def plot():
import matplotlib.pyplot as plt
import matplotlib as mpl
mpl.rcParams['font.family'] = 'serif'
mpl.rcParams['image.interpolation'] = 'none'
with open('results.pkl') as f:
transform_parser, transform_vects, train_losses, tests_losses = pickle.load(f)
RS = RandomState((seed, "plotting"))
fig = plt.figure(0)
fig.clf()
ax = fig.add_subplot(111)
omniglot.show_alphabets(omniglot.load_rotated_alphabets(RS, normalize=False, angle=90), ax=ax)
ax.plot([0, 20 * 28], [5 * 28, 5 * 28], '--k')
ax.text(-15, 5 * 28 * 3 / 2 - 60, "Rotated alphabets", rotation='vertical')
plt.savefig("all_alphabets.png")
# Plotting transformations
names = ['no_sharing', 'full_sharing', 'learned_sharing']
title_strings = {'no_sharing' : 'Independent nets',
'full_sharing' : 'Shared bottom layer',
'learned_sharing' : 'Learned sharing'}
covar_imgs = {name : build_covar_image(transform_vects[name]) for name in names}
for i, name in enumerate(names):
fig = plt.figure(0)
fig.clf()
fig.set_size_inches((2, 6))
ax = fig.add_subplot(111)
ax.matshow(covar_imgs[name], cmap = mpl.cm.binary)
ax.set_xticks([])
ax.set_yticks([])
plt.savefig('learned_corr_{0}.png'.format(i))
plt.savefig('learned_corr_{0}.pdf'.format(i))
def plot():
import matplotlib.pyplot as plt
import matplotlib as mpl
mpl.rcParams['font.family'] = 'serif'
mpl.rcParams['image.interpolation'] = 'none'
with open('results.pkl') as f:
transform_parser, transform_vects, train_losses, tests_losses = pickle.load(
f)
RS = RandomState((seed, "plotting"))
fig = plt.figure(0)
fig.clf()
ax = fig.add_subplot(111)
alphabets = omniglot.load_rotated_alphabets(RS, normalize=False, angle=90)
num_cols = 15
num_rows = 5
omniglot.show_alphabets(alphabets, ax=ax, n_cols=num_cols)
ax.plot([0, num_cols * 28], [num_rows * 28, num_rows * 28], '--k')
#ax.text(-15, 5 * 28 * 3 / 2 - 60, "Rotated alphabets", rotation='vertical')
plt.savefig("all_alphabets.png", bbox_inches='tight')
# Plotting transformations
names = ['no_sharing', 'full_sharing', 'learned_sharing']
title_strings = {
'no_sharing': 'Independent nets',
'full_sharing': 'Shared bottom layer',
'learned_sharing': 'Learned sharing'
}
covar_imgs = {
name: build_covar_image(transform_vects[name])
for name in names
}
for model_ix, model_name in enumerate(names):
image_list = covar_imgs[model_name]
for layer_ix, image in enumerate(image_list):
fig = plt.figure(0)
fig.clf()
fig.set_size_inches((1, 1))
ax = fig.add_subplot(111)
ax.matshow(image, cmap=mpl.cm.binary, vmin=0.0, vmax=1.0)
ax.set_xticks([])
ax.set_yticks([])
plt.savefig('minifigs/learned_corr_{0}_{1}.png'.format(
model_name, layer_ix),
bbox_inches='tight')
plt.savefig('minifigs/learned_corr_{0}_{1}.pdf'.format(
model_name, layer_ix),
bbox_inches='tight')
# Write results to a nice latex table for paper.
with open('results_table.tex', 'w') as f:
f.write(" & No Sharing & Full Sharing & Learned \\\\\n")
f.write("Training loss & {:2.2f} & {:2.2f} & {:2.2f} \\\\\n".format(
train_losses['no_sharing'], train_losses['full_sharing'],
train_losses['learned_sharing']))
f.write("Test loss & {:2.2f} & {:2.2f} & \\bf {:2.2f} ".format(
tests_losses['no_sharing'], tests_losses['full_sharing'],
tests_losses['learned_sharing']))
def plot():
import matplotlib.pyplot as plt
import matplotlib as mpl
mpl.rcParams["font.family"] = "serif"
mpl.rcParams["image.interpolation"] = "none"
with open("results.pkl") as f:
transform_parser, transform_vects, train_losses, tests_losses = pickle.load(f)
RS = RandomState((seed, "plotting"))
fig = plt.figure(0)
fig.clf()
ax = fig.add_subplot(111)
alphabets = omniglot.load_rotated_alphabets(RS, normalize=False, angle=90)
num_cols = 15
num_rows = 5
omniglot.show_alphabets(alphabets, ax=ax, n_cols=num_cols)
ax.plot([0, num_cols * 28], [num_rows * 28, num_rows * 28], "--k")
# ax.text(-15, 5 * 28 * 3 / 2 - 60, "Rotated alphabets", rotation='vertical')
plt.savefig("all_alphabets.png", bbox_inches="tight")
# Plotting transformations
names = ["no_sharing", "full_sharing", "learned_sharing"]
title_strings = {
"no_sharing": "Independent nets",
"full_sharing": "Shared bottom layer",
"learned_sharing": "Learned sharing",
}
covar_imgs = {name: build_covar_image(transform_vects[name]) for name in names}
for model_ix, model_name in enumerate(names):
image_list = covar_imgs[model_name]
for layer_ix, image in enumerate(image_list):
fig = plt.figure(0)
fig.clf()
fig.set_size_inches((1, 1))
ax = fig.add_subplot(111)
ax.matshow(image, cmap=mpl.cm.binary, vmin=0.0, vmax=1.0)
ax.set_xticks([])
ax.set_yticks([])
plt.savefig("minifigs/learned_corr_{0}_{1}.png".format(model_name, layer_ix), bbox_inches="tight")
plt.savefig("minifigs/learned_corr_{0}_{1}.pdf".format(model_name, layer_ix), bbox_inches="tight")
# Write results to a nice latex table for paper.
with open("results_table.tex", "w") as f:
f.write(" & No Sharing & Full Sharing & Learned \\\\\n")
f.write(
"Training loss & {:2.2f} & {:2.2f} & {:2.2f} \\\\\n".format(
train_losses["no_sharing"], train_losses["full_sharing"], train_losses["learned_sharing"]
)
)
f.write(
"Test loss & {:2.2f} & {:2.2f} & \\bf {:2.2f} ".format(
tests_losses["no_sharing"], tests_losses["full_sharing"], tests_losses["learned_sharing"]
)
)
def run():
RS = RandomState((seed, "top_rs"))
all_data = omniglot.load_rotated_alphabets(RS)
train_data, tests_data = random_partition(all_data, RS, [12, 3])
w_parser, pred_fun, loss_fun, frac_err = make_nn_funs(layer_sizes)
N_weights = w_parser.vect.size
script_parser = VectorParser()
for i_script in range(N_scripts):
script_parser[i_script] = np.zeros(N_weights)
transform_parser = make_transform([0] * N_layers)
def get_layers(vect):
layers = []
for i_layer in range(N_layers):
weights_by_scripts = vect.reshape((N_scripts, N_weights))
weights_idxs, _ = w_parser.idxs_and_shapes[('weights', i_layer)]
biases_idxs, _ = w_parser.idxs_and_shapes[('biases', i_layer)]
assert weights_idxs.stop == biases_idxs.start
layer_idxs = slice(weights_idxs.start, biases_idxs.stop)
layers.append(weights_by_scripts[:, layer_idxs])
return layers
def transform_weights(z_vect, transform_vect):
z_layers = get_layers(z_vect)
transform = transform_parser.new_vect(transform_vect)
w_layers = [np.dot(transform[i], z) for i, z in enumerate(z_layers)]
return np.concatenate(w_layers, axis=1).ravel()
def likelihood_loss(w_vect, data):
w = script_parser.new_vect(w_vect)
return sum([loss_fun(w[i], **script_data) for i, script_data in enumerate(data)])
def regularization(z_vect):
return np.dot(z_vect, z_vect) * np.exp(log_L2)
def train_z(data, transform_vect, RS):
def primal_loss(z_vect, transform_vect, i_primal, record_results=False):
w_vect = transform_weights(z_vect, transform_vect)
loss = likelihood_loss(w_vect, data)
reg = regularization(z_vect)
if record_results and i_primal % N_thin == 0:
print "Iter {0}: train: {1}".format(i_primal, getval(loss) / N_scripts)
return loss + reg
z_vect_0 = RS.randn(script_parser.vect.size) * np.exp(log_init_scale)
return sgd(grad(primal_loss), transform_vect, z_vect_0, alpha, beta, N_iters)
def train_sharing():
def hyperloss(transform_vect, i_hyper):
RS = RandomState((seed, i_hyper, "hyperloss"))
cur_train_data, cur_valid_data = random_partition(train_data, RS, [10, 2])
z_vect_final = train_z(cur_train_data, transform_vect, RS)
w_vect_final = transform_weights(z_vect_final, transform_vect)
return likelihood_loss(w_vect_final, cur_valid_data) / N_scripts
hypergrad = grad(hyperloss)
cur_transform_vect = make_transform([init_script_corr] * N_layers).vect
for i_hyper in range(N_meta_iter):
print "Hyper iter {0}".format(i_hyper)
grad_transform = hypergrad(cur_transform_vect, i_hyper)
cur_transform_vect = cur_transform_vect - grad_transform * meta_alpha
return cur_transform_vect
transform_vects, train_losses, tests_losses = {}, {}, {}
transform_vects['no_sharing'] = make_transform([0, 0, 0]).vect
transform_vects['full_sharing'] = make_transform([1, 0, 0]).vect
transform_vects['learned_sharing'] = train_sharing()
for name in transform_vects.keys():
RS = RandomState("final_training")
tv = transform_vects[name]
trained_z = train_z(train_data, tv, RS)
trained_w = transform_weights(trained_z, tv)
train_losses[name] = likelihood_loss(trained_w, train_data) / N_scripts
tests_losses[name] = likelihood_loss(trained_w, tests_data) / N_scripts
print "{0} : train: {1}, test: {2}".format(name, train_losses[name], tests_losses[name])
return transform_parser, transform_vects, train_losses, tests_losses
def run():
"""Three different parsers:
w_parser[('biases', i_layer)] : neural net weights/biases per layer for a single script
script_parser[i_script] : weights vector for each script
transform_parser[i_layer] : transform matrix (scripts x scripts) for each alphabet"""
RS = RandomState((seed, "top_rs"))
train_data, valid_data, tests_data = omniglot.load_rotated_alphabets(
[11, 2, 2], RS)
w_parser, pred_fun, loss_fun, frac_err = make_nn_funs(layer_sizes)
N_weights = w_parser.vect.size
transform_parser = make_transform(N_scripts, script_corr_init)
script_parser = VectorParser()
for i_script in range(N_scripts):
script_parser[i_script] = np.zeros(N_weights)
def get_layers(vect):
layers = []
for i_layer in range(N_layers):
weights_by_scripts = vect.reshape((N_scripts, N_weights))
weights_idxs, _ = w_parser.idxs_and_shapes[('weights', i_layer)]
biases_idxs, _ = w_parser.idxs_and_shapes[('biases', i_layer)]
assert weights_idxs.stop == biases_idxs.start
layer_idxs = slice(weights_idxs.start, biases_idxs.stop)
layers.append(weights_by_scripts[:, layer_idxs])
return layers
def transform_weights(z_vect, transform_vect):
z_layers = get_layers(z_vect)
transform = transform_parser.new_vect(transform_vect)
w_layers = [np.dot(transform[i], z) for i, z in enumerate(z_layers)]
return np.concatenate(w_layers, axis=1).ravel()
def total_loss(w_vect, data):
w = script_parser.new_vect(w_vect)
return sum([
loss_fun(w[i], **script_data) for i, script_data in enumerate(data)
])
def regularization(z_vect):
return np.dot(z_vect, z_vect) * np.exp(log_L2_init)
results = defaultdict(list)
def hyperloss(transform_vect, i_hyper, record_results=True):
RS = RandomState((seed, i_hyper, "hyperloss"))
def primal_loss(z_vect,
transform_vect,
i_primal,
record_results=False):
w_vect = transform_weights(z_vect, transform_vect)
loss = total_loss(w_vect, train_data)
reg = regularization(z_vect)
if VERBOSE and record_results and i_primal % N_thin == 0:
print "Iter {0}: train: {1}, valid: {2}, reg: {3}".format(
i_primal,
getval(loss) / N_scripts,
total_loss(getval(w_vect), valid_data) / N_scripts,
getval(reg))
return loss + reg
z_vect_0 = RS.randn(
script_parser.vect.size) * np.exp(log_initialization_scale)
z_vect_final = sgd(grad(primal_loss),
transform_vect,
z_vect_0,
alpha,
beta,
N_iters,
callback=None)
w_vect_final = transform_weights(z_vect_final, transform_vect)
valid_loss = total_loss(w_vect_final, valid_data)
if record_results:
results['valid_loss'].append(getval(valid_loss) / N_scripts)
results['train_loss'].append(
total_loss(getval(w_vect_final), train_data) / N_scripts)
results['tests_loss'].append(
total_loss(getval(w_vect_final), tests_data) / N_scripts)
return valid_loss
all_transforms = []
hypergrad = grad(hyperloss)
for i_hyper in range(N_meta_iter):
grad_transform = hypergrad(transform_parser.vect,
i_hyper,
record_results=True)
transform_parser.vect = transform_parser.vect - grad_transform * meta_alpha
all_transforms.append(transform_parser.as_dict())
print "Hyper iter {0}".format(i_hyper)
print "Results", {k: v[-1] for k, v in results.iteritems()}
return results, all_transforms
def run():
"""Three different parsers:
w_parser[('biases', i_layer)] : neural net weights/biases per layer for a single script
script_parser[i_script] : weights vector for each script
transform_parser[i_layer] : transform matrix (scripts x scripts) for each alphabet"""
RS = RandomState((seed, "top_rs"))
train_data, valid_data, tests_data = omniglot.load_rotated_alphabets([11, 2, 2], RS)
w_parser, pred_fun, loss_fun, frac_err = make_nn_funs(layer_sizes)
N_weights = w_parser.vect.size
transform_parser = make_transform(N_scripts, script_corr_init)
script_parser = VectorParser()
for i_script in range(N_scripts):
script_parser[i_script] = np.zeros(N_weights)
def get_layers(vect):
layers = []
for i_layer in range(N_layers):
weights_by_scripts = vect.reshape((N_scripts, N_weights))
weights_idxs, _ = w_parser.idxs_and_shapes[('weights', i_layer)]
biases_idxs, _ = w_parser.idxs_and_shapes[('biases', i_layer)]
assert weights_idxs.stop == biases_idxs.start
layer_idxs = slice(weights_idxs.start, biases_idxs.stop)
layers.append(weights_by_scripts[:, layer_idxs])
return layers
def transform_weights(z_vect, transform_vect):
z_layers = get_layers(z_vect)
transform = transform_parser.new_vect(transform_vect)
w_layers = [np.dot(transform[i], z) for i, z in enumerate(z_layers)]
return np.concatenate(w_layers, axis=1).ravel()
def total_loss(w_vect, data):
w = script_parser.new_vect(w_vect)
return sum([loss_fun(w[i], **script_data) for i, script_data in enumerate(data)])
def regularization(z_vect):
return np.dot(z_vect, z_vect) * np.exp(log_L2_init)
results = defaultdict(list)
def hyperloss(transform_vect, i_hyper, record_results=True):
RS = RandomState((seed, i_hyper, "hyperloss"))
def primal_loss(z_vect, transform_vect, i_primal, record_results=False):
w_vect = transform_weights(z_vect, transform_vect)
loss = total_loss(w_vect, train_data)
reg = regularization(z_vect)
if VERBOSE and record_results and i_primal % N_thin == 0:
print "Iter {0}: train: {1}, valid: {2}, reg: {3}".format(
i_primal,
getval(loss) / N_scripts,
total_loss(getval(w_vect), valid_data) / N_scripts,
getval(reg))
return loss + reg
z_vect_0 = RS.randn(script_parser.vect.size) * np.exp(log_initialization_scale)
z_vect_final = sgd(grad(primal_loss), transform_vect, z_vect_0,
alpha, beta, N_iters, callback=None)
w_vect_final = transform_weights(z_vect_final, transform_vect)
valid_loss = total_loss(w_vect_final, valid_data)
if record_results:
results['valid_loss'].append(getval(valid_loss) / N_scripts)
results['train_loss'].append(total_loss(getval(w_vect_final), train_data) / N_scripts)
results['tests_loss'].append(total_loss(getval(w_vect_final), tests_data) / N_scripts)
return valid_loss
all_transforms = []
hypergrad = grad(hyperloss)
for i_hyper in range(N_meta_iter):
grad_transform = hypergrad(transform_parser.vect, i_hyper, record_results=True)
transform_parser.vect = transform_parser.vect - grad_transform * meta_alpha
all_transforms.append(transform_parser.as_dict())
print "Hyper iter {0}".format(i_hyper)
print "Results", {k : v[-1] for k, v in results.iteritems()}
return results, all_transforms