You can try OpFromGraph with inline=True, and specify
override_gradients=... with the right expression.
This is still experimental.
On Wed, Mar 29, 2017, nokunok...@gmail.com wrote:
> Hi guys!
>
> I recently started using theano and am struggling to implement custom
> gradient for stochastic node. Can anyone help me?
>
> What I want is an op that produces one hot vector whose hot element is
> sampled from softmax distribution.
> The op is not differentiable, but I want to "fake" as if its gradient is
> softmax's one ("straight through estimator").
> Below is the minimum code that perform forward path, which raises
> DisconnectedInputError due to missing gradient.
>
>
> import theano
>
> import theano.tensor as T
>
> import numpy as np
>
>
> logits_values = np.random.uniform(-1, 1, size=3)
>
> logits = theano.shared(logits_values, 'logits')
>
> probabilities = T.nnet.softmax(logits)
>
> print('probabilities', probabilities.eval())
>
> # result: probabilities [[ 0.55155489 0.290773 0.15767211]]
>
>
> random_streams = T.shared_randomstreams.RandomStreams()
>
> index = random_streams.choice(size=(1,), a=3, p=probabilities[0])
>
> samples = T.extra_ops.to_one_hot(index, logits.shape[-1])
>
> print('samples', samples.eval())
>
> # result: samples [[ 1. 0. 0.]]
>
>
> # We want to use gradient of probabilities instead of samples!
>
> samples_grad = T.grad(samples[0][0], logits)
>
> # result: raise DisconnectedInputError
>
>
> The node is not the final layer, so I can't use categorical cross entropy
> loss for training it.
>
> I am trying to implement custom op (see attached stochastic_softmax.py), but
> it is not working in practice.
>
> Since I have working expression for forward path, can I simply override
> gradient of existing expression?
>
> --
>
> ---
> You received this message because you are subscribed to the Google Groups
> "theano-users" group.
> To unsubscribe from this group and stop receiving emails from it, send an
> email to theano-users+unsubscr...@googlegroups.com.
> For more options, visit https://groups.google.com/d/optout.
> import numpy as np
> import theano
> import theano.tensor as T
>
>
> class StochasticSoftmax(theano.Op):
> def __init__(self, random_state=np.random.RandomState()):
> self.random_state = random_state
>
> def make_node(self, x):
> x = T.as_tensor_variable(x)
> return theano.Apply(self, [x], [x.type()])
>
> def perform(self, node, inputs, output_storage):
> # Gumbel-max trick
> x, = inputs
> z = self.random_state.gumbel(loc=0, scale=1, size=x.shape)
> indices = (x + z).argmax(axis=-1)
> y = np.eye(x.shape[-1], dtype=np.float32)[indices]
> output_storage[0][0] = y
>
> def grad(self, inp, grads):
> x, = inp
> g_sm, = grads
>
> sm = T.nnet.softmax(x)
> return [T.nnet.softmax_grad(g_sm, sm)]
>
> def infer_shape(self, node, i0_shapes):
> return i0_shapes
--
Pascal
--
---
You received this message because you are subscribed to the Google Groups
"theano-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email
to theano-users+unsubscr...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.
