def test(self, model, cases):
sess = tf.get_default_session()
guarantee_initialized_variables(sess)
embeds = model.compute(model.embeds, cases)
primitive_embeddings = RLongPrimitiveEmbeddings(6)
# compute object embedding after applying projection
object_projection_layer = model._object_projection_layer
W, b = object_projection_layer.get_weights() # shapes [10, 6] and [6]
object_embed = np.ones(10).dot(W) + b
assert_array_almost_equal(
embeds[0],
np.concatenate((np.zeros(6), primitive_embeddings['r'],
primitive_embeddings[-1])))
assert_array_almost_equal(
embeds[1],
np.concatenate(
(np.zeros(6), np.zeros(6), primitive_embeddings['X1/1'])))
assert_array_almost_equal(
embeds[2],
np.concatenate(
(primitive_embeddings['b'], object_embed, object_embed)))
def test_outputs(self, model, inputs, output_tensors, outputs):
"""Test for correct output."""
sess = tf.get_default_session()
guarantee_initialized_variables(sess)
args, kwargs = inputs
test_outputs = model.compute(output_tensors, *args, **kwargs)
assert_array_collections_equal(outputs, test_outputs, decimal=4)
def __init__(self, config, save_dir):
super(Experiment, self).__init__(config, save_dir)
self.workspace.add_file('train_visualize', 'train_visualizer.txt')
self.workspace.add_file('valid_visualize', 'valid_visualizer.txt')
self.workspace.add_file('full_eval', 'full_eval_at_{step}.txt')
self.workspace.add_file('codalab', 'codalab.json')
self._domain = get_domain(config)
self._train_parse_model = self._build_train_parse_model()
self._decoder = self._build_decoder(self.train_parse_model)
self._train_visualizer = Visualizer(self.decoder,
self.workspace.train_visualize,
'train',
train=True)
self._valid_visualizer = Visualizer(self.decoder,
self.workspace.valid_visualize,
'valid',
train=False)
# Reload weights if they exist. Otherwise, initialize weights.
try:
self.saver.restore()
print 'Successfully reloaded the weights'
except IOError:
# NOTE: use this instead of tf.initialize_all_variables()!
# That op will overwrite Keras initializations.
sess = tf.get_default_session()
guarantee_initialized_variables(sess)
print 'Weights initialized'
def test_restore(self, tmpdir, v):
save_100_path = str(tmpdir.join('weights-100'))
save_10_path = str(tmpdir.join('weights-10'))
saver = Saver(str(tmpdir))
assign_op = tf.assign(v, 12)
sess = tf.get_default_session()
guarantee_initialized_variables(sess)
assert v.eval() == 5
saver.save(100) # save as step 100
sess.run(assign_op)
assert v.eval() == 12
saver.save(10) # save as step 10
saver.restore() # restores from the larger step number by default (100)
assert v.eval() == 5 # restored
saver.restore(10) # force restore number 10
assert v.eval() == 12
saver.restore(save_100_path)
assert v.eval() == 5
# latest should be the largest step number, not necessarily last saved
assert saver.latest_checkpoint == save_100_path
assert os.path.exists(save_100_path)
assert saver.checkpoint_paths == {
10: save_10_path,
100: save_100_path,
}
def test_value_function(value_function, weights, dummy_cases, rewards):
sess = tf.InteractiveSession()
guarantee_initialized_variables(sess)
fetch = {
"loss": value_function._loss
}
feed_dict = value_function.inputs_to_feed_dict(dummy_cases, rewards)
# Test that the loss decreases after taking a train step
loss = sess.run(fetch, feed_dict=feed_dict)["loss"]
values = value_function.values(dummy_cases)
for i in xrange(10):
vf_examples = [ValueFunctionExample(c, r) for c, r in zip(dummy_cases, rewards)]
value_function.train_step(vf_examples)
new_loss = sess.run(fetch, feed_dict=feed_dict)["loss"]
new_values = value_function.values(dummy_cases)
assert new_loss < loss
# Test that the weights didn't propagate to the ParseModel
fetch = {
"weights": value_function._parse_model._weights
}
model_weights = sess.run(fetch, feed_dict=feed_dict)["weights"]
assert np.array_equal(model_weights, weights)
def test_restore(self, tmpdir, v):
save_100_path = str(tmpdir.join('weights-100'))
save_10_path = str(tmpdir.join('weights-10'))
saver = Saver(str(tmpdir))
assign_op = tf.assign(v, 12)
sess = tf.get_default_session()
guarantee_initialized_variables(sess)
assert v.eval() == 5
saver.save(100) # save as step 100
sess.run(assign_op)
assert v.eval() == 12
saver.save(10) # save as step 10
saver.restore(
) # restores from the larger step number by default (100)
assert v.eval() == 5 # restored
saver.restore(10) # force restore number 10
assert v.eval() == 12
saver.restore(save_100_path)
assert v.eval() == 5
# latest should be the largest step number, not necessarily last saved
assert saver.latest_checkpoint == save_100_path
assert os.path.exists(save_100_path)
assert saver.checkpoint_paths == {
10: save_10_path,
100: save_100_path,
}
def test(self, model, cases):
sess = tf.get_default_session()
guarantee_initialized_variables(sess)
embeds = model.compute(model.embeds, cases)
primitive_embeddings = RLongPrimitiveEmbeddings(6)
# compute object embedding after applying projection
object_projection_layer = model._object_projection_layer
W, b = object_projection_layer.get_weights() # shapes [10, 6] and [6]
object_embed = np.ones(10).dot(W) + b
assert_array_almost_equal(embeds[0],
np.concatenate((np.zeros(6), primitive_embeddings['r'], primitive_embeddings[-1]))
)
assert_array_almost_equal(embeds[1],
np.concatenate((np.zeros(6), np.zeros(6), primitive_embeddings['X1/1']))
)
assert_array_almost_equal(embeds[2],
np.concatenate((primitive_embeddings['b'], object_embed, object_embed))
)