def add_variable_inputs(sym, sym_input_dict, is_training):
arg_shape, _, aux_shape = sym.infer_shape()
arg_type, _, aux_type = sym.infer_type()
arg_shape_type = zip(sym.list_arguments(), arg_shape, arg_type)
aux_shape_type = zip(sym.list_auxiliary_states(), aux_shape, aux_type)
node_info = json.loads(sym.tojson())["nodes"]
for item in arg_shape_type:
import xdl.python.framework.variable as variable
import xdl.python.backend.mxnet.convert_utils as cu
if item[0] not in sym_input_dict:
initializer_and_args = get_initializer_and_args(item[0], node_info)
xdl_var = variable.Variable(
name=item[0],
shape=item[1],
dtype=cu.MX2XDL.convert_type(item[2]),
trainable=True,
initializer=cu.MX2XDL.convert_initializer(
initializer_and_args[0], initializer_and_args[1]))
sym_input_dict[item[0]] = xdl_var.value
for item in aux_shape_type:
if item[0].endswith('_moving_mean') or \
item[0].endswith('_moving_var'):
initializer_and_args = get_initializer_and_args(item[0], node_info)
xdl_var = variable.Variable(
name=item[0],
shape=item[1],
dtype=cu.MX2XDL.convert_type(item[2]),
trainable=True,
initializer=cu.MX2XDL.convert_initializer(
initializer_and_args[0], initializer_and_args[1]))
if not is_training:
sym_input_dict[item[0]] = xdl_var.value
def merged_embedding(name,
sparse_inputs,
initializer,
emb_dim,
feature_dim,
combiner='sum',
vtype=VarType.Index,
length=50,
reverse=False):
"""xdl embedding
Args:
name: name for embedding, will be used for declaring variable on ps-plus
sparse_inputs: a list of sparse tensors represent input data
initializer: intializer for the weights
emb_dim: embedding dimension
feature_dim: sparse input dimension, for pre-allocate memory
combiner: reduce operator, support sum|mean
Returns:
a tensor represent embedding result
Raises:
None
"""
import xdl.python.framework.variable as variable
var = variable.Variable(name=name,
dtype=DataType.float,
shape=[feature_dim, emb_dim],
initializer=initializer,
vtype=vtype,
trainable=True)
merged_sparse_inputs = merge_sparse(sparse_inputs)
ids = merged_sparse_inputs.ids
unique_ids, idx = xdl.unique(ids, itype=DataType.int32)
embeddings = var.gather(unique_ids, save_ratio=feature_add_probability)
global _EMBEDDING_TENSOR
_EMBEDDING_TENSOR[embeddings] = var
import xdl.python.sparse_engine.embedding_ops as embedding_ops
if combiner == 'sum':
embeddings = embedding_ops.merged_ksum(embeddings, idx,
merged_sparse_inputs.values,
merged_sparse_inputs.segments,
merged_sparse_inputs.groups)
elif combiner == 'mean':
embeddings = embedding_ops.merged_kmean(embeddings, idx,
merged_sparse_inputs.values,
merged_sparse_inputs.segments,
merged_sparse_inputs.groups)
elif combiner == 'tile':
embeddings = embedding_ops.merged_tile(embeddings, idx,
merged_sparse_inputs.values,
merged_sparse_inputs.segments,
merged_sparse_inputs.groups,
length, reverse)
else:
raise Exception("Unrecognized combiner:" + str(combiner))
emb_info = EmbeddingInfo(name, feature_dim, emb_dim, combiner, None, var,
embeddings)
set_embedding_info([var], emb_info)
return embeddings
def merged_embedding(name, sparse_inputs, initializer, emb_dim, feature_dim,
combiner='sum', vtype=VarType.Index, length=50, reverse=False,
batch_read=3000, feature_add_probability=1.0, cbf=0, device='CPU', **device_attr):
"""xdl embedding
Args:
name: name for embedding, will be used for declaring variable on ps-plus
sparse_inputs: a list of sparse tensors represent input data
initializer: intializer for the weights
emb_dim: embedding dimension
feature_dim: sparse input dimension, for pre-allocate memory
combiner: reduce operator, support sum|mean
Returns:
a tensor represent embedding result
Raises:
None
"""
import xdl.python.framework.variable as variable
with variable.variable_info(batch_read=batch_read, save_ratio=feature_add_probability, bloom_filter=cbf):
var = variable.Variable(name=name,
dtype=DataType.float,
shape=[feature_dim, emb_dim],
initializer=initializer,
vtype=vtype,
trainable = True)
if isinstance(sparse_inputs, (list, tuple)):
merged_sparse_inputs = merge_sparse(sparse_inputs)
emb_dim *= len(sparse_inputs)
else:
assert(isinstance(sparse_inputs, MergedSparseTensor))
merged_sparse_inputs = sparse_inputs
if merged_sparse_inputs.has_unique_ids():
unique_ids = merged_sparse_inputs.ids
idx = merged_sparse_inputs.indices
sidx = merged_sparse_inputs.sidx
sseg = merged_sparse_inputs.sseg
else:
with xdl.device(device, **device_attr):
unique_ids, idx, sidx, sseg = xdl.unique(ids, merged_sparse_inputs.groups, itype=DataType.int32)
embeddings = var.gather(unique_ids)
global _EMBEDDING_TENSOR
_EMBEDDING_TENSOR[embeddings] = var
import xdl.python.sparse_engine.embedding_ops as embedding_ops
if combiner == 'sum':
embeddings = embedding_ops.merged_ksum(
embeddings,
idx,
merged_sparse_inputs.values,
merged_sparse_inputs.segments,
merged_sparse_inputs.groups,
sidx,
sseg,
device, **device_attr)
elif combiner == 'mean':
embeddings = embedding_ops.merged_kmean(
embeddings,
idx,
merged_sparse_inputs.values,
merged_sparse_inputs.segments,
merged_sparse_inputs.groups,
sidx,
sseg,
device, **device_attr)
elif combiner == 'tile':
embeddings = embedding_ops.merged_tile(
embeddings,
idx,
merged_sparse_inputs.values,
merged_sparse_inputs.segments,
merged_sparse_inputs.groups,
length,
reverse,
device, **device_attr)
else:
raise Exception("Unrecognized combiner:" + str(combiner))
emb_info = EmbeddingInfo(name, feature_dim, emb_dim, combiner, None, var, length, embeddings)
set_embedding_info([var], emb_info)
return embeddings
def embedding(name, sparse_input, initializer, emb_dim, feature_dim,
combiner='sum',
vtype=VarType.Index,
length=50,
reverse=False,
batch_read=3000,
feature_add_probability=1.0,
cbf=0,
device='CPU',
statis_list=None,
statis_decay=0.07,
statis_decay_period=100,
labels=None,
save=True,
**device_attr):
"""xdl embedding
Args:
name: name for embedding, will be used for declaring variable on ps-plus
sparse_input: a sparse tensor represent input data
initializer: intializer for the variable on ps-plus
emb_dim: embedding dimension
feature_dim: sparse input dimension, for pre-allocate memory
combiner: reduce operator, support sum|mean
Returns:
a tensor represent embedding result
Raises:
None
"""
global EMBEDDING_LIST, EMBEDDING_SET
if name not in EMBEDDING_SET:
EMBEDDING_SET.add(name)
EMBEDDING_LIST.append(name)
import xdl.python.framework.variable as variable
with variable.variable_info(batch_read=batch_read, save_ratio=feature_add_probability, bloom_filter=cbf, save="true" if save else "false"):
var = variable.Variable(name=name,
dtype=DataType.float,
shape=[feature_dim, emb_dim],
initializer=initializer,
vtype=vtype,
trainable=True)
if statis_list is not None:
statis_vars = []
for statis_type in statis_list:
statis_var = variable.Variable(name=name,
dtype=DataType.float,
shape=[feature_dim, 1],
initializer=xdl.Zeros(),
vtype=vtype,
trainable=False,
statis_type=statis_type,
statis_decay=statis_decay,
statis_decay_period=statis_decay_period)
statis_vars.append(statis_var)
if sparse_input.has_unique_ids():
unique_ids = xdl.identity_op(sparse_input.ids)
idx = sparse_input.indices
embeddings = var.gather(unique_ids)
sidx = sparse_input.sidx
sseg = sparse_input.sseg
else:
with xdl.device(device, **device_attr):
unique_ids, idx, sidx, sseg = xdl.unique(sparse_input.ids, sparse_input.segments, itype=DataType.int32)
embeddings = var.gather(unique_ids)
if statis_list is not None:
assert labels is not None
from xdl.python.training.training_utils import get_global_step
global_step = get_global_step()
statis_results = []
for statis_var in statis_vars:
statis_result = statis_var.statis(sparse_input.ids, idx, sparse_input.segments, sidx, sseg, labels, global_step.value)
statis_results.append(statis_result)
global _EMBEDDING_TENSOR
_EMBEDDING_TENSOR[embeddings] = var
import xdl.python.sparse_engine.embedding_ops as embedding_ops
import numpy as np
if combiner == 'sum':
embeddings = embedding_ops.ksum(
embeddings,
idx,
sparse_input.values,
sparse_input.segments,
sidx,
sseg,
device, **device_attr)
elif combiner == 'mean':
embeddings = embedding_ops.kmean(
embeddings,
idx,
sparse_input.values,
sparse_input.segments,
sidx,
sseg,
device, **device_attr)
elif combiner == 'tile':
embeddings = embedding_ops.tile(
embeddings,
idx,
np.array([], dtype=np.float32),
#sparse_input.values,
sparse_input.segments,
length,
reverse,
device, **device_attr)
else:
raise Exception("Unrecognized combiner:" + str(combiner))
if sparse_input.shape is not None and len(sparse_input.shape) > 0:
embeddings.set_shape([sparse_input.shape[0], emb_dim]);
emb_info = EmbeddingInfo(name, feature_dim, emb_dim, combiner, None, var, length, embeddings)
set_embedding_info([var], emb_info)
if statis_list is not None:
return embeddings, statis_results
return embeddings
def embedding(name,
sparse_input,
initializer,
emb_dim,
feature_dim,
combiner='sum',
vtype=VarType.Index,
length=50,
reverse=False,
batch_read=3000,
feature_add_probability=1.0):
"""xdl embedding
Args:
name: name for embedding, will be used for declaring variable on ps-plus
sparse_input: a sparse tensor represent input data
initializer: intializer for the variable on ps-plus
emb_dim: embedding dimension
feature_dim: sparse input dimension, for pre-allocate memory
combiner: reduce operator, support sum|mean
Returns:
a tensor represent embedding result
Raises:
None
"""
import xdl.python.framework.variable as variable
with variable.variable_info(batch_read=batch_read):
var = variable.Variable(name=name,
dtype=DataType.float,
shape=[feature_dim, emb_dim],
initializer=initializer,
vtype=vtype,
trainable=True)
if sparse_input.has_unique_ids():
unique_ids = sparse_input.ids
idx = sparse_input.indices
embeddings = var.gather(unique_ids, save_ratio=feature_add_probability)
else:
unique_ids, idx = xdl.unique(sparse_input.ids, itype=DataType.int32)
embeddings = var.gather(unique_ids, save_ratio=feature_add_probability)
global _EMBEDDING_TENSOR
_EMBEDDING_TENSOR[embeddings] = var
import xdl.python.sparse_engine.embedding_ops as embedding_ops
if combiner == 'sum':
embeddings = embedding_ops.ksum(embeddings, idx, sparse_input.values,
sparse_input.segments)
elif combiner == 'mean':
embeddings = embedding_ops.kmean(embeddings, idx, sparse_input.values,
sparse_input.segments)
elif combiner == 'tile':
embeddings = embedding_ops.tile(embeddings, idx, sparse_input.values,
sparse_input.segments, length, reverse)
else:
raise Exception("Unrecognized combiner:" + str(combiner))
if sparse_input.shape is not None and len(sparse_input.shape) > 0:
embeddings.set_shape([sparse_input.shape[0], emb_dim])
emb_info = EmbeddingInfo(name, feature_dim, emb_dim, combiner, None, var,
embeddings)
set_embedding_info([var], emb_info)
return embeddings
