# Initial Tensor for concatenating the tensors to after each while loop
tf_padded_final = tf.zeros(shape=[1,max_sent_seq_len,output_size * 2])

# 1-D vector of size = document batch size which is the number of sentences in a document (used for the tf.bidirectional_dynamic_rnn sequence lengths argument)
sentence_batch_len = tf.placeholder(shape=[None],dtype=tf.int32,name="sentence_batch_len")

# This is a 2-D array, with the first index being the start sentence index of the document to gather, and the second index is the last sentence index of the document to gather.
**sentence_index_offsets = tf.placeholder(shape=[None,2],dtype=tf.int32,name="sentence_index_offsets")**

i = tf.constant(0)

def while_cond (i, tf_padded_final):

    mb = tf.constant(mini_batch_size)
    return tf.less(i,mb)


# Loop through the mini batch of Documents (not sentences) one at a time
# And rollup the sentence vectors into a single row (1 row per document)
# While loop was used as there can be multiple sentence counts across documents
def body(i,tf_padded_final):

    #tf.print(i,[i])
    end_idx = sentence_index_offsets[i,1]
    st_idx = sentence_index_offsets[i,0]
    tf_range = tf.range(start=st_idx,limit=end_idx)
    pad_len = max_sent_seq_len - sentence_batch_len[i]

    tf_slice = tf.gather(outputs,tf_range)
    tf_slice_padding = [[0, pad_len], [0, 0]]
    tf_slice_padded = tf.pad(tf_slice, tf_slice_padding, 'CONSTANT')
    tf_slice_padded_3D = tf.expand_dims(tf_slice_padded, axis=0)

    tf_padded_final = tf.concat([tf_padded_final,tf_slice_padded_3D],axis=0)

    i = tf.add(i,1)

    return i, tf_padded_final

_, tf_padded_final_2 = tf.while_loop(while_cond, body, [i, tf_padded_final],shape_invariants=[i.get_shape(),tf.TensorShape([None,12,20])])