From: Francois Fleuret Date: Wed, 14 Nov 2018 08:37:14 +0000 (+0100) Subject: Initial commit. X-Git-Url: https://fleuret.org/cgi-bin/gitweb/gitweb.cgi?a=commitdiff_plain;h=5e18633709b82b0739e62d105b2685793be1d014;p=pytorch.git Initial commit. --- diff --git a/mine_mnist.py b/mine_mnist.py new file mode 100755 index 0000000..f573b89 --- /dev/null +++ b/mine_mnist.py @@ -0,0 +1,105 @@ +#!/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 || echo "Nothing killed" + +import math, sys, torch, torchvision + +from torch import nn +from torch.nn import functional as F + +###################################################################### + +# Returns a pair of tensors (x, c), where x is a Nx2x28x28 containing +# pairs of images of same classes (one per channel), and p is a 1d +# long tensor with the count of pairs per class used + +def create_pair_set(used_classes, input, target): + u = [] + + 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) + + x = torch.cat(u, 0).contiguous() + c = torch.tensor([x.size(0) for x in u]) + + return x, c + +###################################################################### + +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) + + def forward(self, x): + 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 + +###################################################################### + +train_set = torchvision.datasets.MNIST('./data/mnist/', train = True, download = True) +train_input = train_set.train_data.view(-1, 1, 28, 28).float() +train_target = train_set.train_labels + +mu, std = train_input.mean(), train_input.std() +train_input.sub_(mu).div_(std) + +###################################################################### + +# The information bound is the log of the number of classes in there + +# used_classes = torch.tensor([ 0, 1, 3, 5, 6, 7, 8, 9]) +used_classes = torch.tensor([ 3, 4, 7, 0 ]) + +nb_epochs, batch_size = 50, 100 + +model = Net() +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() + +for e in range(nb_epochs): + input, count = create_pair_set(used_classes, train_input, train_target) + + class_proba = count.float() + class_proba /= class_proba.sum() + class_entropy = - (class_proba.log() * class_proba).sum().item() + + input = input[torch.randperm(input.size(0))] + indep_input = input.clone() + indep_input[:, 1] = input[torch.randperm(input.size(0)), 1] + + mi = 0.0 + + for batch, indep_batch in zip(input.split(batch_size), indep_input.split(batch_size)): + loss = - (model(batch).mean() - model(indep_batch).exp().mean().log()) + mi -= loss.item() + optimizer.zero_grad() + loss.backward() + optimizer.step() + + mi /= (input.size(0) // batch_size) + + print('%d %.04f %.04f'%(e, mi / math.log(2), class_entropy / math.log(2))) + + sys.stdout.flush() + +######################################################################