def test_build(self):
with self.test_session(config=self.sess_config) as sess:
m = model.SeqModel(check_feed_dict=False)
opt = {
'emb:vocab_size': 20,
'emb:dim': 5,
'cell:num_units': 10,
'cell:cell_class': 'tensorflow.nn.rnn_cell.BasicLSTMCell',
'logit:output_size': 2
}
expected_vars = {
't/embedding:0': (20, 5),
't/rnn/basic_lstm_cell/kernel:0': (10 + 5, 10 * 4),
't/rnn/basic_lstm_cell/bias:0': (10 * 4, ),
't/logit_w:0': (2, 10),
't/logit_b:0': (2, )
}
n = m.build_graph(opt, name='t')
for v in tf.global_variables():
self.assertTrue(v.name in expected_vars,
'expected variable scope/name')
self.assertEqual(v.shape, expected_vars[v.name],
'shape is correct')
for k, v in m._fetches.items():
if k is not None:
self.assertNotEqual(v[0], v[1], 'fetch array is set')
def test_train(self):
logger = util.get_logger(level='warning')
data = generator.read_seq_data(
self.gen(), self.vocab, self.vocab, keep_sentence=False, seq_len=20)
batch_iter = partial(generator.seq_batch_iter, *data,
batch_size=13, shuffle=True, keep_sentence=False)
with self.test_session(config=self.sess_config) as sess:
m = tfm.SeqModel()
n = m.build_graph({'rnn:fn': 'seqmodel.graph.scan_rnn',
'cell:num_layers': 2})
m.set_default_feed('train_loss_denom', 13)
optimizer = tf.train.AdamOptimizer()
train_op = optimizer.minimize(m.training_loss)
sess.run(tf.global_variables_initializer())
train_run_fn = partial(run.run_epoch, sess, m, batch_iter, train_op=train_op)
eval_run_fn = partial(run.run_epoch, sess, m, batch_iter)
def stop_early(*args, **kwargs):
pass
eval_info = eval_run_fn()
train_state = run.train(train_run_fn, logger, max_epoch=3,
valid_run_epoch_fn=eval_run_fn,
end_epoch_fn=stop_early)
self.assertLess(train_state.best_eval, eval_info.eval_loss,
'after training, eval loss is lower.')
self.assertEqual(train_state.cur_epoch, 3, 'train for max epoch')
def test_run_epoch(self):
data = generator.read_seq_data(self.gen(),
self.vocab,
self.vocab,
keep_sentence=False,
seq_len=20)
batch_iter = partial(generator.seq_batch_iter,
*data,
batch_size=13,
shuffle=True,
keep_sentence=False)
with self.test_session(config=self.sess_config) as sess:
m = tfm.SeqModel()
n = m.build_graph()
optimizer = tf.train.AdamOptimizer()
train_op = optimizer.minimize(m.training_loss)
sess.run(tf.global_variables_initializer())
run_fn = partial(run.run_epoch, sess, m, batch_iter)
eval_info = run_fn()
self.assertEqual(eval_info.num_tokens,
self.num_lines + self.num_tokens,
'run uses all tokens')
self.assertAlmostEqual(eval_info.eval_loss,
np.log(self.vocab.vocab_size),
places=1,
msg='eval loss is close to uniform.')
for __ in range(3):
train_info = run_fn(train_op=train_op)
self.assertLess(train_info.eval_loss, eval_info.eval_loss,
'after training, eval loss is lower.')
def test_build_reuse(self):
with self.test_session(config=self.sess_config) as sess:
m = model.SeqModel(check_feed_dict=False)
n = m.build_graph()
num_vars = len(tf.global_variables())
n = m.build_graph(reuse=True, no_dropout=True)
self.assertEqual(type(n['cell']), tf.nn.rnn_cell.BasicLSTMCell,
'overwrite default options with kwargs')
num_vars_ = len(tf.global_variables())
self.assertEqual(num_vars, num_vars,
'no new variables when reuse is True')
def test_build_no_output(self):
with self.test_session(config=self.sess_config) as sess:
m = model.SeqModel(check_feed_dict=False)
n = m.build_graph(**{'out:logit': False, 'out:loss': False})
self.assertTrue('logit' not in n, 'logit is not in nodes')
for k, v in m._fetches.items():
if k is None or k == m._TRAIN_ or k == m._EVAL_:
self.assertEqual(v[0], v[1], 'fetch array is not set')
if k == m._PREDICT_:
self.assertNotEqual(v[0], v[1],
'predict fetch array is set')
self.assertRaises(ValueError, m.build_graph,
**{'out:logit': False})
def test_build_overwrite_opt(self):
with self.test_session(config=self.sess_config) as sess:
m = model.SeqModel(check_feed_dict=False)
opt = {
'emb:vocab_size': 20,
'cell:in_keep_prob': 0.5,
'logit:output_size': 2
}
n = m.build_graph(opt)
self.assertEqual(n['emb_vars'].get_shape()[0], 20,
'overwrite default options')
self.assertEqual(type(n['cell']), tf.nn.rnn_cell.DropoutWrapper,
'overwrite default options')
n = m.build_graph(opt, reuse=True, **{'cell:in_keep_prob': 1.0})
self.assertEqual(type(n['cell']), tf.nn.rnn_cell.BasicLSTMCell,
'overwrite default options with kwargs')
def test_build_decode(self):
with self.test_session(config=self.sess_config) as sess:
m = model.SeqModel(check_feed_dict=False)
n = m.build_graph(
**{
'out:logit': True,
'out:decode': True,
'decode:add_greedy': True,
'decode:add_sampling': True
})
self.assertTrue('decode_greedy' in n, 'decode_greedy in nodes')
self.assertTrue('decode_greedy' in m._predict_fetch,
'decode_greedy in predict dict')
self.assertTrue('decode_sampling' in n, 'decode_sampling in nodes')
self.assertTrue('decode_sampling' in m._predict_fetch,
'decode_sampling in predict dict')