Recurrent shop issues

data.py

import numpy as np


def get_data(n_samples, max_len):
    data = np.zeros([n_samples, max_len])
    labels = np.zeros([n_samples, 1])

    for row, label in zip(data, labels):
        length = np.random.randint(1, max_len+1)
        seq = np.random.normal(size=length)
        row[:length] += seq
        if length >= max_len/2:
            label += 1

    data = np.expand_dims(data, axis=-1)
    return data, labels

main.py

from RWACell import RWA
import data
from keras.layers import Input
from keras.models import Model

input_dim = 1
output_dim = 1
timesteps = 100

rwa = RWA(input_dim, output_dim)

inp = Input((timesteps, input_dim))
out = rwa(inp)
RWAModel = Model(inp, out)
RWAModel.compile(loss='binary_crossentropy', optimizer='adam')

train_data, train_labels = data.get_data(n_samples=1000, max_len=100)
RWAModel.fit(train_data, train_labels)

RWACell.py

'''
Machine Learning on Sequential Data Using a Recurrent
Weighted Average
- Jared Ostmeyer
- Lindsay Cowell
https://arxiv.org/pdf/1703.01253.pdf
'''

import tensorflow as tf
from recurrentshop import RecurrentModel
from keras.layers import Dense, Activation, Lambda, Input, concatenate
from keras.layers import add, multiply
from keras.backend import exp


def RWA(input_dim, output_dim):
    x = Input((input_dim, ))
    h_tm1 = Input((output_dim, ))
    n_tm1 = Input((output_dim, ))
    d_tm1 = Input((output_dim, ))

    x_h = concatenate([x, h_tm1])  # Concatenated vector

    u = Dense(output_dim)(x)
    g = Dense(output_dim, activation='tanh')(x_h)

    a = Dense(output_dim, use_bias=False)(x_h)
    e_a = Lambda(lambda x: exp(x))(a)

    z = multiply([u, g])
    nt = add([n_tm1, multiply([z, e_a])])
    dt = add([d_tm1, e_a])
    ht = Lambda(lambda x: tf.divide(1.0, x))(dt)
    ht = Activation('tanh')(ht)

    return RecurrentModel(input=x, output=ht,
                          initial_states=[h_tm1, n_tm1, d_tm1],
                          final_states=[ht, nt, dt],)