Deep undertanding about tutorials

[MorvanZhou]

Hi Gisselle,

Your understanding is right, there are three Weight matrixes all the time.
The weights['in'] you have seen in the example is W_hx, weights['out'] is
W_yh. The third one W_hh is in the cell, we just need to define the hidden
units size for the cell instead of defining W_hh.

Hope this would help you.

Regards,
Morvan

On 28 October 2016 at 11:45, Gissella Bejarano [email protected]
wrote:

Hi,
Can you recommend any source to try to understand the real architecture
done with tensorflow in the classification of RNN (MNIST). I thought that
the simplest RNN would work with 3 Weight matrixes. W_hx, W_hh and W_yh but
I just saw two in the tutorial and I don't understand very well why.

Regards,
Gissella B.

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
#19, or mute the thread
https://github.com/notifications/unsubscribe-auth/ASgmNEi3BByT-mLPOOsCKV-mBGvCavzXks5q4UXBgaJpZM4Ki-Qd
.

[gissemari]

Oh, I see..thanks for the explanation!
So BasicLSTMCell() would create a hidden recurrent layer which would create the W_hh matrix? One more question. I have a testSet of different size of a batch size... is there a way to change the batchsize we used here: lstm_cell.zero_state(batch_size, dtype=tf.float32)?

Regards,
Gissella B.

[MorvanZhou]

BasicLSTMCell() would create a hidden recurrent layer which would create
the W_hh matrix?

Answer: Yes, it will.

I have a testSet of different size of a batch size... is there a way to
change the batchsize we used here: lstm_cell.zero_state(batch_size,
dtype=tf.float32)?

Answer: I also faced this issue (my issue is the different time_steps in
training and testing), and my solution is to create a test-RNN with new
parameters and restore the Weights and biases from a saved Train-RNN. I
don't think this is a good solution but I don't see other better answers.

On 29 October 2016 at 01:07, Gissella Bejarano [email protected]
wrote:

Oh, I see..thanks for the explanation!
So BasicLSTMCell() would create a hidden recurrent layer which would
create the W_hh matrix? One more question. I have a testSet of different
size of a batch size... is there a way to change the batchsize we used
here: lstm_cell.zero_state(batch_size, dtype=tf.float32)?

Regards,
Gissella B.

—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
#19 (comment),
or mute the thread
https://github.com/notifications/unsubscribe-auth/ASgmNKc5kHMWrllKjhC3QqjPZBxnFi2hks5q4gG4gaJpZM4Ki-Qd
.

[gissemari]

Thank you. I managed to fixed the time_steps of all my instances and I am doing cross validation so I can have equal-size training batchs and test batch. I'm still looking for an example to manage different time_steps, there should be a more formal solution than the one you mention right? What if instances in the same training set has different time_steps?

[MorvanZhou]

I figured out a new method to deal with the different time steps by using variable_scope: https://github.com/MorvanZhou/tutorials/blob/master/tensorflowTUT/tf22_scope/tf22_RNN_scope.py