init

models/EntropyMinimizationPrinciple.py

@@ -0,0 +1,46 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torch.autograd import Variable
+
+
+def _assert_no_grad(variable):
+    assert not variable.requires_grad, \
+        "nn criterions don't compute the gradient w.r.t. targets - please " \
+        "mark these variables as volatile or not requiring gradients"
+
+
+class _Loss(nn.Module):
+    def __init__(self, size_average=True):
+        super(_Loss, self).__init__()
+        self.size_average = size_average
+
+
+class _WeightedLoss(_Loss):
+    def __init__(self, weight=None, size_average=True):
+        super(_WeightedLoss, self).__init__(size_average)
+        self.register_buffer('weight', weight)
+
+
+class EMLossForTarget(_WeightedLoss):
+
+    def __init__(self, weight=None, size_average=True, ignore_index=-100, reduce=True, nClass=10):
+        super(EMLossForTarget, self).__init__(weight, size_average)
+        self.nClass = nClass
+
+    def forward(self, input):
+        batch_size = input.size(0)
+        prob = F.softmax(input, dim=1)
+        # prob = F.sigmoid(input)
+        prob_source = prob[:, :self.nClass]
+        prob_target = prob[:, self.nClass:]
+        prob_sum = prob_target + prob_source
+        if (prob_sum.data.cpu() == 0).sum() != 0:
+            weight_sum = torch.FloatTensor(batch_size, self.nClass).fill_(0)
+            weight_sum[prob_sum.data.cpu() == 0] = 1e-6
+            weight_sum = Variable(weight_sum).cuda()
+            loss_sum = -(prob_sum + weight_sum).log().mul(prob_sum).sum(1).mean()
+        else:
+            loss_sum = -prob_sum.log().mul(prob_sum).sum(1).mean()
+
+        return loss_sum