--- /dev/null
+#!/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
+
+import torch, torchvision
+
+from torch import nn
+import torch.nn.functional as F
+
+######################################################################
+
+if torch.cuda.is_available():
+ torch.backends.cudnn.benchmark = True
+ device = torch.device('cuda')
+else:
+ device = torch.device('cpu')
+
+######################################################################
+
+parser = argparse.ArgumentParser(
+ 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: image_pair, image_values_pair, sequence_pair')
+
+parser.add_argument('--seed',
+ type = int, default = 0,
+ help = 'Random seed (default 0, < 0 is no seeding)')
+
+parser.add_argument('--mnist_classes',
+ type = str, default = '0, 1, 3, 5, 6, 7, 8, 9',
+ help = 'What MNIST classes to use')
+
+parser.add_argument('--nb_classes',
+ type = int, default = 2,
+ help = 'How many classes for sequences')
+
+parser.add_argument('--nb_epochs',
+ type = int, default = 50,
+ help = 'How many epochs')
+
+parser.add_argument('--batch_size',
+ type = int, default = 100,
+ help = 'Batch size')
+
+parser.add_argument('--learning_rate',
+ type = float, default = 1e-3,
+ help = 'Batch size')
+
+parser.add_argument('--independent', action = 'store_true',
+ help = 'Should the pair components be independent')
+
+######################################################################
+
+args = parser.parse_args()
+
+if args.seed >= 0:
+ torch.manual_seed(args.seed)
+
+used_MNIST_classes = torch.tensor(eval('[' + args.mnist_classes + ']'), device = device)
+
+######################################################################
+
+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)
+
+test_set = torchvision.datasets.MNIST('./data/mnist/', train = False, download = True)
+test_input = test_set.test_data.view(-1, 1, 28, 28).to(device).float()
+test_target = test_set.test_labels.to(device)
+
+mu, std = train_input.mean(), train_input.std()
+train_input.sub_(mu).div_(std)
+test_input.sub_(mu).div_(std)
+
+######################################################################
+
+# Returns a triplet of tensors (a, b, c), where a and b contain each
+# half of the samples, with a[i] and b[i] of same class for any i, and
+# c is a 1d long tensor real classes
+
+def create_image_pairs(train = False):
+ ua, ub, uc = [], [], []
+
+ if train:
+ input, target = train_input, train_target
+ else:
+ input, target = test_input, test_target
+
+ for i in used_MNIST_classes:
+ used_indices = torch.arange(input.size(0), device = target.device)\
+ .masked_select(target == i.item())
+ x = input[used_indices]
+ x = x[torch.randperm(x.size(0))]
+ hs = x.size(0)//2
+ ua.append(x.narrow(0, 0, hs))
+ ub.append(x.narrow(0, hs, hs))
+ uc.append(target[used_indices])
+
+ a = torch.cat(ua, 0)
+ b = torch.cat(ub, 0)
+ 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, with for every n:
+#
+# b[n, 0] ~ Uniform(0, 10)
+# b[n, 1] ~ b[n, 0] + Uniform(0, 0.5) + c[n]
+
+def create_image_values_pairs(train = False):
+ ua, ub = [], []
+
+ if train:
+ input, target = train_input, train_target
+ else:
+ input, target = test_input, test_target
+
+ m = torch.zeros(used_MNIST_classes.max() + 1, dtype = torch.uint8, device = target.device)
+ m[used_MNIST_classes] = 1
+ m = m[target]
+ used_indices = torch.arange(input.size(0), device = target.device).masked_select(m)
+
+ input = input[used_indices].contiguous()
+ target = target[used_indices].contiguous()
+
+ a = input
+ c = target
+
+ b = a.new(a.size(0), 2)
+ b[:, 0].uniform_(0.0, 10.0)
+ b[:, 1].uniform_(0.0, 0.5)
+
+ 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
+
+######################################################################
+
+def create_sequences_pairs(train = False):
+ nb, length = 10000, 1024
+ noise_level = 2e-2
+
+ ha = torch.randint(args.nb_classes, (nb, ), device = device) + 1
+ 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)
+ a = a - pos.view(nb, 1)
+ a = (a >= 0).float() * torch.exp(-a * math.log(2) / 0.1)
+ a = a * ha.float().view(-1, 1).expand_as(a) / (1 + args.nb_classes)
+ noise = a.new(a.size()).normal_(0, noise_level)
+ a = a + noise
+
+ pos = torch.empty(nb, device = device).uniform_(0.0, 0.5)
+ b1 = 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
+
+ b = b1 + b2
+ noise = b.new(b.size()).normal_(0, noise_level)
+ b = b + noise
+
+ # a = (a - a.mean()) / a.std()
+ # b = (b - b.mean()) / b.std()
+
+ return a, b, ha
+
+######################################################################
+
+class NetForImagePair(nn.Module):
+ def __init__(self):
+ super(NetForImagePair, self).__init__()
+ self.features_a = nn.Sequential(
+ nn.Conv2d(1, 16, kernel_size = 5),
+ nn.MaxPool2d(3), nn.ReLU(),
+ nn.Conv2d(16, 32, kernel_size = 5),
+ nn.MaxPool2d(2), nn.ReLU(),
+ )
+
+ self.features_b = nn.Sequential(
+ nn.Conv2d(1, 16, kernel_size = 5),
+ nn.MaxPool2d(3), nn.ReLU(),
+ nn.Conv2d(16, 32, kernel_size = 5),
+ nn.MaxPool2d(2), nn.ReLU(),
+ )
+
+ self.fully_connected = nn.Sequential(
+ nn.Linear(256, 200),
+ nn.ReLU(),
+ nn.Linear(200, 1)
+ )
+
+ def forward(self, a, b):
+ a = self.features_a(a).view(a.size(0), -1)
+ b = self.features_b(b).view(b.size(0), -1)
+ x = torch.cat((a, b), 1)
+ return self.fully_connected(x)
+
+######################################################################
+
+class NetForImageValuesPair(nn.Module):
+ def __init__(self):
+ super(NetForImageValuesPair, self).__init__()
+ self.features_a = nn.Sequential(
+ nn.Conv2d(1, 16, kernel_size = 5),
+ nn.MaxPool2d(3), nn.ReLU(),
+ nn.Conv2d(16, 32, kernel_size = 5),
+ nn.MaxPool2d(2), nn.ReLU(),
+ )
+
+ self.features_b = nn.Sequential(
+ nn.Linear(2, 32), nn.ReLU(),
+ nn.Linear(32, 32), nn.ReLU(),
+ nn.Linear(32, 128), nn.ReLU(),
+ )
+
+ self.fully_connected = nn.Sequential(
+ nn.Linear(256, 200),
+ nn.ReLU(),
+ nn.Linear(200, 1)
+ )
+
+ def forward(self, a, b):
+ a = self.features_a(a).view(a.size(0), -1)
+ b = self.features_b(b).view(b.size(0), -1)
+ x = torch.cat((a, b), 1)
+ return self.fully_connected(x)
+
+######################################################################
+
+class NetForSequencePair(nn.Module):
+
+ def feature_model(self):
+ kernel_size = 11
+ pooling_size = 4
+ return nn.Sequential(
+ nn.Conv1d( 1, self.nc, kernel_size = kernel_size),
+ nn.AvgPool1d(pooling_size),
+ nn.LeakyReLU(),
+ nn.Conv1d(self.nc, self.nc, kernel_size = kernel_size),
+ nn.AvgPool1d(pooling_size),
+ nn.LeakyReLU(),
+ nn.Conv1d(self.nc, self.nc, kernel_size = kernel_size),
+ nn.AvgPool1d(pooling_size),
+ nn.LeakyReLU(),
+ nn.Conv1d(self.nc, self.nc, kernel_size = kernel_size),
+ nn.AvgPool1d(pooling_size),
+ nn.LeakyReLU(),
+ )
+
+ def __init__(self):
+ super(NetForSequencePair, self).__init__()
+
+ self.nc = 32
+ self.nh = 256
+
+ self.features_a = self.feature_model()
+ self.features_b = self.feature_model()
+
+ self.fully_connected = nn.Sequential(
+ nn.Linear(2 * self.nc, self.nh),
+ nn.ReLU(),
+ nn.Linear(self.nh, 1)
+ )
+
+ def forward(self, a, b):
+ a = a.view(a.size(0), 1, a.size(1))
+ a = self.features_a(a)
+ a = F.avg_pool1d(a, a.size(2))
+
+ b = b.view(b.size(0), 1, b.size(1))
+ b = self.features_b(b)
+ b = F.avg_pool1d(b, b.size(2))
+
+ x = torch.cat((a.view(a.size(0), -1), b.view(b.size(0), -1)), 1)
+ return self.fully_connected(x)
+
+######################################################################
+
+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'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()
+ ######################
+
+else:
+ raise Exception('Unknown data ' + args.data)
+
+######################################################################
+# Train
+
+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):
+
+ optimizer = torch.optim.Adam(model.parameters(), lr = args.learning_rate)
+
+ input_br = input_b[torch.randperm(input_b.size(0))]
+
+ acc_mi = 0.0
+
+ 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)):
+ mi = model(batch_a, batch_b).mean() - model(batch_a, batch_br).exp().mean().log()
+ acc_mi += mi.item()
+ loss = - mi
+ optimizer.zero_grad()
+ loss.backward()
+ optimizer.step()
+
+ acc_mi /= (input_a.size(0) // args.batch_size)
+
+ 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)
+
+input_br = input_b[torch.randperm(input_b.size(0))]
+
+acc_mi = 0.0
+
+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)):
+ mi = model(batch_a, batch_b).mean() - model(batch_a, batch_br).exp().mean().log()
+ acc_mi += mi.item()
+
+acc_mi /= (input_a.size(0) // args.batch_size)
+
+print(f'test {acc_mi / math.log(2):.04f} {entropy(classes) / math.log(2):.04f}')
+
+######################################################################