Let me be more specific. As to multiplying, I know that the result of
gradient is something like [array(), array(), ...], the number of elements
in this list is the number of derivative parameters.
If T.grad() * 2, it would double the number of elements in the list, but
not the value of elements.
So what need to do if I want to multiply something with every element in
the gradient list?
Same as the problem of summing
On Tuesday, July 19, 2016 at 3:50:45 PM UTC+8, 祝士祥 wrote:
>
> Hi everybody,
>
> I am trying to calculate the product between a result of gradient and a
> formula, like this:
> T.grad(T.log(hl_mean).sum(), self.params, disconnected_inputs='warn') * T.
> log(vn_mean).sum()
> The error is:
> theano.tensor.var.AsTensorError: ('Cannot convert [dot.0,
> Elemwise{mul}.0, Elemwise{second,no_inplace}.0] to TensorType', <type
> 'list'>)
>
> And to sum multiple results of gradient, like this:
> (Rls, updates) = theano.scan(
> calculate_Rl,
> outputs_info=None,
> non_sequences=v_input,
> n_steps=hidden_sample_l
> )
> Rn = Rls.sum()
> calculate_Rl returns a result of gradient:
> def calculate_Rl(v_input):
> # Sample a h_sample according to one v_input
> _, hl_mean, hl_sample = self.sample_h_given_v(v_input)
> # Calculate the probability of visible output according to h_sample
> _, vn_mean = self.propdown(hl_sample)
> # return T.grad(T.log(hl_mean).sum(), self.params,
> disconnected_inputs='warn') \
> # * T.log(vn_mean).sum() \
> # + T.grad((T.log(vn_mean).sum()), self.params,
> consider_constant=[vn_mean], disconnected_inputs='warn')
> return T.grad(T.log(hl_mean).sum(), self.params,
> disconnected_inputs='warn') * T.log(vn_mean).sum()
> The error is:
> AttributeError: 'list' object has no attribute 'sum'
>
> So, the question is how to operate the results of gradients in Theano?
> Such as calculating product or sum?
>
>
>
> Any help is appreciated! Thanks!
>
>
>
--
---
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 [email protected].
For more options, visit https://groups.google.com/d/optout.
