def test_compute_sampled_logits_grad(self):
cs_config = cs_config_builder.build_candidate_sampler_config(
cs_config_builder.negative_sampler(unique=True,
algorithm='UNIFORM'))
de_config = test_util.default_de_config(3, cs_config=cs_config)
# Add a few embeddings into knowledge bank.
de_ops.dynamic_embedding_update(['key1', 'key2', 'key3'],
tf.constant([[1.0, 2.0, 3.0],
[4.0, 5.0, 6.0],
[7.0, 8.0, 9.0]]),
de_config,
'emb',
service_address=self._kbs_address)
# A simple one layer NN model.
# Input data: x = [[1, 2], [3, 4]].
# Weights from input to logit output layer: W = [[1, 2, 3], [4, 5, 6]].
# Input activation at output layer i = x*W = [[9, 12, 15], [19, 26, 33]].
# Logits output therefore becomes E*i, where E are the embeddings of output
# keys, i.e., E = [[1, 2, 3], [4, 5, 6], [7, 8, 9]].
# Then the logits output becomes [[78, 186, 294], [170, 404, 638]]
#
# If we define the loss to be L = tf.reduced_sum(Logits), then
# dL/dE = sum_by_key(i) = [[28, 38, 48], [28, 38, 48], [28, 38, 48]].
# So the expected new embeddings become
# E - 0.1 * dL/dE = [[-1.8, -1.8, -1.8], [1.2, 1.2, 1.2], [4.2, 4.2, 4.2]].
weights = tf.Variable([[1, 2, 3], [4, 5, 6]], dtype=tf.float32)
inputs = tf.constant([[1.0, 2.0], [3.0, 4.0]])
with tf.GradientTape() as tape:
logits, _, _, _, _ = cs_ops.compute_sampled_logits(
[['key1', ''], ['key2', 'key3']],
tf.matmul(inputs, weights),
3,
de_config,
'emb',
service_address=self._kbs_address)
loss = tf.reduce_sum(logits)
# Applies the gradient descent.
grads = tape.gradient(loss, weights)
# The gradients updated by the knowledge bank.
updated_embedding = de_ops.dynamic_embedding_lookup(
['key1', 'key2', 'key3'],
de_config,
'emb',
service_address=self._kbs_address)
self.assertAllClose(
updated_embedding,
[[-1.8, -1.8, -1.8], [1.2, 1.2, 1.2], [4.2, 4.2, 4.2]])
# The gradients w.r.t. the weight W is calculated as
# dL/dw = dL/di * di/dW = sum_by_dim(E) * x =
# [12, 15, 18] * [[4, 4, 4], [6, 6, 6]] = [[48, 60, 72], [72, 90, 108]]
self.assertAllClose(grads, [[48, 60, 72], [72, 90, 108]])
def test_build_candidate_sampler_config_success(self):
self.assertProtoEquals(
"""
extension {
[type.googleapis.com/carls.candidate_sampling.BruteForceTopkSamplerConfig] {
similarity_type: COSINE
}
}
""",
cs_config_builder.build_candidate_sampler_config(
cs_config_builder.brute_force_topk_sampler('COSINE')))
self.assertProtoEquals(
"""
extension {
[type.googleapis.com/carls.candidate_sampling.LogUniformSamplerConfig] {
unique: true
}
}
""",
cs_config_builder.build_candidate_sampler_config(
cs_config_builder.log_uniform_sampler(True)))
def test_build_candidate_sampler_config_success(self):
self.assertProtoEquals(
"""
extension {
[type.googleapis.com/carls.candidate_sampling.BruteForceTopkSamplerConfig] {
similarity_type: COSINE
}
}
""",
cs_config_builder.build_candidate_sampler_config(
cs_config_builder.brute_force_topk_sampler('COSINE')))
self.assertProtoEquals(
"""
extension {
[type.googleapis.com/carls.candidate_sampling.NegativeSamplerConfig] {
unique: true
sampler: UNIFORM
}
}
""",
cs_config_builder.build_candidate_sampler_config(
cs_config_builder.negative_sampler(True, 'UNIFORM')))
def test_brute_force_topk(self):
cs_config = cs_config_builder.build_candidate_sampler_config(
cs_config_builder.brute_force_topk_sampler('DOT_PRODUCT'))
de_config = test_util.default_de_config(2, cs_config=cs_config)
# Add a few embeddings into knowledge bank.
de_ops.dynamic_embedding_update(['key1', 'key2', 'key3'],
tf.constant([[2.0, 4.0], [4.0, 8.0],
[8.0, 16.0]]),
de_config,
'emb',
service_address=self._kbs_address)
keys, logits = cs_ops.top_k([[1.0, 2.0], [-1.0, -2.0]],
3,
de_config,
'emb',
service_address=self._kbs_address)
self.assertAllEqual(
keys.numpy(),
[[b'key3', b'key2', b'key1'], [b'key1', b'key2', b'key3']])
self.assertAllClose(logits.numpy(), [[40, 20, 10], [-10, -20, -40]])
def test_build_candidate_sampler_config_failed(self):
with self.assertRaises(ValueError):
cs_config_builder.build_candidate_sampler_config(100)
with self.assertRaises(ValueError):
cs_config_builder.build_candidate_sampler_config('invalid')
def test_compute_sampled_logits(self):
cs_config = cs_config_builder.build_candidate_sampler_config(
cs_config_builder.negative_sampler(unique=True,
algorithm='UNIFORM'))
de_config = test_util.default_de_config(3, cs_config=cs_config)
# Add a few embeddings into knowledge bank.
de_ops.dynamic_embedding_update(['key1', 'key2', 'key3'],
tf.constant([[1.0, 2.0, 3.0],
[4.0, 5.0, 6.0],
[7.0, 8.0, 9.0]]),
de_config,
'emb',
service_address=self._kbs_address)
# Sample logits.
logits, labels, keys, mask, weights = cs_ops.compute_sampled_logits(
[['key1', ''], ['key2', 'key3']],
tf.constant([[2.0, 4.0, 1], [-2.0, -4.0, 1]]),
3,
de_config,
'emb',
service_address=self._kbs_address)
# Expected results:
# - Example one returns one positive key {'key2'} and two negative keys
# {'key2', 'key3'}.
# - Example two returns two positive keys {'key2', 'key3'} and one
# positive key {'key1'}.
expected_weights = {
b'key1': [1, 2, 3],
b'key2': [4, 5, 6],
b'key3': [7, 8, 9]
}
expected_labels = [{
b'key1': 1,
b'key2': 0,
b'key3': 0
}, {
b'key1': 0,
b'key2': 1,
b'key3': 1
}]
# Logit for example one:
# - 'key1': [2, 4, 1] * [1, 2, 3] = 13
# - 'key2': [2, 4, 1] * [4, 5, 6] = 34
# - 'key3': [2, 4, 1] * [7, 8, 9] = 55
# Logit for example two:
# - 'key1': [-2, -4, 1] * [1, 2, 3] = -7
# - 'key2': [-2, -4, 1] * [4, 5, 6] = -22
# - 'key3': [-2, -4, 1] * [7, 8, 9] = -37
expected_logits = [{
b'key1': 13,
b'key2': 34,
b'key3': 55
}, {
b'key1': -7,
b'key2': -22,
b'key3': -37
}]
# Check keys and weights.
for b in range(2):
self.assertEqual(1, mask.numpy()[b])
for key in {b'key1', b'key2', b'key3'}:
self.assertIn(key, keys.numpy()[b])
for i in range(3):
key = keys.numpy()[b][i]
self.assertAllClose(expected_weights[key],
weights.numpy()[b][i])
self.assertAllClose(expected_labels[b][key],
labels.numpy()[b][i])
self.assertAllClose(expected_logits[b][key],
logits.numpy()[b][i])