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 plot():
import matplotlib.pyplot as plt
import matplotlib as mpl
mpl.rcParams['font.family'] = 'serif'
with open('results.pkl') as f:
transform_parser, transform_vects, train_losses, tests_losses = pickle.load(
f)
omniglot.show_alphabets(omniglot.load_flipped_alphabets(normalize=False))
# 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
}
prop = {'family': 'serif', 'size': '12'}
fig = plt.figure(0)
fig.clf()
fig.set_size_inches((6, 6))
for i, name in enumerate(names):
ax = fig.add_subplot(1, 3, i + 1)
ax.imshow(covar_imgs[name], cmap=mpl.cm.binary)
ax.set_title(title_strings[name])
ax.set_xticks([])
ax.set_yticks([])
if i == 0:
labels = ["Layer {0}".format(layer) for layer in [3, 2, 1]]
ypos = [5, 15, 25]
for s, y in zip(labels, ypos):
ax.text(-2, y, s, rotation='vertical')
plt.tight_layout()
plt.savefig('learned_corr.png')
plt.savefig('learned_corr.pdf')
def plot():
import matplotlib.pyplot as plt
import matplotlib as mpl
mpl.rcParams["font.family"] = "serif"
with open("results.pkl") as f:
transform_parser, transform_vects, train_losses, tests_losses = pickle.load(f)
omniglot.show_alphabets(omniglot.load_flipped_alphabets(normalize=False))
# 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}
prop = {"family": "serif", "size": "12"}
fig = plt.figure(0)
fig.clf()
fig.set_size_inches((6, 6))
for i, name in enumerate(names):
ax = fig.add_subplot(1, 3, i + 1)
ax.imshow(covar_imgs[name], cmap=mpl.cm.binary)
ax.set_title(title_strings[name])
ax.set_xticks([])
ax.set_yticks([])
if i == 0:
labels = ["Layer {0}".format(layer) for layer in [3, 2, 1]]
ypos = [5, 15, 25]
for s, y in zip(labels, ypos):
ax.text(-2, y, s, rotation="vertical")
plt.tight_layout()
plt.savefig("learned_corr.png")
plt.savefig("learned_corr.pdf")
def plot():
import matplotlib.pyplot as plt
import matplotlib as mpl
mpl.rcParams['font.family'] = 'serif'
with open('results.pkl') as f:
transform_parser, transform_vects, train_losses, tests_losses = pickle.load(f)
RS = RandomState((seed, "top_rs"))
omniglot.show_alphabets(omniglot.load_flipped_alphabets(RS, normalize=False))
# Plotting transformations
names = ['no_sharing', 'full_sharing', 'learned_sharing']
title_strings = {'no_sharing' : 'Independent\nnets',
'full_sharing' : 'Shared\nbottom layer',
'learned_sharing' : 'Learned\nsharing'}
covar_imgs = {name : build_covar_image(transform_vects[name]) for name in names}
prop={'family':'serif', 'size':'12'}
fig = plt.figure(0)
fig.clf()
fig.set_size_inches((4,4))
for i, name in enumerate(names):
ax = fig.add_subplot(1, 3, i + 1)
ax.matshow(covar_imgs[name], cmap = mpl.cm.binary)
ax.set_title(title_strings[name])
ax.set_xticks([])
ax.set_yticks([])
if i == 0:
labels = ["Layer {0}".format(layer) for layer in [3, 2, 1]]
ypos = [5, 15, 25]
for s, y in zip(labels, ypos):
ax.text(-3, y, s, rotation='vertical')
plt.tight_layout()
plt.savefig('learned_corr.png')
plt.savefig('learned_corr.pdf')
