Hi everyone.
I am trying to check how the error changes and how the network works when
the weights initialization and bias are changed. I have change those in
Lenet deeplearning.net example.
I have changed in the convolutional class layer:
self.W = theano.shared(numpy.asarray(
rng.normal(mean, std, size=filter_shape),
dtype=theano.config.floatX
),
borrow=True
)
# the bias is a 1D tensor -- one bias per output feature map
b_values = numpy.ones((filter_shape[0],), dtype=theano.config.floatX)
And in the hidden layer the same:
if W is None:
W_values = numpy.asarray(rng.normal(mean, std, size = (n_in, n_out)
),
dtype=theano.config.floatX
)
if activation == theano.tensor.nnet.sigmoid:
W_values *= 4
W = theano.shared(value=W_values, name='W', borrow=True)
if b is None:
b_values = numpy.ones((n_out,), dtype=theano.config.floatX)
b = theano.shared(value=b_values, name='b', borrow=True)
The values of mean and std are 0 and 0.01 respectively. I have run the code
and in training/validation I am getting errors about 89-90% in first epochs
and then it gets stuck at 90.85% error rate
I am doing something wrong when I am initializing weights? or anybody could
help me to understand the results?
Thanks you in advance.
Beatriz.
--
---
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.
