Mercurial > pylearn

comparison sparse_random_autoassociator/graph.py @ 381:e4473d9697d7

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
Added xent loss
author Joseph Turian <turian@gmail.com>
date Tue, 08 Jul 2008 01:57:33 -0400
parents a1bbcde6b456
children b4efd192d880
comparison
equal deleted inserted replaced
373:42cc94cf6c12 381:e4473d9697d7
29 # xnonzero sensitive loss: 29 # xnonzero sensitive loss:
30 #nonzeroloss = hingeloss(ynonzero - t.max(yzero) - MARGIN - xnonzero) 30 #nonzeroloss = hingeloss(ynonzero - t.max(yzero) - MARGIN - xnonzero)
31 #zeroloss = hingeloss(-t.max(-(ynonzero - xnonzero)) - yzero - MARGIN) 31 #zeroloss = hingeloss(-t.max(-(ynonzero - xnonzero)) - yzero - MARGIN)
32 loss = t.sum(nonzeroloss) + t.sum(zeroloss) 32 loss = t.sum(nonzeroloss) + t.sum(zeroloss)
33 33
34 def binary_crossentropy(output, target):
35 """
36 Compute the crossentropy of binary output wrt binary target.
37 @note: We do not sum, crossentropy is computed by component.
38 @todo: Rewrite as a scalar, and then broadcast to tensor.
39 """
40 return -(target * t.log(output) + (1 - target) * t.log(1 - output))
41 #loss = t.sum(binary_crossentropy(ynonzero, xnonzero)) + t.sum(binary_crossentropy(yzero, t.constant(0)))
42
34 (gw1nonzero, gb1, gw2nonzero, gw2zero, gb2nonzero, gb2zero) = t.grad(loss, [w1nonzero, b1, w2nonzero, w2zero, b2nonzero, b2zero]) 43 (gw1nonzero, gb1, gw2nonzero, gw2zero, gb2nonzero, gb2zero) = t.grad(loss, [w1nonzero, b1, w2nonzero, w2zero, b2nonzero, b2zero])
35 44
36 import theano.compile 45 import theano.compile
37 46
38 inputs = [xnonzero, w1nonzero, b1, w2nonzero, w2zero, b2nonzero, b2zero] 47 inputs = [xnonzero, w1nonzero, b1, w2nonzero, w2zero, b2nonzero, b2zero]