Hi,

You don't use input_var or target_var in your Theano function. So Theano
ignore there value. Did you wanted to define the model with input_var and
target_var instead of X and Y? If so, that could work by calling
set_value().

Frédéric

On Thu, May 11, 2017 at 5:55 PM Tara <[email protected]> wrote:

> I am trying to combine pymc3 with Theano for a simple recurrent neural
> network.However, when I complete training and change the input of the
> shared variables to the test set, the values are not updated in the graph
> even though the shared variables are updated.
> Any ideas will be appreciated.
> Here is the code :
>
> # CREATE PYMC3 + THEANO IMPLEMENTATION OF A SIMPLE RECURRENT NETWORK
> import timeit
> start = timeit.default_timer()
> import theano
> import theano.tensor as T
> import numpy as np
> import pymc3 as pm
> from scipy.stats import mode
> theano.config.compute_test_value = 'ignore'
>
> input_dim = 2
> output_dim = 2
> ### PARAMETERS OF THE MODEL ###
> hidden_dim = 64
> learning_rate = 0.1
> nb_epochs = 10
>
> np.random.seed(0)
>
> # Initialization /placeholder values
> X = T.dtensor3('X')
> Y = T.dtensor3('Y')
>
> # begin by generating dataset so we have an array of lists
> # ....
>
> NUM_EXAMPLES = 1500
> test_input = X_data[NUM_EXAMPLES:]
> test_output = y_data[NUM_EXAMPLES:]
>
> train_input = X_data[:NUM_EXAMPLES]
> train_output = y_data[:NUM_EXAMPLES]
>
> input_var = theano.shared(np.asarray(train_input).astype(np.float64),
> borrow = True)
> target_var = theano.shared(np.asarray(train_output).astype(np.float64),
> borrow = True)
>
> # Reference
> # From paper :IMPROVING PERFORMANCE OF RECURRENT NEURAL NETWORK WITH RELU
> NONLINEARITY
> def norm_positive_definite(r):
>     A = np.dot(r, r.transpose())/hidden_dim
>     values, vectors = np.linalg.eig(A)
>     e = np.amax(values)
>
>

-- 

--- 
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.

Reply via email to