def pyramid_hash_layer(x, hash_size, dict_size, num_hash=16, emb_size=256, min_win_size=2, max_win_size=4, shared=True):
emb_list = []
for win_size in six.moves.xrange(min_win_size, max_win_size + 1):
print("pyramid_hash_layer input shape ", x.shape)
seq_enum = fluid.layers.sequence_enumerate(x, win_size)
print("seq_enum.shape ", seq_enum.shape)
xxhash = fluid.layers.hash(seq_enum, hash_size=hash_size, num_hash=num_hash)
print("xxhash.shape ", xxhash.shape)
pool = layers.fused_embedding_seq_pool(
input=xxhash, size=[hash_size, emb_size // num_hash], is_sparse=True,
param_attr=fluid.ParamAttr(
name='PyramidHash_emb_0',
learning_rate=640,
initializer=fluid.initializer.Uniform(low=-math.sqrt(6.0 / (emb_size * dict_size)), high=math.sqrt(6.0 / (emb_size * dict_size)), seed=0)) if shared else None,
dtype='float32')
# emb = layers.embedding(
# input=xxhash, size=[hash_size, emb_size // num_hash], is_sparse=True,
# param_attr=fluid.ParamAttr(
# name='PyramidHash_emb_0',
# learning_rate=640,
# initializer=fluid.initializer.Uniform(low=-math.sqrt(6.0 / (emb_size * dict_size)), high=math.sqrt(6.0 / (emb_size * dict_size)), seed=0)) if shared else None,
# dtype='float32')
# print(emb.name, emb.shape)
# reshape_emb = layers.reshape(x=emb, shape=[-1, emb_size], inplace=True)
# pool = fluid.layers.sequence_pool(input=reshape_emb, pool_type='sum')
emb_list.append(pool)
return emb_list
def emb_layer(x, dict_size, emb_size=256, shared=True):
print("emb_layer input shape ", x.shape)
pool = layers.fused_embedding_seq_pool(
input=x,
size=[dict_size, emb_size],
is_sparse=True,
param_attr=fluid.ParamAttr(name='EmbeddingWithVSum_emb_0',
learning_rate=640,
initializer=fluid.initializer.Uniform(
low=-math.sqrt(6.0 /
(emb_size * dict_size)),
high=math.sqrt(6.0 /
(emb_size * dict_size)),
seed=0)) if shared else None,
dtype='float32')
# emb = layers.embedding(
# input=x, size=[dict_size, emb_size], is_sparse=True,
# param_attr=fluid.ParamAttr(
# name='EmbeddingWithVSum_emb_0',
# learning_rate=640,
# initializer=fluid.initializer.Uniform(low=-math.sqrt(6.0 / (emb_size * dict_size)), high=math.sqrt(6.0 / (emb_size * dict_size)), seed=0)) if shared else None,
# dtype='float32')
# pool = fluid.layers.sequence_pool(input=emb, pool_type='sum')
return pool
def emb_layer(self, x, dict_size, emb_size=256, shared=True):
pool = layers.fused_embedding_seq_pool(
input=x,
size=[dict_size, emb_size],
is_sparse=True,
param_attr=fluid.ParamAttr(
name='EmbeddingWithVSum_emb_0',
learning_rate=640,
initializer=fluid.initializer.Uniform(
low=-math.sqrt(6.0 / (emb_size * dict_size)),
high=math.sqrt(6.0 / (emb_size * dict_size)),
seed=0)) if shared else None,
dtype='float32')
return pool
def pyramid_hash_layer(self,
x,
hash_size,
dict_size,
num_hash=16,
emb_size=256,
min_win_size=2,
max_win_size=4,
shared=True,
name=''):
emb_list = []
for win_size in six.moves.xrange(min_win_size, max_win_size + 1):
seq_enum = fluid.layers.sequence_enumerate(x, win_size)
xxhash = fluid.layers.hash(seq_enum,
hash_size=hash_size * num_hash,
num_hash=num_hash)
pool = layers.fused_embedding_seq_pool(
input=xxhash,
size=[hash_size * num_hash, emb_size // num_hash],
is_sparse=True,
param_attr=fluid.ParamAttr(
name='PyramidHash_emb_0',
learning_rate=640,
initializer=fluid.initializer.Uniform(
low=-math.sqrt(6.0 / (emb_size * dict_size)),
high=math.sqrt(6.0 / (emb_size * dict_size)),
seed=0)) if shared else None,
dtype='float32')
self.var_append(seq_enum, name + "seq_enum_w" + str(win_size))
self.var_append(xxhash, name + "xxhash_w" + str(win_size))
self.var_append(pool, name + "pool_w" + str(win_size))
emb_list.append(pool)
return emb_list