#!/usr/bin/env python
-# @XREMOTE_HOST: elk.fleuret.org
-# @XREMOTE_EXEC: ~/conda/bin/python
-# @XREMOTE_PRE: ln -s ~/data/pytorch ./data
-# @XREMOTE_PRE: killall -q -9 python || true
+import argparse
import math, sys, torch, torchvision
######################################################################
-# Returns a pair of tensors (a, b, c), where a and b are Nx1x28x28
-# tensors containing images, with a[i] and b[i] of same class for any
-# i, and c is a 1d long tensor with the count of pairs per class used.
+parser = argparse.ArgumentParser(
+ description = 'An implementation of Mutual Information estimator with a deep model',
+ formatter_class = argparse.ArgumentDefaultsHelpFormatter
+)
-def create_pair_set(used_classes, input, target):
- u = []
+parser.add_argument('--data',
+ type = str, default = 'image_pair',
+ help = 'What data')
- for i in used_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))]
- # Careful with odd numbers of samples in a class
- x = x[0:2 * (x.size(0) // 2)].reshape(-1, 2, 28, 28)
- u.append(x)
+parser.add_argument('--seed',
+ type = int, default = 0,
+ help = 'Random seed (default 0, < 0 is no seeding)')
- x = torch.cat(u, 0)
- x = x[torch.randperm(x.size(0))]
- c = torch.tensor([x.size(0) for x in u])
-
- return x.narrow(1, 0, 1).contiguous(), x.narrow(1, 1, 1).contiguous(), c
+parser.add_argument('--mnist_classes',
+ type = str, default = '0, 1, 3, 5, 6, 7, 8, 9',
+ help = 'What MNIST classes to use')
######################################################################
-class Net(nn.Module):
- def __init__(self):
- super(Net, self).__init__()
- self.conv1 = nn.Conv2d(2, 32, kernel_size = 5)
- self.conv2 = nn.Conv2d(32, 64, kernel_size = 5)
- self.fc1 = nn.Linear(256, 200)
- self.fc2 = nn.Linear(200, 1)
+args = parser.parse_args()
- def forward(self, a, b):
- x = torch.cat((a, b), 1)
- x = F.relu(F.max_pool2d(self.conv1(x), kernel_size = 3))
- x = F.relu(F.max_pool2d(self.conv2(x), kernel_size = 2))
- x = x.view(x.size(0), -1)
- x = F.relu(self.fc1(x))
- x = self.fc2(x)
- return x
+if args.seed >= 0:
+ torch.manual_seed(args.seed)
+
+used_MNIST_classes = torch.tensor(eval('[' + args.mnist_classes + ']'))
######################################################################
test_input = test_set.test_data.view(-1, 1, 28, 28).float()
test_target = test_set.test_labels
+if torch.cuda.is_available():
+ used_MNIST_classes = used_MNIST_classes.cuda()
+ train_input, train_target = train_input.cuda(), train_target.cuda()
+ test_input, test_target = test_input.cuda(), test_target.cuda()
+
mu, std = train_input.mean(), train_input.std()
train_input.sub_(mu).div_(std)
test_input.sub_(mu).div_(std)
######################################################################
-# The information bound is the log of the number of classes in there
+# 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 with the count of pairs per class used.
+
+def create_image_pairs(train = False):
+ ua, ub = [], []
+
+ 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))
+
+ a = torch.cat(ua, 0)
+ b = torch.cat(ub, 0)
+ perm = torch.randperm(a.size(0))
+ a = a[perm].contiguous()
+ b = b[perm].contiguous()
+ c = torch.tensor([x.size(0) for x in ua])
+
+ return a, b, c
+
+######################################################################
+
+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)
-# used_classes = torch.tensor([ 0, 1, 3, 5, 6, 7, 8, 9])
-used_classes = torch.tensor([ 3, 4, 7, 0 ])
+ input = input[used_indices].contiguous()
+ target = target[used_indices].contiguous()
+
+ a = input
+
+ b = a.new(a.size(0), 2)
+ b[:, 0].uniform_(10)
+ b[:, 1].uniform_(0.5)
+ b[:, 1] += b[:, 0] + target.float()
+
+ c = torch.tensor([(target == k).sum().item() for k in used_MNIST_classes])
+
+ return a, b, c
+
+######################################################################
+
+class NetImagePair(nn.Module):
+ def __init__(self):
+ super(NetImagePair, 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 NetImageValuesPair(nn.Module):
+ def __init__(self):
+ super(NetImageValuesPair, 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)
+
+######################################################################
+
+if args.data == 'image_pair':
+ create_pairs = create_image_pairs
+ model = NetImagePair()
+elif args.data == 'image_values_pair':
+ create_pairs = create_image_values_pairs
+ model = NetImageValuesPair()
+else:
+ raise Exception('Unknown data ' + args.data)
+
+######################################################################
nb_epochs, batch_size = 50, 100
-model = Net()
+print('nb_parameters %d' % sum(x.numel() for x in model.parameters()))
+
optimizer = torch.optim.Adam(model.parameters(), lr = 1e-3)
if torch.cuda.is_available():
model.cuda()
- train_input, train_target = train_input.cuda(), train_target.cuda()
- test_input, test_target = test_input.cuda(), test_target.cuda()
for e in range(nb_epochs):
- input_a, input_b, count = create_pair_set(used_classes, train_input, train_target)
+ input_a, input_b, count = create_pairs(train = True)
+ # The information bound is the entropy of the class distribution
class_proba = count.float()
class_proba /= class_proba.sum()
class_entropy = - (class_proba.log() * class_proba).sum().item()
input_br = input_b[torch.randperm(input_b.size(0))]
- mi = 0.0
+ acc_mi = 0.0
for batch_a, batch_b, batch_br in zip(input_a.split(batch_size),
input_b.split(batch_size),
input_br.split(batch_size)):
- loss = - (model(batch_a, batch_b).mean() - model(batch_a, batch_br).exp().mean().log())
- mi -= loss.item()
+ mi = model(batch_a, batch_b).mean() - model(batch_a, batch_br).exp().mean().log()
+ loss = - mi
+ acc_mi += mi.item()
optimizer.zero_grad()
loss.backward()
optimizer.step()
- mi /= (input_a.size(0) // batch_size)
+ acc_mi /= (input_a.size(0) // batch_size)
- print('%d %.04f %.04f'%(e, mi / math.log(2), class_entropy / math.log(2)))
+ print('%d %.04f %.04f' % (e, acc_mi / math.log(2), class_entropy / math.log(2)))
sys.stdout.flush()
######################################################################
-input_a, input_b, count = create_pair_set(used_classes, test_input, test_target)
+input_a, input_b, count = create_pairs(train = False)
for e in range(nb_epochs):
class_proba = count.float()
input_br = input_b[torch.randperm(input_b.size(0))]
- mi = 0.0
+ acc_mi = 0.0
for batch_a, batch_b, batch_br in zip(input_a.split(batch_size),
input_b.split(batch_size),
input_br.split(batch_size)):
- loss = - (model(batch_a, batch_b).mean() - model(batch_a, batch_br).exp().mean().log())
- mi -= loss.item()
+ mi = model(batch_a, batch_b).mean() - model(batch_a, batch_br).exp().mean().log()
+ acc_mi += mi.item()
- mi /= (input_a.size(0) // batch_size)
+ acc_mi /= (input_a.size(0) // batch_size)
-print('test %.04f %.04f'%(mi / math.log(2), class_entropy / math.log(2)))
+print('test %.04f %.04f'%(acc_mi / math.log(2), class_entropy / math.log(2)))
######################################################################