as mpl
mpl.rcParams['font.family'] = 'serif'
# Fake data
fig = plt.figure(0)
fig.clf()
ax = plt.Axes(fig, [0., 0., 1., 1.])
fig.set_size_inches((4.5, 1.75))
ax.set_axis_off()
fig.add_axes(ax)
images = all_fakedata[-1]
immin = np.min(images.ravel())
immax = np.max(images.ravel())
cax = plot_mnist_images(images, ax, ims_per_row=5, padding=2)
cbar = fig.colorbar(cax, ticks=[immin, 0, immax], shrink=.7)
cbar.ax.set_yticklabels(
['{:2.2f}'.format(immin), '0', '{:2.2f}'.format(immax)])
plt.savefig('fake_data.png', bbox_inches='tight')
plt.savefig('fake_data.pdf', bbox_inches='tight')
# Everything else plot.
fig = plt.figure(0)
fig.clf()
N_figs = 3
ax = fig.add_subplot(N_figs, 1, 1)
ax.set_title("Learning Curves")
subsample = np.ceil(float(len(all_learning_curves)) / 50)
for i, log_alphas in enumerate(all_learning_curves):
if i % subsample == 0:
ax.plot(log_alphas, 'o-')
ax.set_ylabel("Loss")
ax.set_xlabel("Step number")
#ax.legend(loc=4)
ax = fig.add_subplot(N_figs, 1, 2)
ax.set_title("Meta Learning Curve")
all_train_loss = [curve[-1] for curve in all_learning_curves]
ax.plot(all_train_loss, 'o-', label="Train Loss")
ax.plot(all_val_loss, 'o-', label="Validation Loss")
ax.plot(all_test_loss, 'o-', label="Test Loss")
ax.plot(all_fakedatasize, 'o-', label="Fake Data Scale")
ax.plot(all_L2, 'o-', label="L2_regularization")
ax.set_ylabel("Validation Loss")
ax.set_xlabel("Meta Iteration Number")
ax.legend(loc=2)
ax = fig.add_subplot(N_figs, 1, 3)
ax.set_title("Fake Data")
images = all_fakedata[-1]
plot_mnist_images(images, ax, ims_per_row=10)
fig.set_size_inches((8, 12))
plt.savefig("fig.png")
def plot():
with open('results.pkl') as f:
all_learning_curves, all_val_loss, all_test_loss,\
all_fakedata, all_fakedatasize, all_L2 = zip(*pickle.load(f))
import matplotlib as mpl
mpl.rcParams['font.family'] = 'serif'
# Fake data
fig = plt.figure(0)
fig.clf()
ax = plt.Axes(fig, [0., 0., 1., 1.])
fig.set_size_inches((4.5,1.75))
ax.set_axis_off()
fig.add_axes(ax)
images = all_fakedata[-1]
immin = np.min(images.ravel())
immax = np.max(images.ravel())
cax = plot_mnist_images(images, ax, ims_per_row=5, padding=2)
cbar = fig.colorbar(cax, ticks=[immin, 0, immax], shrink=.7)
cbar.ax.set_yticklabels(['{:2.2f}'.format(immin), '0', '{:2.2f}'.format(immax)])
plt.savefig('fake_data.png', bbox_inches='tight')
plt.savefig('fake_data.pdf', bbox_inches='tight')
# Everything else plot.
fig = plt.figure(0)
fig.clf()
N_figs = 3
ax = fig.add_subplot(N_figs, 1, 1)
ax.set_title("Learning Curves")
subsample = np.ceil(float(len(all_learning_curves)) / 50)
for i, log_alphas in enumerate(all_learning_curves):
if i % subsample == 0:
ax.plot(log_alphas, 'o-')
ax.set_ylabel("Loss")
ax.set_xlabel("Step number")
#ax.legend(loc=4)
ax = fig.add_subplot(N_figs, 1, 2)
ax.set_title("Meta Learning Curve")
all_train_loss = [curve[-1] for curve in all_learning_curves]
ax.plot(all_train_loss, 'o-', label="Train Loss")
ax.plot(all_val_loss, 'o-', label="Validation Loss")
ax.plot(all_test_loss, 'o-', label="Test Loss")
ax.plot(all_fakedatasize, 'o-', label="Fake Data Scale")
ax.plot(all_L2, 'o-', label="L2_regularization")
ax.set_ylabel("Validation Loss")
ax.set_xlabel("Meta Iteration Number")
ax.legend(loc=2)
ax = fig.add_subplot(N_figs, 1, 3)
ax.set_title("Fake Data")
images = all_fakedata[-1]
plot_mnist_images(images, ax, ims_per_row=10)
fig.set_size_inches((8,12))
plt.savefig("fig.png")
def plot():
with open('results.pkl') as f:
all_learning_curves, all_val_loss, all_test_loss, all_weights, all_L2 = zip(
*pickle.load(f))
fig = plt.figure(0)
fig.clf()
N_figs = 2
ax = fig.add_subplot(N_figs, 1, 1)
ax.set_title("Learning Curves")
subsample = np.ceil(float(len(all_learning_curves)) / 50)
for i, log_alphas in enumerate(all_learning_curves):
if i % subsample == 0:
ax.plot(log_alphas, 'o-')
ax.set_ylabel("Loss")
ax.set_xlabel("Step number")
#ax.legend(loc=4)
ax = fig.add_subplot(N_figs, 1, 2)
ax.set_title("Meta Learning Curve")
all_train_loss = [curve[-1] for curve in all_learning_curves]
ax.plot(all_train_loss, 'o-', label="Train Loss")
ax.plot(all_val_loss, 'o-', label="Validation Loss")
ax.plot(all_test_loss, 'o-', label="Test Loss")
#ax.plot(all_L2, 'o-', label="L2_regularization")
ax.set_ylabel("Validation Loss")
ax.set_xlabel("Meta Iteration Number")
ax.legend(loc=2)
plt.savefig("fig.png")
fig = plt.figure(0)
fig.clf()
N_figs = 1
ax = fig.add_subplot(N_figs, 1, 1)
images = all_weights[-1].T
plot_mnist_images(images, ax, ims_per_row=10)
fig.set_size_inches((8, 12))
fig.tight_layout()
plt.savefig("weights.png")
plt.savefig("weights.pdf", pad_inches=0.05, bbox_inches='tight')
fig = plt.figure(0)
fig.clf()
N_figs = 1
ax = fig.add_subplot(N_figs, 1, 1)
images = all_L2[-1].T
plot_mnist_images(images, ax, ims_per_row=10)
fig.set_size_inches((8, 12))
fig.tight_layout()
plt.savefig("penalties.png")
plt.savefig("penalties.pdf", pad_inches=0.05, bbox_inches='tight')
def plot():
with open('results.pkl') as f:
all_learning_curves, all_val_loss, all_test_loss, all_weights, all_L2 = zip(*pickle.load(f))
fig = plt.figure(0)
fig.clf()
N_figs = 2
ax = fig.add_subplot(N_figs, 1, 1)
ax.set_title("Learning Curves")
subsample = np.ceil(float(len(all_learning_curves)) / 50)
for i, log_alphas in enumerate(all_learning_curves):
if i % subsample == 0:
ax.plot(log_alphas, 'o-')
ax.set_ylabel("Loss")
ax.set_xlabel("Step number")
#ax.legend(loc=4)
ax = fig.add_subplot(N_figs, 1, 2)
ax.set_title("Meta Learning Curve")
all_train_loss = [curve[-1] for curve in all_learning_curves]
ax.plot(all_train_loss, 'o-', label="Train Loss")
ax.plot(all_val_loss, 'o-', label="Validation Loss")
ax.plot(all_test_loss, 'o-', label="Test Loss")
#ax.plot(all_L2, 'o-', label="L2_regularization")
ax.set_ylabel("Validation Loss")
ax.set_xlabel("Meta Iteration Number")
ax.legend(loc=2)
plt.savefig("fig.png")
fig = plt.figure(0)
fig.clf()
N_figs = 1
ax = fig.add_subplot(N_figs, 1, 1)
images = all_weights[-1].T
plot_mnist_images(images, ax, ims_per_row=10)
fig.set_size_inches((8,12))
fig.tight_layout()
plt.savefig("weights.png")
plt.savefig("weights.pdf", pad_inches=0.05, bbox_inches='tight')
fig = plt.figure(0)
fig.clf()
N_figs = 1
ax = fig.add_subplot(N_figs, 1, 1)
images = all_L2[-1].T
plot_mnist_images(images, ax, ims_per_row=10)
fig.set_size_inches((8,12))
fig.tight_layout()
plt.savefig("penalties.png")
plt.savefig("penalties.pdf", pad_inches=0.05, bbox_inches='tight')
def plot():
with open('results.pkl') as f:
output, fakedata = pickle.load(f)
all_learning_curves, all_val_loss, all_test_loss,\
all_fakedatasize, all_L2, test_err = zip(*output)
# Fake data
fig = plt.figure(0)
fig.clf()
ax = fig.add_subplot(1, 1, 1)
images = fakedata
plot_mnist_images(images, ax, ims_per_row=5)
fig.set_size_inches((8,12))
plt.savefig('fake_data.pdf', pad_inches=0.05, bbox_inches='tight')
fig = plt.figure(0)
fig.clf()
N_figs = 3
ax = fig.add_subplot(N_figs, 1, 1)
ax.set_title("Learning Curves")
subsample = np.ceil(float(len(all_learning_curves)) / 50)
for i, log_alphas in enumerate(all_learning_curves):
if i % subsample == 0:
ax.plot(log_alphas, 'o-')
ax.set_ylabel("Loss")
ax.set_xlabel("Step number")
#ax.legend(loc=4)
ax = fig.add_subplot(N_figs, 1, 2)
ax.set_title("Meta Learning Curve")
all_train_loss = [curve[-1] for curve in all_learning_curves]
ax.plot(all_train_loss, 'o-', label="Train Loss")
ax.plot(all_val_loss, 'o-', label="Validation Loss")
ax.plot(all_test_loss, 'o-', label="Test Loss")
ax.plot(all_fakedatasize, 'o-', label="Fake Data Scale")
ax.plot(all_L2, 'o-', label="L2_regularization")
ax.set_ylabel("Validation Loss")
ax.set_xlabel("Meta Iteration Number")
ax.legend(loc=2)
ax = fig.add_subplot(N_figs, 1, 3)
ax.set_title("Fake Data")
images = fakedata
plot_mnist_images(images, ax, ims_per_row=10)
fig.set_size_inches((8,12))
plt.savefig("fig.png")
def plot():
with open('results.pkl') as f:
all_learning_curves, all_val_loss, all_test_loss,\
all_fakedata, all_fakedatasize = zip(*pickle.load(f))
fig = plt.figure(0)
fig.clf()
N_figs = 3
ax = fig.add_subplot(N_figs, 1, 1)
ax.set_title("Learning Curves")
subsample = np.ceil(float(len(all_learning_curves)) / 50)
for i, log_alphas in enumerate(all_learning_curves):
if i % subsample == 0:
ax.plot(log_alphas, 'o-')
ax.set_ylabel("Loss")
ax.set_xlabel("Step number")
#ax.legend(loc=4)
ax = fig.add_subplot(N_figs, 1, 2)
ax.set_title("Meta Learning Curve")
all_train_loss = [curve[-1] for curve in all_learning_curves]
ax.plot(all_train_loss, 'o-', label="Train Loss")
ax.plot(all_val_loss, 'o-', label="Validation Loss")
ax.plot(all_test_loss, 'o-', label="Test Loss")
ax.plot(all_fakedatasize, 'o-', label="Fake Data Scale")
ax.set_ylabel("Validation Loss")
ax.set_xlabel("Meta Iteration Number")
ax.legend(loc=2)
ax = fig.add_subplot(N_figs, 1, 3)
ax.set_title("Fake Data")
images = all_fakedata[-1]
plot_mnist_images(images, ax, ims_per_row=10)
fig.set_size_inches((8, 12))
plt.savefig("/tmp/fig.png")
plt.savefig("fig.png")
plt.show()
def plot():
with open('results.pkl') as f:
all_learning_curves, all_val_loss, all_fakedata = zip(*pickle.load(f))
fig = plt.figure(0)
fig.clf()
N_figs = 3
ax = fig.add_subplot(N_figs, 1, 1)
ax.set_title("Learning Curves")
subsample = len(all_learning_curves) / 50
for i, log_alphas in enumerate(all_learning_curves):
if i % subsample == 0:
ax.plot(log_alphas, 'o-')
ax.set_ylabel("Loss")
ax.set_xlabel("Step number")
#ax.legend(loc=4)
ax = fig.add_subplot(N_figs, 1, 2)
ax.set_title("Meta Learning Curve")
all_train_loss = [curve[-1] for curve in all_learning_curves]
ax.plot(all_train_loss, 'o-', label="Train Loss")
ax.plot(all_val_loss, 'o-', label="Validation Loss")
ax.set_ylabel("Validation Loss")
ax.set_xlabel("Meta Iteration Number")
ax.legend(loc=2)
ax = fig.add_subplot(N_figs, 1, 3)
ax.set_title("Fake Data")
images = all_fakedata[-1]
plot_mnist_images(images, ax, ims_per_row=10)
fig.set_size_inches((8,12))
plt.savefig("/tmp/fig.png")
plt.savefig("fig.png")
plt.show()
def plot():
import matplotlib.pyplot as plt
with open('results.pkl') as f:
A, valid_losses, test_losses = pickle.load(f)
fig = plt.figure(0)
fig.clf()
fig.set_size_inches((8, 12))
ax = fig.add_subplot(211)
ax.set_title("Meta learning curves")
ax.plot(valid_losses, 'o-', label="Validation")
ax.plot(test_losses, 'o-', label="Test")
ax.set_ylabel("Negative log prob")
ax.set_xlabel("Step number")
ax.legend(loc=1, frameon=False)
ax = fig.add_subplot(212)
test_images = build_test_images()
transformed_images = np.dot(test_images, A.T)
cat_images = np.concatenate((test_images, transformed_images))
plot_mnist_images(cat_images, ax, ims_per_row=test_images.shape[0])
plt.savefig("/tmp/fig.png")
plt.savefig("fig.png")
def plot():
import matplotlib.pyplot as plt
with open('results.pkl') as f:
A, valid_losses, test_losses = pickle.load(f)
fig = plt.figure(0)
fig.clf()
fig.set_size_inches((8,12))
ax = fig.add_subplot(211)
ax.set_title("Meta learning curves")
ax.plot(valid_losses, 'o-', label="Validation")
ax.plot(test_losses , 'o-', label="Test")
ax.set_ylabel("Negative log prob")
ax.set_xlabel("Step number")
ax.legend(loc=1, frameon=False)
ax = fig.add_subplot(212)
test_images = build_test_images()
transformed_images = np.dot(test_images, A.T)
cat_images = np.concatenate((test_images, transformed_images))
plot_mnist_images(cat_images, ax, ims_per_row=test_images.shape[0])
plt.savefig("/tmp/fig.png")
plt.savefig("fig.png")
def plot():
import matplotlib.pyplot as plt
with open('results.pkl') as f:
results, parser, parsed_init_hypergrad = pickle.load(f)
# Show final ARD in the first layer filters.
L2s = results['log_L2_reg']
L2parser = parser.empty_copy()
L2parser.vect = L2s
l2_images = L2parser[('weights', 0)].T
fig = plt.figure(0)
fig.clf()
fig.set_size_inches((6, 8))
ax = fig.add_subplot(111)
plot_mnist_images(l2_images, ax, ims_per_row=30, padding=0)
fig.set_size_inches((8, 12))
#fig.tight_layout()
plt.savefig("penalties.png")
plt.savefig("penalties.pdf", pad_inches=0.05, bbox_inches='tight')
fig = plt.figure(0)
fig.clf()
fig.set_size_inches((6, 8))
ax = fig.add_subplot(111)
plt.hist(l2_images.ravel(), 100)
plt.savefig("penalty_histogram.png")
# Show first layer filters from the last meta-iteration.
weights = results['example_weights']
parser.vect = weights
weight_images = parser[('weights', 0)].T
fig = plt.figure(0)
fig.clf()
fig.set_size_inches((6, 8))
ax = fig.add_subplot(111)
plot_mnist_images(weight_images, ax, ims_per_row=30, padding=0)
fig.set_size_inches((8, 12))
#fig.tight_layout()
plt.savefig("weights.png")
plt.savefig("weights.pdf", pad_inches=0.05, bbox_inches='tight')
fig = plt.figure(0)
fig.clf()
fig.set_size_inches((6, 8))
# ----- Primal learning curves -----
ax = fig.add_subplot(311)
ax.set_title('Primal learning curves')
for i, y in enumerate(results['learning_curves']):
ax.plot(y['learning_curve'], 'o-', label='Meta iter {0}'.format(i))
ax.set_xlabel('Epoch number')
ax.set_ylabel('Negative log prob')
#ax.legend(loc=1, frameon=False)
ax = fig.add_subplot(312)
ax.set_title('Meta learning curves')
losses = ['train_loss', 'valid_loss', 'tests_loss']
for loss_type in losses:
ax.plot(results[loss_type], 'o-', label=loss_type)
ax.set_xlabel('Meta iter number')
ax.set_ylabel('Negative log prob')
ax.legend(loc=1, frameon=False)
ax = fig.add_subplot(313)
ax.set_title('Meta-gradient magnitude')
ax.plot(results['meta_grad_magnitude'],
'o-',
label='Meta-gradient magnitude')
ax.plot(results['meta_grad_angle'], 'o-', label='Meta-gradient angle')
ax.set_xlabel('Meta iter number')
ax.set_ylabel('Meta-gradient Magnitude')
ax.legend(loc=1, frameon=False)
plt.savefig('learning_curves.png')
# ----- Learning curve info -----
fig.clf()
ax = fig.add_subplot(311)
ax.set_title('Primal learning curves')
for i, y in enumerate(results['learning_curves']):
ax.plot(y['grad_norm'], 'o-', label='Meta iter {0}'.format(i))
ax.set_xlabel('Epoch number')
#ax.legend(loc=1, frameon=False)
ax.set_title('Grad norm')
ax = fig.add_subplot(312)
for i, y in enumerate(results['learning_curves']):
ax.plot(y['weight_norm'], 'o-', label='Meta iter {0}'.format(i))
ax.set_xlabel('Epoch number')
ax.legend(loc=1, frameon=False)
ax.set_title('Weight norm')
ax = fig.add_subplot(313)
for i, y in enumerate(results['learning_curves']):
ax.plot(y['velocity_norm'], 'o-', label='Meta iter {0}'.format(i))
ax.set_xlabel('Epoch number')
ax.set_title('Velocity norm')
ax.legend(loc=1, frameon=False)
plt.savefig('extra_learning_curves.png')
def plot():
with open('results.pkl') as f:
all_learning_curves, all_val_loss, all_test_loss, all_weights, all_L2 = zip(*pickle.load(f))
fig = plt.figure(0)
fig.clf()
N_figs = 4
ax = fig.add_subplot(N_figs, 1, 1)
ax.set_title("Learning Curves")
subsample = np.ceil(float(len(all_learning_curves)) / 50)
for i, log_alphas in enumerate(all_learning_curves):
if i % subsample == 0:
ax.plot(log_alphas, 'o-')
ax.set_ylabel("Loss")
ax.set_xlabel("Step number")
#ax.legend(loc=4)
ax = fig.add_subplot(N_figs, 1, 2)
ax.set_title("Meta Learning Curve")
all_train_loss = [curve[-1] for curve in all_learning_curves]
ax.plot(all_train_loss, 'o-', label="Train Loss")
ax.plot(all_val_loss, 'o-', label="Validation Loss")
ax.plot(all_test_loss, 'o-', label="Test Loss")
#ax.plot(all_L2, 'o-', label="L2_regularization")
ax.set_ylabel("Validation Loss")
ax.set_xlabel("Meta Iteration Number")
ax.legend(loc=2)
ax = fig.add_subplot(N_figs, 1, 3)
ax.set_title("Weights")
images = all_weights[-1].T
plot_mnist_images(images, ax, ims_per_row=10)
fig.set_size_inches((8,12))
ax = fig.add_subplot(N_figs, 1, 4)
ax.set_title("Per-weight L2 penalty")
images = all_L2[-1].T
plot_mnist_images(images, ax, ims_per_row=10)
fig.set_size_inches((8,12))
plt.savefig("fig.png")
fig = plt.figure(0)
fig.clf()
N_figs = 1
ax = fig.add_subplot(N_figs, 1, 1)
#ax.set_title("Weights")
images = all_weights[-1].T
plot_mnist_images(images, ax, ims_per_row=10)
fig.set_size_inches((8,12))
fig.tight_layout()
plt.savefig("weights.png")
plt.savefig("weights.pdf", pad_inches=0.05, bbox_inches='tight')
from matplotlib import rc
rc('font',**{'family':'serif'})
# Show ARD penalties.
fig = plt.figure(0)
plt.clf()
ax = plt.Axes(fig, [0., 0., 1., 1.])
#ax = fig.add_subplot(1, 1, 1)
fig.set_size_inches((4.5,1.75))
ax.set_axis_off()
fig.add_axes(ax)
images = all_L2[-1].T.copy() # Get rid of spikes
newmax = np.percentile(images.ravel(), 98.0)
over_ixs = images > newmax
images[over_ixs] = newmax
cax = plot_mnist_images(images, ax, ims_per_row=5, padding=2, vmin=0.0)
cbar = fig.colorbar(cax, ticks=[0, newmax], shrink=.7)
cbar.ax.set_yticklabels(['0', '{:2.2f}'.format(newmax)])
plt.savefig("penalties.png")
plt.savefig("penalties.pdf", bbox_inches='tight')
def plot():
import matplotlib.pyplot as plt
with open('results.pkl') as f:
results, parser = pickle.load(f)
# Fake data
fig = plt.figure(0)
fig.clf()
ax = fig.add_subplot(1, 1, 1)
ax.set_title("Fake Data")
images = results['fake_data'][-1]
plot_mnist_images(images, ax, ims_per_row=10)
fig.set_size_inches((8,12))
plt.savefig('fake_data.pdf', pad_inches=0.05, bbox_inches='tight')
# Show first layer filters from the last meta-iteration.
fig = plt.figure(0)
fig.clf()
ax = fig.add_subplot(1, 1, 1)
ax.set_title("Weights")
weights = results['example_weights']
parser.vect = weights
weight_images = parser[('weights', 0)].T
plot_mnist_images(weight_images, ax, ims_per_row=10)
fig.set_size_inches((8,12))
plt.savefig('first_layer_weights.pdf', pad_inches=0.05, bbox_inches='tight')
# ----- Nice versions of Alpha and beta schedules for paper -----
fig = plt.figure(0)
fig.clf()
ax = fig.add_subplot(411)
#ax.set_title('Alpha learning curves')
for cur_results, name in zip(results['log_alphas'][-1].T, parser.names):
if name[0] == 'weights':
ax.plot(np.exp(cur_results), 'o-', label=name)
#ax.set_xlabel('Learning Iteration', fontproperties='serif')
low, high = ax.get_ylim()
ax.set_ylim([0, high])
ax.set_ylabel('Step size', fontproperties='serif')
ax.set_xticklabels([])
ax.legend(numpoints=1, loc=1, frameon=False, bbox_to_anchor=(1.0, 0.5),
prop={'family':'serif', 'size':'12'})
ax = fig.add_subplot(412)
#ax.set_title('Alpha learning curves')
for cur_results, name in zip(results['invlogit_betas'][-1].T, parser.names):
if name[0] == 'weights':
ax.plot(logit(cur_results), 'o-', label=name)
low, high = ax.get_ylim()
ax.set_ylim([0, 1])
ax.set_xlabel('Learning Iteration', fontproperties='serif')
ax.set_ylabel('Momentum', fontproperties='serif')
ax = fig.add_subplot(413)
#ax.set_title('Alpha learning curves')
for cur_results, name in zip(results['log_alphas'][-1].T, parser.names):
if name[0] == 'biases':
ax.plot(np.exp(cur_results), 'o-', label=name)
#ax.set_xlabel('Learning Iteration', fontproperties='serif')
low, high = ax.get_ylim()
ax.set_ylim([0, high])
ax.set_ylabel('Step size', fontproperties='serif')
ax.set_xticklabels([])
ax.legend(numpoints=1, loc=1, frameon=False, bbox_to_anchor=(1.0, 0.5),
prop={'family':'serif', 'size':'12'})
ax = fig.add_subplot(414)
#ax.set_title('Alpha learning curves')
for cur_results, name in zip(results['invlogit_betas'][-1].T, parser.names):
if name[0] == 'biases':
ax.plot(logit(cur_results), 'o-', label=name)
low, high = ax.get_ylim()
ax.set_ylim([0, 1])
ax.set_xlabel('Learning Iteration', fontproperties='serif')
ax.set_ylabel('Momentum', fontproperties='serif')
fig.set_size_inches((6,8))
#plt.show()
plt.savefig('alpha_beta_paper.png')
plt.savefig('alpha_beta_paper.pdf', pad_inches=0.05, bbox_inches='tight')
fig.clf()
fig.set_size_inches((6,8))
# ----- Primal learning curves -----
ax = fig.add_subplot(311)
#ax.set_title('Primal learning curves')
#for i, y in enumerate(results['learning_curves']):
# ax.plot(y['learning_curve'], 'o-', label='Meta iter {0}'.format(i))
#ax.set_xlabel('Epoch number')
#ax.set_ylabel('Negative log prob')
#ax.legend(loc=1, frameon=False)
ax = fig.add_subplot(312)
ax.set_title('Meta learning curves')
losses = ['train_loss', 'valid_loss', 'tests_loss']
for loss_type in losses:
ax.plot(results[loss_type], 'o-', label=loss_type)
ax.set_xlabel('Meta iter number')
ax.set_ylabel('Negative log prob')
ax.legend(loc=1, frameon=False)
ax = fig.add_subplot(313)
ax.set_title('Meta-gradient magnitude')
ax.plot(results['meta_grad_magnitude'], 'o-', label='Meta-gradient magnitude')
ax.plot(results['meta_grad_angle'], 'o-', label='Meta-gradient angle')
ax.set_xlabel('Meta iter number')
ax.set_ylabel('Meta-gradient Magnitude')
ax.legend(loc=1, frameon=False)
plt.savefig('learning_curves.png')
# ----- Learning curve info -----
fig.clf()
ax = fig.add_subplot(311)
ax.set_title('Primal learning curves')
for i, y in enumerate(results['learning_curves']):
ax.plot(y['grad_norm'], 'o-', label='Meta iter {0}'.format(i))
ax.set_xlabel('Epoch number')
#ax.legend(loc=1, frameon=False)
ax.set_title('Grad norm')
ax = fig.add_subplot(312)
for i, y in enumerate(results['learning_curves']):
ax.plot(y['weight_norm'], 'o-', label='Meta iter {0}'.format(i))
ax.set_xlabel('Epoch number')
ax.legend(loc=1, frameon=False)
ax.set_title('Weight norm')
ax = fig.add_subplot(313)
for i, y in enumerate(results['learning_curves']):
ax.plot(y['velocity_norm'], 'o-', label='Meta iter {0}'.format(i))
ax.set_xlabel('Epoch number')
ax.set_title('Velocity norm')
ax.legend(loc=1, frameon=False)
plt.savefig('extra_learning_curves.png')
# ----- Alpha and beta schedules -----
fig.clf()
ax = fig.add_subplot(211)
ax.set_title('Alpha learning curves')
for i, y in enumerate(results['log_alphas']):
ax.plot(y, 'o-', label="Meta iter {0}".format(i))
ax.set_xlabel('Primal iter number')
#ax.set_ylabel('Log alpha')
ax.legend(loc=1, frameon=False)
ax = fig.add_subplot(212)
ax.set_title('Beta learning curves')
for y in results['invlogit_betas']:
ax.plot(y, 'o-')
ax.set_xlabel('Primal iter number')
ax.set_ylabel('Inv logit beta')
plt.savefig('alpha_beta_curves.png')
# ----- Init scale and L2 reg -----
fig.clf()
ax = fig.add_subplot(211)
ax.set_title('Init scale learning curves')
for i, y in enumerate(zip(*results['log_param_scale'])):
if parser.names[i][0] == 'weights':
ax.plot(y, 'o-', label=parser.names[i])
ax.set_xlabel('Meta iter number')
ax.set_ylabel('Log param scale')
ax.legend(loc=1, frameon=False)
ax = fig.add_subplot(212)
ax.set_title('Init scale learning curves')
for i, y in enumerate(zip(*results['log_param_scale'])):
if parser.names[i][0] == 'biases':
ax.plot(y, 'o-', label=parser.names[i])
ax.set_xlabel('Meta iter number')
ax.set_ylabel('Log param scale')
ax.legend(loc=1, frameon=False)
plt.savefig('scale_and_reg.png')
def plot():
import matplotlib.pyplot as plt
with open('results.pkl') as f:
results, parser, parsed_init_hypergrad = pickle.load(f)
# Show final ARD in the first layer filters.
L2s = results['log_L2_reg']
L2parser = parser.empty_copy()
L2parser.vect = L2s
l2_images = L2parser[('weights', 0)].T
fig = plt.figure(0)
fig.clf()
fig.set_size_inches((6,8))
ax = fig.add_subplot(111)
plot_mnist_images(l2_images, ax, ims_per_row=30, padding=0)
fig.set_size_inches((8,12))
#fig.tight_layout()
plt.savefig("penalties.png")
plt.savefig("penalties.pdf", pad_inches=0.05, bbox_inches='tight')
fig = plt.figure(0)
fig.clf()
fig.set_size_inches((6,8))
ax = fig.add_subplot(111)
plt.hist(l2_images.ravel(), 100)
plt.savefig("penalty_histogram.png")
# Show first layer filters from the last meta-iteration.
weights = results['example_weights']
parser.vect = weights
weight_images = parser[('weights', 0)].T
fig = plt.figure(0)
fig.clf()
fig.set_size_inches((6,8))
ax = fig.add_subplot(111)
plot_mnist_images(weight_images, ax, ims_per_row=30, padding=0)
fig.set_size_inches((8,12))
#fig.tight_layout()
plt.savefig("weights.png")
plt.savefig("weights.pdf", pad_inches=0.05, bbox_inches='tight')
fig = plt.figure(0)
fig.clf()
fig.set_size_inches((6,8))
# ----- Primal learning curves -----
ax = fig.add_subplot(311)
ax.set_title('Primal learning curves')
for i, y in enumerate(results['learning_curves']):
ax.plot(y['learning_curve'], 'o-', label='Meta iter {0}'.format(i))
ax.set_xlabel('Epoch number')
ax.set_ylabel('Negative log prob')
#ax.legend(loc=1, frameon=False)
ax = fig.add_subplot(312)
ax.set_title('Meta learning curves')
losses = ['train_loss', 'valid_loss', 'tests_loss']
for loss_type in losses:
ax.plot(results[loss_type], 'o-', label=loss_type)
ax.set_xlabel('Meta iter number')
ax.set_ylabel('Negative log prob')
ax.legend(loc=1, frameon=False)
ax = fig.add_subplot(313)
ax.set_title('Meta-gradient magnitude')
ax.plot(results['meta_grad_magnitude'], 'o-', label='Meta-gradient magnitude')
ax.plot(results['meta_grad_angle'], 'o-', label='Meta-gradient angle')
ax.set_xlabel('Meta iter number')
ax.set_ylabel('Meta-gradient Magnitude')
ax.legend(loc=1, frameon=False)
plt.savefig('learning_curves.png')
# ----- Learning curve info -----
fig.clf()
ax = fig.add_subplot(311)
ax.set_title('Primal learning curves')
for i, y in enumerate(results['learning_curves']):
ax.plot(y['grad_norm'], 'o-', label='Meta iter {0}'.format(i))
ax.set_xlabel('Epoch number')
#ax.legend(loc=1, frameon=False)
ax.set_title('Grad norm')
ax = fig.add_subplot(312)
for i, y in enumerate(results['learning_curves']):
ax.plot(y['weight_norm'], 'o-', label='Meta iter {0}'.format(i))
ax.set_xlabel('Epoch number')
ax.legend(loc=1, frameon=False)
ax.set_title('Weight norm')
ax = fig.add_subplot(313)
for i, y in enumerate(results['learning_curves']):
ax.plot(y['velocity_norm'], 'o-', label='Meta iter {0}'.format(i))
ax.set_xlabel('Epoch number')
ax.set_title('Velocity norm')
ax.legend(loc=1, frameon=False)
plt.savefig('extra_learning_curves.png')
def plot():
import matplotlib.pyplot as plt
with open('results.pkl') as f:
results, parser = pickle.load(f)
# Fake data
fig = plt.figure(0)
fig.clf()
ax = fig.add_subplot(1, 1, 1)
ax.set_title("Fake Data")
images = results['fake_data'][-1]
plot_mnist_images(images, ax, ims_per_row=10)
fig.set_size_inches((8, 12))
plt.savefig('fake_data.pdf', pad_inches=0.05, bbox_inches='tight')
# Show first layer filters from the last meta-iteration.
fig = plt.figure(0)
fig.clf()
ax = fig.add_subplot(1, 1, 1)
ax.set_title("Weights")
weights = results['example_weights']
parser.vect = weights
weight_images = parser[('weights', 0)].T
plot_mnist_images(weight_images, ax, ims_per_row=10)
fig.set_size_inches((8, 12))
plt.savefig('first_layer_weights.pdf',
pad_inches=0.05,
bbox_inches='tight')
# ----- Nice versions of Alpha and beta schedules for paper -----
fig = plt.figure(0)
fig.clf()
ax = fig.add_subplot(411)
#ax.set_title('Alpha learning curves')
for cur_results, name in zip(results['log_alphas'][-1].T, parser.names):
if name[0] == 'weights':
ax.plot(np.exp(cur_results), 'o-', label=name)
#ax.set_xlabel('Learning Iteration', fontproperties='serif')
low, high = ax.get_ylim()
ax.set_ylim([0, high])
ax.set_ylabel('Step size', fontproperties='serif')
ax.set_xticklabels([])
ax.legend(numpoints=1,
loc=1,
frameon=False,
bbox_to_anchor=(1.0, 0.5),
prop={
'family': 'serif',
'size': '12'
})
ax = fig.add_subplot(412)
#ax.set_title('Alpha learning curves')
for cur_results, name in zip(results['invlogit_betas'][-1].T,
parser.names):
if name[0] == 'weights':
ax.plot(logit(cur_results), 'o-', label=name)
low, high = ax.get_ylim()
ax.set_ylim([0, 1])
ax.set_xlabel('Learning Iteration', fontproperties='serif')
ax.set_ylabel('Momentum', fontproperties='serif')
ax = fig.add_subplot(413)
#ax.set_title('Alpha learning curves')
for cur_results, name in zip(results['log_alphas'][-1].T, parser.names):
if name[0] == 'biases':
ax.plot(np.exp(cur_results), 'o-', label=name)
#ax.set_xlabel('Learning Iteration', fontproperties='serif')
low, high = ax.get_ylim()
ax.set_ylim([0, high])
ax.set_ylabel('Step size', fontproperties='serif')
ax.set_xticklabels([])
ax.legend(numpoints=1,
loc=1,
frameon=False,
bbox_to_anchor=(1.0, 0.5),
prop={
'family': 'serif',
'size': '12'
})
ax = fig.add_subplot(414)
#ax.set_title('Alpha learning curves')
for cur_results, name in zip(results['invlogit_betas'][-1].T,
parser.names):
if name[0] == 'biases':
ax.plot(logit(cur_results), 'o-', label=name)
low, high = ax.get_ylim()
ax.set_ylim([0, 1])
ax.set_xlabel('Learning Iteration', fontproperties='serif')
ax.set_ylabel('Momentum', fontproperties='serif')
fig.set_size_inches((6, 8))
#plt.show()
plt.savefig('alpha_beta_paper.png')
plt.savefig('alpha_beta_paper.pdf', pad_inches=0.05, bbox_inches='tight')
fig.clf()
fig.set_size_inches((6, 8))
# ----- Primal learning curves -----
ax = fig.add_subplot(311)
#ax.set_title('Primal learning curves')
#for i, y in enumerate(results['learning_curves']):
# ax.plot(y['learning_curve'], 'o-', label='Meta iter {0}'.format(i))
#ax.set_xlabel('Epoch number')
#ax.set_ylabel('Negative log prob')
#ax.legend(loc=1, frameon=False)
ax = fig.add_subplot(312)
ax.set_title('Meta learning curves')
losses = ['train_loss', 'valid_loss', 'tests_loss']
for loss_type in losses:
ax.plot(results[loss_type], 'o-', label=loss_type)
ax.set_xlabel('Meta iter number')
ax.set_ylabel('Negative log prob')
ax.legend(loc=1, frameon=False)
ax = fig.add_subplot(313)
ax.set_title('Meta-gradient magnitude')
ax.plot(results['meta_grad_magnitude'],
'o-',
label='Meta-gradient magnitude')
ax.plot(results['meta_grad_angle'], 'o-', label='Meta-gradient angle')
ax.set_xlabel('Meta iter number')
ax.set_ylabel('Meta-gradient Magnitude')
ax.legend(loc=1, frameon=False)
plt.savefig('learning_curves.png')
# ----- Learning curve info -----
fig.clf()
ax = fig.add_subplot(311)
ax.set_title('Primal learning curves')
for i, y in enumerate(results['learning_curves']):
ax.plot(y['grad_norm'], 'o-', label='Meta iter {0}'.format(i))
ax.set_xlabel('Epoch number')
#ax.legend(loc=1, frameon=False)
ax.set_title('Grad norm')
ax = fig.add_subplot(312)
for i, y in enumerate(results['learning_curves']):
ax.plot(y['weight_norm'], 'o-', label='Meta iter {0}'.format(i))
ax.set_xlabel('Epoch number')
ax.legend(loc=1, frameon=False)
ax.set_title('Weight norm')
ax = fig.add_subplot(313)
for i, y in enumerate(results['learning_curves']):
ax.plot(y['velocity_norm'], 'o-', label='Meta iter {0}'.format(i))
ax.set_xlabel('Epoch number')
ax.set_title('Velocity norm')
ax.legend(loc=1, frameon=False)
plt.savefig('extra_learning_curves.png')
# ----- Alpha and beta schedules -----
fig.clf()
ax = fig.add_subplot(211)
ax.set_title('Alpha learning curves')
for i, y in enumerate(results['log_alphas']):
ax.plot(y, 'o-', label="Meta iter {0}".format(i))
ax.set_xlabel('Primal iter number')
#ax.set_ylabel('Log alpha')
ax.legend(loc=1, frameon=False)
ax = fig.add_subplot(212)
ax.set_title('Beta learning curves')
for y in results['invlogit_betas']:
ax.plot(y, 'o-')
ax.set_xlabel('Primal iter number')
ax.set_ylabel('Inv logit beta')
plt.savefig('alpha_beta_curves.png')
# ----- Init scale and L2 reg -----
fig.clf()
ax = fig.add_subplot(211)
ax.set_title('Init scale learning curves')
for i, y in enumerate(zip(*results['log_param_scale'])):
if parser.names[i][0] == 'weights':
ax.plot(y, 'o-', label=parser.names[i])
ax.set_xlabel('Meta iter number')
ax.set_ylabel('Log param scale')
ax.legend(loc=1, frameon=False)
ax = fig.add_subplot(212)
ax.set_title('Init scale learning curves')
for i, y in enumerate(zip(*results['log_param_scale'])):
if parser.names[i][0] == 'biases':
ax.plot(y, 'o-', label=parser.names[i])
ax.set_xlabel('Meta iter number')
ax.set_ylabel('Log param scale')
ax.legend(loc=1, frameon=False)
plt.savefig('scale_and_reg.png')
def plot():
with open("results.pkl") as f:
all_learning_curves, all_val_loss, all_test_loss, all_weights, all_L2 = zip(*pickle.load(f))
fig = plt.figure(0)
fig.clf()
N_figs = 4
ax = fig.add_subplot(N_figs, 1, 1)
ax.set_title("Learning Curves")
subsample = np.ceil(float(len(all_learning_curves)) / 50)
for i, log_alphas in enumerate(all_learning_curves):
if i % subsample == 0:
ax.plot(log_alphas, "o-")
ax.set_ylabel("Loss")
ax.set_xlabel("Step number")
# ax.legend(loc=4)
ax = fig.add_subplot(N_figs, 1, 2)
ax.set_title("Meta Learning Curve")
all_train_loss = [curve[-1] for curve in all_learning_curves]
ax.plot(all_train_loss, "o-", label="Train Loss")
ax.plot(all_val_loss, "o-", label="Validation Loss")
ax.plot(all_test_loss, "o-", label="Test Loss")
# ax.plot(all_L2, 'o-', label="L2_regularization")
ax.set_ylabel("Validation Loss")
ax.set_xlabel("Meta Iteration Number")
ax.legend(loc=2)
ax = fig.add_subplot(N_figs, 1, 3)
ax.set_title("Weights")
images = all_weights[-1].T
plot_mnist_images(images, ax, ims_per_row=10)
fig.set_size_inches((8, 12))
ax = fig.add_subplot(N_figs, 1, 4)
ax.set_title("Per-weight L2 penalty")
images = all_L2[-1].T
plot_mnist_images(images, ax, ims_per_row=10)
fig.set_size_inches((8, 12))
plt.savefig("fig.png")
fig = plt.figure(0)
fig.clf()
N_figs = 1
ax = fig.add_subplot(N_figs, 1, 1)
# ax.set_title("Weights")
images = all_weights[-1].T
plot_mnist_images(images, ax, ims_per_row=10)
fig.set_size_inches((8, 12))
fig.tight_layout()
plt.savefig("weights.png")
plt.savefig("weights.pdf", pad_inches=0.05, bbox_inches="tight")
from matplotlib import rc
rc("font", **{"family": "serif"})
# Show ARD penalties.
fig = plt.figure(0)
plt.clf()
ax = plt.Axes(fig, [0.0, 0.0, 1.0, 1.0])
# ax = fig.add_subplot(1, 1, 1)
fig.set_size_inches((4.5, 1.75))
ax.set_axis_off()
fig.add_axes(ax)
images = all_L2[-1].T.copy() # Get rid of spikes
newmax = np.percentile(images.ravel(), 98.0)
over_ixs = images > newmax
images[over_ixs] = newmax
cax = plot_mnist_images(images, ax, ims_per_row=5, padding=2, vmin=0.0)
cbar = fig.colorbar(cax, ticks=[0, newmax], shrink=0.7)
cbar.ax.set_yticklabels(["0", "{:2.2f}".format(newmax)])
plt.savefig("penalties.png")
plt.savefig("penalties.pdf", bbox_inches="tight")