Possible sign error in DeepLearning gradient descent code

[flipdazed]

Issue There may be a sign error in the RBM gradient descent code provided on [deeplearning.net]()

Affected

rbm.py
dbn.py
test_rbm.py
test_dbn.py
train_rbm.py
train_dbn.py

Reproduction

Code

The cost is calculated in models.rbm.RBM.getCostUpdates as,

    cost = T.mean(self.freeEnergy(self.inputs)) - T.mean(
        self.freeEnergy(chain_end))

with the gradient given by

    # We must not compute the gradient through the gibbs sampling
    gparams = T.grad(cost, self.params, consider_constant=[chain_end])

but in the theory this is the negative gradient. However, this gradient is still subtracted from params in the update loop,

    for gparam, param in zip(gparams, self.params):
        updates[param] = param - gparam * T.cast(lr, dtype=theano.config.floatX)

[flipdazed]

see equation (1) in deeplearning.net where $l(\theta, \mathcal{D}) = - \langle \ln p(x) \rangle $