def __init__(self, gpus):
super(OriginSparseModel, self).__init__()
self.embedding_layer = OriginalEmbedding(vocabulary_size=1737710,
embedding_vec_size=32,
initializer='uniform',
combiner='sum',
gpus=gpus)
def _fprop_v4_VS_tf():
print("[INFO]: Testing fprop_v4 vs tf...")
if vocabulary_size < slot_num:
raise RuntimeError("vocabulary_size must > slot.")
with tf.GradientTape(persistent=True) as tape:
# initial embedding table
init_value = np.float32(
np.random.normal(loc=0,
scale=1,
size=(vocabulary_size, embedding_vec_size)))
# input keys
# TODO: Keys in different slots should be unique.
input_keys = np.ones(shape=(batch_size, slot_num, max_nnz),
dtype=np.int64) * -1
each_slot = vocabulary_size // slot_num
nnz_0_num = 0
for batch_id in range(batch_size):
for slot_id in range(slot_num):
nnz = np.random.randint(
low=nnz_0_num, high=max_nnz + 1,
size=1)[0] # how many keys in this slot
if nnz == 0:
nnz_0_num = 1
if (embedding_type == 'distributed'):
keys = np.random.randint(low=slot_id * each_slot,
high=(slot_id + 1) *
each_slot,
size=nnz)
elif (embedding_type == "localized"):
# TODO: key should belong to that slot.
keys = []
while len(keys) < nnz:
key = np.random.randint(low=slot_id * each_slot,
high=(slot_id + 1) *
each_slot,
size=1)
if key % slot_num == slot_id:
keys.append(key)
input_keys[batch_id, slot_id, 0:nnz] = keys
# hugectr ops
hugectr_tf_ops.init(visiable_gpus=gpus,
key_type='int64',
value_type='float',
batch_size=batch_size,
batch_size_eval=len(gpus))
embedding_name = hugectr_tf_ops.create_embedding(
init_value=init_value,
opt_hparams=[0.1, 0.9, 0.99, 1e-5],
name_='hugectr_embedding',
max_vocabulary_size_per_gpu=(vocabulary_size // len(gpus)) * 2
+ 1,
slot_num=slot_num,
embedding_vec_size=embedding_vec_size,
max_feature_num=slot_num * max_nnz,
embedding_type=embedding_type,
max_nnz=max_nnz,
update_type='Global')
reshape_input_keys = np.reshape(input_keys, [-1, max_nnz])
indices = tf.where(reshape_input_keys != -1)
values = tf.gather_nd(reshape_input_keys, indices)
row_indices = tf.transpose(indices, perm=[1, 0])[0]
bp_trigger = tf.Variable(initial_value=1.0,
trainable=True,
dtype=tf.float32)
hugectr_forward = hugectr_tf_ops.fprop_v4(
embedding_name=embedding_name,
row_indices=row_indices,
values=values,
bp_trigger=bp_trigger,
is_training=True,
output_shape=[batch_size, slot_num, max_nnz])
# print("hugectr_results=\n", hugectr_forward)
# tf ops
reshape_input_keys = np.reshape(input_keys, [-1, max_nnz])
tf_indices = tf.where(reshape_input_keys != -1)
tf_values = tf.gather_nd(reshape_input_keys, tf_indices)
sparse_tensor = tf.sparse.SparseTensor(tf_indices, tf_values,
reshape_input_keys.shape)
# FIXME: if there are too more nnz=0 slots, tf.nn.embedding_lookup_sparse may get wrong results?
tf_embedding_layer = OriginalEmbedding(
vocabulary_size=vocabulary_size,
embedding_vec_size=embedding_vec_size,
initializer=init_value,
combiner='sum',
gpus=gpus)
tf_forward = tf_embedding_layer(
sparse_tensor,
output_shape=[batch_size, slot_num, embedding_vec_size])
# print("tf_results=\n", tf_forward)
# compare first forward result
try:
tf.debugging.assert_near(hugectr_forward, tf_forward)
except tf.errors.InvalidArgumentError as error:
raise error
print(
"[INFO]: The results from HugeCTR and tf in the first forward propagation are the same."
)
# backward
hugectr_grads = tape.gradient(hugectr_forward, bp_trigger)
tf_opt = tf.keras.optimizers.Adam(learning_rate=0.1,
beta_1=0.9,
beta_2=0.99,
epsilon=1e-5)
tf_grads = tape.gradient(tf_forward,
tf_embedding_layer.trainable_weights)
tf_opt.apply_gradients(
zip(tf_grads, tf_embedding_layer.trainable_weights))
# compare second forward result
hugectr_forward_2 = hugectr_tf_ops.fprop_v4(
embedding_name=embedding_name,
row_indices=row_indices,
values=values,
bp_trigger=bp_trigger,
is_training=True,
output_shape=[batch_size, slot_num, max_nnz])
tf_forward_2 = tf_embedding_layer(
sparse_tensor,
output_shape=[batch_size, slot_num, embedding_vec_size])
# print("hugectr 2:\n", hugectr_forward_2)
# print("tf 2:\n", tf_forward_2)
try:
tf.debugging.assert_near(hugectr_forward_2,
tf_forward_2,
rtol=1e-4,
atol=1e-5)
except tf.errors.InvalidArgumentError as error:
raise error
print(
"[INFO]: The results from HugeCTR and tf in the second forward propagation are the same."
)
hugectr_tf_ops.reset()
def _v2_fprop_v1_test():
print("[INFO]: Testing plugin_v2 fprop_experimental vs tf..")
if vocabulary_size < slot_num:
raise ValueError("vocabulary_size must > slot_num.")
# generate initial values
init_value, input_keys = generate_embedding_init_value_and_inputs()
# -------------------------------- hugectr ops ------------------------------------ #
class TestModel(tf.keras.models.Model):
def __init__(self, init_value, name_, embedding_type,
optimizer_type, max_vocabulary_size_per_gpu,
opt_hparams, update_type, atomic_update, scaler,
slot_num, max_nnz, max_feature_num,
embedding_vec_size, combiner):
super(TestModel, self).__init__()
self.input_buffer_reset = True if "distributed" == embedding_type else False
self.embedding_name = hugectr_tf_ops_v2.create_embedding(
init_value=init_value,
name_=name_,
embedding_type=embedding_type,
optimizer_type=optimizer_type,
max_vocabulary_size_per_gpu=max_vocabulary_size_per_gpu,
opt_hparams=opt_hparams,
update_type=update_type,
atomic_update=atomic_update,
scaler=scaler,
slot_num=slot_num,
max_nnz=max_nnz,
max_feature_num=max_feature_num,
embedding_vec_size=embedding_vec_size,
combiner=combiner)
def build(self, _):
self.bp_trigger = self.add_weight(name="bp_trigger",
shape=(1, ),
dtype=tf.float32,
trainable=True)
@tf.function
def call(self, row_offset, values, nnz, training=True):
replica_ctx = tf.distribute.get_replica_context()
result = hugectr_tf_ops_v2.fprop_experimental(
self.embedding_name,
replica_ctx.replica_id_in_sync_group,
row_offset,
values,
nnz,
self.bp_trigger,
input_buffer_reset=self.input_buffer_reset)
return result
hugectr_tf_ops_v2.init(visible_gpus=gpus,
seed=0,
key_type='int64',
value_type='float',
batch_size=batch_size,
batch_size_eval=len(gpus))
strategy = tf.distribute.MirroredStrategy(
devices=['/GPU:' + str(i) for i in gpus])
with strategy.scope():
hugectr_model = TestModel(
init_value=init_value,
name_='test_embedding',
embedding_type=embedding_type,
optimizer_type='Adam',
max_vocabulary_size_per_gpu=(vocabulary_size // len(gpus)) * 2
+ 1,
opt_hparams=[0.1, 0.9, 0.99, 1e-5],
update_type='Global',
atomic_update=True,
scaler=1.0,
slot_num=slot_num,
max_nnz=max_nnz,
max_feature_num=slot_num * max_nnz,
embedding_vec_size=embedding_vec_size,
combiner='sum')
opt = tf.keras.optimizers.Adam(learning_rate=0.1,
beta_1=0.9,
beta_2=0.99,
epsilon=1e-5)
# preprocess inputs
dataset_utils = CreateDataset(dataset_names=None,
feature_desc=None,
batch_size=batch_size,
n_epochs=None,
slot_num=slot_num,
max_nnz=max_nnz,
convert_to_csr=None,
gpu_count=len(gpus),
embedding_type=embedding_type,
get_row_indices=None)
if "distributed" == embedding_type:
row_offsets, value_tensors, nnz_array = dataset_utils._distribute_keys_for_distributed(
input_keys)
elif "localized" == embedding_type:
row_offsets, value_tensors, nnz_array = dataset_utils._distribute_keys_for_localized(
input_keys)
else:
raise ValueError("Not supported embedding_type %s." %
embedding_type)
# forward function
@tf.function
def hugectr_train_step(row_offset, values, nnz):
with tf.GradientTape() as tape:
forward_result = hugectr_model(row_offset, values, nnz)
grads = tape.gradient(forward_result,
hugectr_model.trainable_weights)
opt.apply_gradients(zip(grads, hugectr_model.trainable_weights))
return forward_result
# -------------------------------- tf ops ------------------------------------------- #
reshape_input_keys = np.reshape(input_keys, [-1, max_nnz])
tf_indices = tf.where(reshape_input_keys != -1)
tf_values = tf.gather_nd(reshape_input_keys, tf_indices)
sparse_tensor = tf.sparse.SparseTensor(tf_indices, tf_values,
reshape_input_keys.shape)
tf_embedding_layer = OriginalEmbedding(
vocabulary_size=vocabulary_size,
embedding_vec_size=embedding_vec_size,
initializer=init_value,
combiner='sum',
gpus=gpus)
tf_opt = tf.keras.optimizers.Adam(learning_rate=0.1,
beta_1=0.9,
beta_2=0.99,
epsilon=1e-5)
@tf.function
def tf_train_step(sparse_tensor):
with tf.GradientTape() as tape:
tf_forward = tf_embedding_layer(
sparse_tensor,
output_shape=[batch_size, slot_num, embedding_vec_size])
grads = tape.gradient(tf_forward,
tf_embedding_layer.trainable_weights)
tf_opt.apply_gradients(
zip(grads, tf_embedding_layer.trainable_weights))
return tf_forward
# ------------------ comparison ---------------------------------------------------- #
for iteration in range(2):
replica_row_offsets = PerReplica(row_offsets)
replica_values = PerReplica(value_tensors)
replica_nnz = PerReplica(nnz_array)
hugectr_forward = strategy.run(hugectr_train_step,
args=(replica_row_offsets,
replica_values, replica_nnz))
if len(gpus) > 1:
hugectr_forward = tf.concat(hugectr_forward.values, axis=0)
tf_forward = tf_train_step(sparse_tensor)
try:
tf.debugging.assert_near(hugectr_forward,
tf_forward,
rtol=1e-4,
atol=1e-5)
except tf.errors.InvalidArgumentError as error:
raise error
else:
print(
"[INFO]: The results from HugeCTR and tf in %d iteration are the same"
% (iteration + 1))
# --------------------- release resources -------------------------------------- #
hugectr_tf_ops_v2.reset()