def test_single_feature_lookup_1D(self, sigma_dimension):
emb_dim = 5 + sigma_dimension
config = test_util.default_de_config(emb_dim, [1] * emb_dim)
fea_embed = sparse_features.SparseFeatureEmbedding(
config, {'fea': (5, sigma_dimension)},
op_name='single_feature_%d' % sigma_dimension,
service_address=self._kbs_address)
embed, _, _, embed_map = fea_embed.lookup(['input1', 'input2'])
if sigma_dimension > 0:
self.assertEqual((2, 5), embed.shape)
else:
self.assertAllClose([[1] * 5, [1] * 5], embed)
self.assertEqual(['fea'], list(embed_map.keys()))
self.assertEqual((2, 5), embed_map['fea'].shape)
self.assertEqual(['fea'], list(fea_embed._variable_map.keys()))
def test_training_logistic(self):
self._config.gradient_descent_config.learning_rate = 0.05
fea_embed = sparse_features.SparseFeatureEmbedding(
self._config, {
'weather': (10, 2),
'day_of_week': (10, 2)
},
op_name='multiple_feature',
service_address=self._kbs_address)
model = tf.keras.models.Sequential(
[fea_embed,
tf.keras.layers.Dense(2, activation='softmax')])
model.compile(optimizer='sgd',
loss='categorical_crossentropy',
metrics=['accuracy'])
# Create an optimizer.
optimizer = tf.keras.optimizers.SGD(learning_rate=0.05)
x_train = {
'weather': [['cold', 'day'], ['hot', ''], ['warm', 'day'],
['warm', '']],
'day_of_week': [['monday', 'day'], ['tuesday', 'day'],
['sunday', ''], ['saturday', '']],
}
y_train = np.array([[0.0, 1.0], [0.0, 1.0], [1.0, 0.0], [1.0, 0.0]])
# Test the shape of model's output.
self.assertEqual((4, 2), model(x_train).shape)
loss_layer = tf.keras.losses.CategoricalCrossentropy(from_logits=True)
logits = model(x_train)
init_loss = loss_layer(logits, y_train)
for _ in range(10):
with tf.GradientTape() as tape:
logits = model(x_train)
loss = loss_layer(logits, y_train)
grads = tape.gradient(loss, model.trainable_weights)
# Update the trainable variables w.r.t. the logistic loss
optimizer.apply_gradients(zip(grads, model.trainable_weights))
print('===>loss: ', loss_layer(logits, y_train).numpy())
# Checks the loss is dropped after 10 steps of training.
logits = model(x_train)
final_loss = loss_layer(logits, y_train)
self.assertLess(final_loss.numpy(), init_loss.numpy())
def test_multiple_feature_lookup_2D_with_sigma(self):
fea_embed = sparse_features.SparseFeatureEmbedding(
self._config, {
'fea1': (5, 1),
'fea2': (10, 1)
},
op_name='multiple_feature2',
service_address=self._kbs_address)
embed, _, _, embed_map = fea_embed.lookup({
'fea1': [['input1', ''], ['input2', '']],
'fea2': [['input3', 'input5'], ['input4', 'input6']]
})
self.assertEqual((2, 15), embed.shape)
self.assertLen(embed_map.keys(), 2)
self.assertIn('fea1', embed_map.keys())
self.assertIn('fea2', embed_map.keys())
self.assertEqual((2, 2, 5), embed_map['fea1'].shape)
self.assertEqual((2, 2, 10), embed_map['fea2'].shape)
self.assertLen(fea_embed._variable_map.keys(), 2)
self.assertIn('fea1', fea_embed._variable_map.keys())
self.assertIn('fea2', fea_embed._variable_map.keys())
def test_multiple_feature_lookup_2D_without_sigma(self):
config = test_util.default_de_config(5, [1] * 5)
fea_embed = sparse_features.SparseFeatureEmbedding(
config, {
'fea1': (5, 0),
'fea2': (5, 0)
},
op_name='multiple_feature3',
service_address=self._kbs_address)
embed, _, _, embed_map = fea_embed.lookup({
'fea1': [['input1', ''], ['input2', '']],
'fea2': [['input3', 'input5'], ['input4', 'input6']]
})
self.assertAllClose([[1] * 10, [1] * 10], embed.numpy())
self.assertLen(embed_map.keys(), 2)
self.assertIn('fea1', embed_map.keys())
self.assertIn('fea2', embed_map.keys())
self.assertEqual((2, 2, 5), embed_map['fea1'].shape)
self.assertEqual((2, 2, 5), embed_map['fea2'].shape)
self.assertLen(fea_embed._variable_map.keys(), 2)
self.assertIn('fea1', fea_embed._variable_map.keys())
self.assertIn('fea2', fea_embed._variable_map.keys())