Hello, thanks for making this code available. It's useful because I'm trying to create a custom, LSTM cell (with additional logic for the "forget" gate), and you've clearly coded out the internal logic of the cell.

This isn't so much an issue on the repo I'm writing about. I'm wondering if you could explain and possibly show how to adapt the model for predicting a binary outcome. Let's say the new outcome variable is made like this:

from sklearn.preprocessing import OneHotEncoder, LabelBinarizer, LabelEncoder

_, Y = make_classification(n_samples = 118929, n_classes = 2, n_features=2, n_redundant=0, n_informative=1, n_clusters_per_class=1)
label_encoder = LabelEncoder()
integer_encoded = label_encoder.fit_transform(Y)
onehot_encoder = OneHotEncoder(sparse=False)
integer_encoded = integer_encoded.reshape(len(integer_encoded), 1)
onehot_encoded = onehot_encoder.fit_transform(integer_encoded)
Y = onehot_encoded

I feel very close but I'm getting errors here: tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=ys_)

Thank you!

thanks for this awesome implementation.

fyi, the input gate i = tf.sigmoid(tf.matmul(x,U[0]) + tf.matmul(st_1,W[0])) typically uses tanh as the activation function (unlike other gates, which should use sigmoid).

Thanks for the tutorial code. I wonder if the tf.scan version of a recurrent network is faster than its RNN implementation?

shape of ys_ is (?)
while shape of ys is (128, 20)

Am I missing something ?

UnicodeDecodeError: 'cp949' codec can't decode byte 0xe2 in position 9290: illegal multibyte sequence

Hi, I am following your vanilla RNN code. and I encounter an error above.
so, goggled and found the solution below.

• encoded file via 'cp949' makes this error sometimes
• fyi, I use anaconda, python 3, in windows 7

# line 25 of data.py
def read_lines(filename):
    #with open(filename) as f:
    with open(filename, 'rt', encoding='UTF8') as f:
        return f.read().split('\n')