Using your awesome book and your github repository, I try to cover the very nicely elaborated neural networks.

Perhaps this is why I am rather a newbie to python: But when I assess the predicted variables after your code in In 17:

correct_predictions = tf.equal(tf.argmax(y_,0), tf.argmax(Y,0))
accuracy = tf.reduce_mean(tf.cast(correct_predictions, "float"))
print ("Accuracy:", accuracy.eval({X: test, Y: labels_test_, learning_rate: 0.001}, session = sess))
prediction = y_
mypreds = (prediction.eval(feed_dict={X: train}, session = sess))
cs = np.sum(mypreds,axis=0)
rs = np.sum(mypreds,axis=1)
print(cs)
print(rs)

I actually want to test whether the normalization of the outputs has properly worked, e.g. that probabilities among the 10 items 0-9 add up to 1. For all models before (I tested two of them), this works, but for the present case, I see that ROW sums are 1 and not column sums. Perhaps it is a very minor issue or I missed something.

Kind regards

Ciao Umberto,

I am sure you have locally the working example. I think the one in github what in the middle of test while playing with different length (21 versus 16).

I manage to run until section "Normal network"

train_size = 30768
test_size = 2000

sess = tf.Session()
sess.run(tf.global_variables_initializer())

training_epochs = 100
    
cost_history = []
for epoch in range(training_epochs+1):

    sess.run(optimizer, feed_dict = {X: np.asarray(train_input).reshape(15, train_size), 
                                     Y: np.asarray(train_output).reshape(16, train_size), learning_rate: .1})
    cost_ = sess.run(cost, feed_dict={ X:np.asarray(train_input).reshape(15, train_size), 
                                      Y: np.asarray(train_output).reshape(15, train_size), learning_rate: .1})

and I got:

ValueError: cannot reshape array of size 492288 into shape (15,30768)

I try to play with the size of training and testing set but in case of RNN I am not sure I understad what the shape should be.

Thanks
Cheers
Fabien