#!/usr/bin/env python
+#########################################################################
+# This program is free software: you can redistribute it and/or modify #
+# it under the terms of the version 3 of the GNU General Public License #
+# as published by the Free Software Foundation. #
+# #
+# This program is distributed in the hope that it will be useful, but #
+# WITHOUT ANY WARRANTY; without even the implied warranty of #
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU #
+# General Public License for more details. #
+# #
+# You should have received a copy of the GNU General Public License #
+# along with this program. If not, see <http://www.gnu.org/licenses/>. #
+# #
+# Written by and Copyright (C) Francois Fleuret #
+# Contact <francois.fleuret@idiap.ch> for comments & bug reports #
+#########################################################################
+
import argparse, math, sys
from copy import deepcopy
######################################################################
parser = argparse.ArgumentParser(
- description = 'An implementation of Mutual Information estimator with a deep model',
+ description = '''An implementation of a Mutual Information estimator with a deep model
+
+Three different toy data-sets are implemented:
+
+ (1) Two MNIST images of same class. The "true" MI is the log of the
+ number of used MNIST classes.
+
+ (2) One MNIST image and a pair of real numbers whose difference is
+ the class of the image. The "true" MI is the log of the number of
+ used MNIST classes.
+
+ (3) Two 1d sequences, the first with a single peak, the second with
+ two peaks, and the height of the peak in the first is the
+ difference of timing of the peaks in the second. The "true" MI is
+ the log of the number of possible peak heights.''',
+
formatter_class = argparse.ArgumentDefaultsHelpFormatter
)
parser.add_argument('--data',
type = str, default = 'image_pair',
- help = 'What data')
+ help = 'What data: image_pair, image_values_pair, sequence_pair')
parser.add_argument('--seed',
type = int, default = 0,
type = int, default = 100,
help = 'Batch size')
-######################################################################
-
-def entropy(target):
- probas = []
- for k in range(target.max() + 1):
- n = (target == k).sum().item()
- if n > 0: probas.append(n)
- probas = torch.tensor(probas).float()
- probas /= probas.sum()
- return - (probas * probas.log()).sum().item()
+parser.add_argument('--learning_rate',
+ type = float, default = 1e-3,
+ help = 'Batch size')
-def robust_log_mean_exp(x):
- # a = x.max()
- # return (x-a).exp().mean().log() + a
- # a = x.max()
- return x.exp().mean().log()
+parser.add_argument('--independent', action = 'store_true',
+ help = 'Should the pair components be independent')
######################################################################
######################################################################
+def entropy(target):
+ probas = []
+ for k in range(target.max() + 1):
+ n = (target == k).sum().item()
+ if n > 0: probas.append(n)
+ probas = torch.tensor(probas).float()
+ probas /= probas.sum()
+ return - (probas * probas.log()).sum().item()
+
+######################################################################
+
train_set = torchvision.datasets.MNIST('./data/mnist/', train = True, download = True)
train_input = train_set.train_data.view(-1, 1, 28, 28).to(device).float()
train_target = train_set.train_labels.to(device)
c = torch.cat(uc, 0)
perm = torch.randperm(a.size(0))
a = a[perm].contiguous()
+
+ if args.independent:
+ perm = torch.randperm(a.size(0))
b = b[perm].contiguous()
return a, b, c
######################################################################
# Returns a triplet a, b, c where a are the standard MNIST images, c
-# the classes, and b is a Nx2 tensor, eith for every n:
+# the classes, and b is a Nx2 tensor, with for every n:
#
# b[n, 0] ~ Uniform(0, 10)
# b[n, 1] ~ b[n, 0] + Uniform(0, 0.5) + c[n]
b = a.new(a.size(0), 2)
b[:, 0].uniform_(0.0, 10.0)
b[:, 1].uniform_(0.0, 0.5)
- b[:, 1] += b[:, 0] + target.float()
+
+ if args.independent:
+ b[:, 1] += b[:, 0] + \
+ used_MNIST_classes[torch.randint(len(used_MNIST_classes), target.size())]
+ else:
+ b[:, 1] += b[:, 0] + target.float()
return a, b, c
noise_level = 2e-2
ha = torch.randint(args.nb_classes, (nb, ), device = device) + 1
- # hb = torch.randint(args.nb_classes, (nb, ), device = device)
- hb = ha
+ if args.independent:
+ hb = torch.randint(args.nb_classes, (nb, ), device = device)
+ else:
+ hb = ha
pos = torch.empty(nb, device = device).uniform_(0.0, 0.9)
a = torch.linspace(0, 1, length, device = device).view(1, -1).expand(nb, -1)
b1 = b1 - pos.view(nb, 1)
b1 = (b1 >= 0).float() * torch.exp(-b1 * math.log(2) / 0.1) * 0.25
pos = pos + hb.float() / (args.nb_classes + 1) * 0.5
+ # pos += pos.new(hb.size()).uniform_(0.0, 0.01)
b2 = torch.linspace(0, 1, length, device = device).view(1, -1).expand(nb, -1)
b2 = b2 - pos.view(nb, 1)
b2 = (b2 >= 0).float() * torch.exp(-b2 * math.log(2) / 0.1) * 0.25
if args.data == 'image_pair':
create_pairs = create_image_pairs
model = NetForImagePair()
+
elif args.data == 'image_values_pair':
create_pairs = create_image_values_pairs
model = NetForImageValuesPair()
+
elif args.data == 'sequence_pair':
create_pairs = create_sequences_pairs
model = NetForSequencePair()
- ######################################################################
+
+ ######################
+ ## Save for figures
a, b, c = create_pairs()
for k in range(10):
- file = open(f'/tmp/train_{k:02d}.dat', 'w')
+ file = open(f'train_{k:02d}.dat', 'w')
for i in range(a.size(1)):
file.write(f'{a[k, i]:f} {b[k,i]:f}\n')
file.close()
- # exit(0)
- ######################################################################
+ ######################
+
else:
raise Exception('Unknown data ' + args.data)
######################################################################
+# Train
-print('nb_parameters %d' % sum(x.numel() for x in model.parameters()))
+print(f'nb_parameters {sum(x.numel() for x in model.parameters())}')
model.to(device)
+input_a, input_b, classes = create_pairs(train = True)
+
for e in range(args.nb_epochs):
- input_a, input_b, classes = create_pairs(train = True)
+ optimizer = torch.optim.Adam(model.parameters(), lr = args.learning_rate)
input_br = input_b[torch.randperm(input_b.size(0))]
acc_mi = 0.0
- optimizer = torch.optim.Adam(model.parameters(), lr = 1e-4)
-
for batch_a, batch_b, batch_br in zip(input_a.split(args.batch_size),
input_b.split(args.batch_size),
input_br.split(args.batch_size)):
acc_mi /= (input_a.size(0) // args.batch_size)
- print('%d %.04f %.04f' % (e + 1, acc_mi / math.log(2), entropy(classes) / math.log(2)))
+ print(f'{e+1} {acc_mi / math.log(2):.04f} {entropy(classes) / math.log(2):.04f}')
sys.stdout.flush()
######################################################################
+# Test
input_a, input_b, classes = create_pairs(train = False)
acc_mi /= (input_a.size(0) // args.batch_size)
-print('test %.04f %.04f'%(acc_mi / math.log(2), entropy(classes) / math.log(2)))
+print(f'test {acc_mi / math.log(2):.04f} {entropy(classes) / math.log(2):.04f}')
######################################################################