268710944c3b60049188ed2764aebc08de9bd417,pymc3/variational/ADVI.py,,inner_gradients,#Any#Any#Any#,78

Before Change

    duw = gradient(logp(q), [uw])

    return theano.tensor.set_subtensor(duw[l:], duw[l:] + 1)

tprint = theano.printing.Print()

def adagrad(grad, param, learning_rate, epsilon, n):

After Change

    duw = gradient(logp(q), [uw])

    // Add gradient of entropy term (just equal to element-wise 1 here), formula 6
    duw = theano.tensor.set_subtensor(duw[l:], duw[l:] + 1)
    return duw

def adagrad(grad, param, learning_rate, epsilon, n):
    // Compute windowed adagrad using last n gradients
    i = theano.shared(np.array(0), "i")

Instances