def do_train(session, model):
return train_on_copy_task(session, model,
length_from=3, length_to=8,
vocab_lower=2, vocab_upper=10,
batch_size=100,
max_batches=5000,
batches_in_epoch=1000,
verbose=False)
# 2x encoder state size
model = Seq2SeqModel(encoder_cell=LSTMCell(10),
decoder_cell=LSTMCell(20),
vocab_size=10,
embedding_size=10,
attention=True,
bidirectional=True,
debug=False)
session.run(tf.global_variables_initializer())
train_on_copy_task(session,
model,
length_from=3,
length_to=8,
vocab_lower=2,
vocab_upper=10,
batch_size=100,
max_batches=3000,
batches_in_epoch=1000,
verbose=True)
# ## Fun exercise, compare performance of different seq2seq variants.
#
# Comparison will be done using train loss tracks, since the task is algorithmic and data is generated directly from true distribution and out-of-sample testing doesn't really make sense.
# In[4]:
loss_tracks = dict()
def do_train(session, model):