In [1]:

    

import tensorflow as tf
import numpy as np


    

In [2]:

    

h1 = tf.placeholder("float", 1)
h2 = tf.placeholder("float", 1)


    

In [3]:

    

def step(h_tm1, h_tm2):
    h_t = h_tm1 + h_tm2
    return h_t, h_tm1


    

In [4]:

    

outputs = []
h_tm1 = h1
h_tm2 = h2
for i in range(10):
    h_tm1, h_tm2 = step(h_tm1, h_tm2)
    outputs.append(h_tm1)


    

In [4]:

    

sess = tf.Session()


    

In [6]:

    

r = sess.run(outputs, feed_dict=({h2:np.zeros(1), h1:np.ones(1)}))


    

In [7]:

    

r


    

    
Out[7]:





[array([ 1.], dtype=float32),
 array([ 2.], dtype=float32),
 array([ 3.], dtype=float32),
 array([ 5.], dtype=float32),
 array([ 8.], dtype=float32),
 array([ 13.], dtype=float32),
 array([ 21.], dtype=float32),
 array([ 34.], dtype=float32),
 array([ 55.], dtype=float32),
 array([ 89.], dtype=float32)]

In [ ]:

    

"""
We can write our own "scan" for tensorflow so that we have a unifying API.

For LSTM and GRU, TF already provides a fast implementation
"""


    

In [21]:

    

def scan(step, sequences=None, outputs_info=None, non_sequences=None, n_steps=None):
    #if sequences is None:
    #    assert n_steps is not None
    
    leng = n_steps
    outputs = outputs_info
    results = [[] for i in outputs_info]
    for i in range(leng):
        seq = [s[i] for s in sequences]
        nseq = [n[i] for n in non_sequences]
        input_list = seq + outputs + nseq
        outputs = list(step(*input_list))
        for i, o in enumerate(outputs):
            results[i].append(outputs[i])
    results = [tf.concat(0, r) for r in results]
    return results


    

In [22]:

    

outputs = scan(step, sequences=[], outputs_info=[h1, h2], non_sequences=[], n_steps=10)


    

In [23]:

    

r = sess.run(outputs, feed_dict=({h2:np.zeros(1), h1:np.ones(1)}))


    

In [24]:

    

r


    

    
Out[24]:





[array([  1.,   2.,   3.,   5.,   8.,  13.,  21.,  34.,  55.,  89.], dtype=float32),
 array([  1.,   1.,   2.,   3.,   5.,   8.,  13.,  21.,  34.,  55.], dtype=float32)]