Sometimes all the examples in a batch contain only two labels, resulting

ValueError: Cannot feed value of shape (40, 2) for Tensor u'labels:0', which has shape '(?, 3)'

I tried to run the code in Win10 and my computer installed python3.6 with Anaconda 3, but I get trouble in twtokenizer, so I wonder whether the whole project is built on python2?When run the code the error comes in twtokenizer:
E:\tdlstm\src>python run.py
E:\software\Anaconda\lib\site-packages\gensim\utils.py:865: UserWarning: detected Windows; aliasing chunkize to chunkize_serial
warnings.warn("detected Windows; aliasing chunkize to chunkize_serial")
Traceback (most recent call last):
File "run.py", line 7, in
from optimise import TRAIN, TUNE, hyperoptTUNE, skoptTUNE
File "E:\tdlstm\src\optimise.py", line 9, in
from utils import load_data
File "E:\tdlstm\src\utils.py", line 6, in
import data.dataprocessor as dp
File "..\data\dataprocessor.py", line 8, in
from data.twtokenize import tokenize
File "..\data\twtokenize.py", line 166, in
unicode(regex_or(
NameError: name 'unicode' is not defined

The TD LSTM gives very inconsistent results everytime we train the model. What is the protocol used for reporting the results. Do we run it multiple times and report the best test accuracy?