def testEmbeddingVariableForGradientDescent(self):
ev = embedding_variable_ops.EmbeddingVariable(
embedding_dim=3,
ktype=dtypes.int64,
initializer=init_ops.ones_initializer(dtypes.float32))
def loss_fn(ev):
emb = embedding_ops.embedding_lookup(
ev, math_ops.cast([0, 1, 2, 5, 6, 7], dtypes.int64))
fun = math_ops.multiply(emb, 2.0, name='multiply')
loss = math_ops.reduce_sum(fun, name='reduce_sum')
return loss
gs = training_util.get_or_create_global_step()
opt = gradient_descent.GradientDescentOptimizer(0.1)
g_v = opt.compute_gradients(lambda: loss_fn(ev), [ev])
train_op = opt.apply_gradients(g_v)
emb = embedding_ops.embedding_lookup(
ev, math_ops.cast([0, 1, 2, 5, 6, 7], dtypes.int64))
init = variables.global_variables_initializer()
self.assertEqual(None, self.evaluate(init))
self.assertEqual(None, self.evaluate(train_op))
emb_result = self.evaluate(emb)
grad_result = self.evaluate(g_v[0][0])
for i in range(6):
for j in range(3):
self.assertAlmostEqual(.8, emb_result[i][j], delta=1e-05)
self.assertAlmostEqual(2.,
grad_result.values[i][j],
delta=1e-05)
def testEmbeddingVariableForSaveRestore(self):
ev = embedding_variable_ops.EmbeddingVariable(
embedding_dim=2,
initializer=init_ops.random_normal_initializer(),
ktype=dtypes.int64)
var_emb = embedding_ops.embedding_lookup(
ev, math_ops.cast([0, 1, 2], dtypes.int64))
loss = math_ops.reduce_sum(var_emb)
optimizer = gradient_descent.GradientDescentOptimizer(0.1)
with ops.control_dependencies([var_emb]):
opt = optimizer.minimize(loss)
saver = saver_module.Saver()
init = variables.global_variables_initializer()
with session.Session() as sess:
sess.run([init])
sess.run([opt, var_emb])
sess.run([opt, var_emb])
sess.run([opt, var_emb])
save = sess.run(var_emb)
saver.save(sess, "ckpt")
with session.Session() as sess:
saver.restore(sess, "ckpt")
restore = sess.run(var_emb)
for i in range(3):
for j in range(2):
self.assertAlmostEqual(save[i][j], restore[i][j], delta=1e-05)
def testEmbeddingVariableForTypeNotMatch(self):
with self.assertRaises(errors.InvalidArgumentError):
ev = embedding_variable_ops.EmbeddingVariable(
embedding_dim=3,
ktype=dtypes.int32,
initializer=init_ops.ones_initializer(dtypes.float32))
emb = embedding_ops.embedding_lookup(
ev, math_ops.cast([0, 1, 2, 5, 6, 7], dtypes.int64))
def testEmbeddingVariableForGeneralConstInitializer(self):
ev = embedding_variable_ops.EmbeddingVariable(
embedding_dim=3,
ktype=dtypes.int64,
initializer=init_ops.ones_initializer(dtypes.float32))
emb = embedding_ops.embedding_lookup(
ev, math_ops.cast([1, 6], dtypes.int64))
init = variables.global_variables_initializer()
self.assertEqual(None, self.evaluate(init))
self.assertAllEqual([[1., 1., 1.]] * 2, self.evaluate(emb))
def testEmbeddingVariableForGetShape(self):
ev = embedding_variable_ops.EmbeddingVariable(
embedding_dim=3, initializer=init_ops.ones_initializer(dtypes.float32))
emb = embedding_ops.embedding_lookup(
ev, math_ops.cast([0, 1, 2, 5, 6, 7], dtypes.int32))
shape = ev.total_count()
init = variables.global_variables_initializer()
self.assertEqual(None, self.evaluate(init))
self.evaluate(emb)
self.assertAllEqual([6, 3], self.evaluate(shape))