模型代码
https://github.com/NVIDIA/DeepLearningExamples.git
DeepLearningExamples/TensorFlow/Recommendation/WideAndDeep/
参考代码库说明下载数据集和进行训练得到ckpt
保存saved model
(在task.py修改)
# infer input shapes and types from feature_columns as a parse_example_spec
parse_example_spec = tf.feature_column.make_parse_example_spec(deep_columns + wide_columns)
print("parse_example_spec:", parse_example_spec)
# expose serialized Example protobuf string input and parse to feature tensors
# with a serving_input_receiver_fn
serving_input_receiver_fn = tf.estimator.export.build_parsing_serving_input_receiver_fn(parse_example_spec)
print("serving_input_receiver_fn:", serving_input_receiver_fn)
# export in saved_model format
estimator.export_saved_model(export_dir_base='./', serving_input_receiver_fn=serving_input_receiver_fn)
注意：feature columns里面tf.feature_column创建里面tf.int32改成tf.int64，否则保存模型时出现dtype错误。
查看saved model输入输出信息
模型保存在文件夹1633915266 （示例）
saved_model_cli show --dir 1633915266 --tag_set serve --signature_def serving_default
saved_model_cli show --dir saved_model --tag_set serve --all
The given SavedModel SignatureDef contains the following input(s):
inputs['inputs'] tensor_info:
dtype: DT_STRING
shape: (-1)
name: input_example_tensor:0
The given SavedModel SignatureDef contains the following output(s):
outputs['classes'] tensor_info:
dtype: DT_STRING
shape: (-1, 2)
name: head/Tile:0
outputs['scores'] tensor_info:
dtype: DT_FLOAT
shape: (-1, 2)
name: head/predictions/probabilities:0
Method name is: tensorflow/serving/classify
saved model 转frozen pb
from tensorflow.python.tools import freeze_graph
from tensorflow.python.saved_model import tag_constants
input_saved_model_dir = "./1633915266/"
output_node_names = "head/Tile,head/predictions/probabilities"
input_binary = False
input_saver_def_path = False
restore_op_name = None
filename_tensor_name = None
clear_devices = False
input_meta_graph = False
checkpoint_path = None
input_graph_filename = None
saved_model_tags = tag_constants.SERVING
output_graph_filename='frozen_graph.pb'
freeze_graph.freeze_graph(input_graph_filename,
input_saver_def_path,
input_binary,
checkpoint_path,
output_node_names,
restore_op_name,
filename_tensor_name,
output_graph_filename,
clear_devices,
"", "", "",
input_meta_graph,
input_saved_model_dir,
saved_model_tags)saved model推理
难点在于构造输入，这里给出基于estimator input_fn的简单构造example。
(在task.py修改)
parse_example_ids=[
'doc_event_id',
'doc_id',
'doc_event_source_id',
'event_geo_location',
'event_country_state',
'doc_event_publisher_id',
'event_country',
'event_hour',
'event_platform',
'traffic_source',
'event_weekend',
'user_has_already_viewed_doc',
'doc_event_entity_id',
'doc_event_topic_id',
'doc_event_category_id',
'pop_document_id_conf',
'pop_publisher_id_conf',
'pop_source_id_conf',
'pop_entity_id_conf',
'pop_topic_id_conf',
'pop_category_id_conf',
'pop_document_id_log_01scaled',
'pop_publisher_id_log_01scaled',
'pop_source_id_log_01scaled',
'pop_entity_id_log_01scaled',
'pop_topic_id_log_01scaled',
'pop_category_id_log_01scaled',
'user_views_log_01scaled',
'doc_views_log_01scaled',
'doc_event_days_since_published_log_01scaled',
'doc_event_hour_log_01scaled',
'ad_id',
'doc_ad_source_id',
'ad_advertiser',
'doc_ad_publisher_id',
'doc_ad_topic_id',
'doc_ad_entity_id',
'doc_ad_category_id',
'pop_ad_id_conf',
'user_doc_ad_sim_categories_conf',
'user_doc_ad_sim_topics_conf',
'pop_advertiser_id_conf',
'pop_campain_id_conf_multipl_log_01scaled',
'pop_ad_id_log_01scaled',
'pop_advertiser_id_log_01scaled',
'pop_campain_id_log_01scaled',
'user_doc_ad_sim_categories_log_01scaled',
'user_doc_ad_sim_topics_log_01scaled',
'user_doc_ad_sim_entities_log_01scaled',
'doc_event_doc_ad_sim_categories_log_01scaled',
'doc_event_doc_ad_sim_topics_log_01scaled',
'doc_event_doc_ad_sim_entities_log_01scaled',
'ad_views_log_01scaled',
'doc_ad_days_since_published_log_01scaled',
]
def _float_feature(tensor):
"""Returns a float_list from a float / double."""
return tf.train.Feature(float_list=tf.train.FloatList(value=tensor))
def _int64_feature(tensor):
"""Returns an int64_list from a bool / enum / int / uint."""
return tf.train.Feature(int64_list=tf.train.Int64List(value=tensor))
warmup_time = 100
measure_time = 100
eval_batch = 4
_dataset = eval_input_fn()
iterator = _dataset.make_one_shot_iterator()
one_element = iterator.get_next()
with tf.Session() as sess:
data = sess.run(one_element)
# print("data:", data)
exported_path = "./1633937564"
predictor = tf.contrib.predictor.from_saved_model(exported_path)
warmup_time = 100
measure_time = 100
eval_batch = 4
model_path = "./frozen_graph.pb"
input_names = ["input_example_tensor"]
output_names = ["head/Tile", "head/predictions/probabilities"]
graph = tf.Graph()
with graph.as_default():
with tf.gfile.FastGFile(model_path, "rb") as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
tf.import_graph_def(graph_def, name='')
inputs_tf = [graph.get_tensor_by_name(name + ":0") for name in input_names]
outputs_tf = [graph.get_tensor_by_name(name + ":0") for name in output_names]
config = tf.ConfigProto(allow_soft_placement=True)
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
data0 = data[0]
feature_list = [{} for i in range(eval_batch)]
for key in data0:
if key in parse_example_ids:
tensor = data0[key] # this is a batch of data
for i in range(eval_batch):
tensor0 = tensor[i]
dtype = str(tensor.dtype)
if dtype == "float32":
feature_list[i][key] = _float_feature(tensor0)
elif dtype == "int64":
feature_list[i][key] = _int64_feature(tensor0)
else:
raise ValueError("dtype not supported")
# print("feature_list:", feature_list)
examples = [tf.train.Example(features=tf.train.Features(feature=feature)) for feature in feature_list]
example_protos = [example.SerializeToString() for example in examples]
# print("examples:", examples)
# print("example_protos:", example_protos)
feed_dict = {"inputs": example_protos}
for i in range(warmup_time):
output_dict = predictor(feed_dict)
time1 = time.time()
for i in range(measure_time):
output_dict = predictor(feed_dict)
time2 = time.time()
print("avg time:", (time2 - time1) / measure_time)
print("output_dict:", output_dict)
parse_example_ids来自于上面打印的parse_example_spec
frozen pb推理
(在task.py修改)
parse_example_ids = [
'doc_event_id',
'doc_id',
'doc_event_source_id',
'event_geo_location',
'event_country_state',
'doc_event_publisher_id',
'event_country',
'event_hour',
'event_platform',
'traffic_source',
'event_weekend',
'user_has_already_viewed_doc',
'doc_event_entity_id',
'doc_event_topic_id',
'doc_event_category_id',
'pop_document_id_conf',
'pop_publisher_id_conf',
'pop_source_id_conf',
'pop_entity_id_conf',
'pop_topic_id_conf',
'pop_category_id_conf',
'pop_document_id_log_01scaled',
'pop_publisher_id_log_01scaled',
'pop_source_id_log_01scaled',
'pop_entity_id_log_01scaled',
'pop_topic_id_log_01scaled',
'pop_category_id_log_01scaled',
'user_views_log_01scaled',
'doc_views_log_01scaled',
'doc_event_days_since_published_log_01scaled',
'doc_event_hour_log_01scaled',
'ad_id',
'doc_ad_source_id',
'ad_advertiser',
'doc_ad_publisher_id',
'doc_ad_topic_id',
'doc_ad_entity_id',
'doc_ad_category_id',
'pop_ad_id_conf',
'user_doc_ad_sim_categories_conf',
'user_doc_ad_sim_topics_conf',
'pop_advertiser_id_conf',
'pop_campain_id_conf_multipl_log_01scaled',
'pop_ad_id_log_01scaled',
'pop_advertiser_id_log_01scaled',
'pop_campain_id_log_01scaled',
'user_doc_ad_sim_categories_log_01scaled',
'user_doc_ad_sim_topics_log_01scaled',
'user_doc_ad_sim_entities_log_01scaled',
'doc_event_doc_ad_sim_categories_log_01scaled',
'doc_event_doc_ad_sim_topics_log_01scaled',
'doc_event_doc_ad_sim_entities_log_01scaled',
'ad_views_log_01scaled',
'doc_ad_days_since_published_log_01scaled',
]
def _float_feature(tensor):
"""Returns a float_list from a float / double."""
return tf.train.Feature(float_list=tf.train.FloatList(value=tensor))
def _int64_feature(tensor):
"""Returns an int64_list from a bool / enum / int / uint."""
return tf.train.Feature(int64_list=tf.train.Int64List(value=tensor))
_dataset = eval_input_fn()
iterator = _dataset.make_one_shot_iterator()
one_element = iterator.get_next()
with tf.Session() as sess:
data = sess.run(one_element)
# print("data:", data)
warmup_time = 100
measure_time = 100
eval_batch = 4
model_path = "./frozen_graph.pb"
input_names = ["input_example_tensor"]
output_names = ["head/Tile", "head/predictions/probabilities"]
graph = tf.Graph()
with graph.as_default():
with tf.gfile.FastGFile(model_path, "rb") as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
tf.import_graph_def(graph_def, name='')
inputs_tf = [graph.get_tensor_by_name(name + ":0") for name in input_names]
outputs_tf = [graph.get_tensor_by_name(name + ":0") for name in output_names]
config = tf.ConfigProto(allow_soft_placement=True)
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
data0 = data[0]
feature_list = [{} for i in range(eval_batch)]
for key in data0:
if key in parse_example_ids:
tensor = data0[key] # this is a batch of data
for i in range(eval_batch):
tensor0 = tensor[i]
dtype = str(tensor.dtype)
if dtype == "float32":
feature_list[i][key] = _float_feature(tensor0)
elif dtype == "int64":
feature_list[i][key] = _int64_feature(tensor0)
else:
raise ValueError("dtype not supported")
# print("feature_list:", feature_list)
examples = [tf.train.Example(features=tf.train.Features(feature=feature)) for feature in feature_list]
example_protos = [example.SerializeToString() for example in examples]
# print("examples:", examples)
# print("example_protos:", example_protos)
feed_dict = {inputs_tf[0]: example_protos}
for i in range(warmup_time):
output_dict = sess.run(outputs_tf, feed_dict=feed_dict)
time1 = time.time()
for i in range(measure_time):
output_dict = sess.run(outputs_tf, feed_dict=feed_dict)
time2 = time.time()
print("avg time:", (time2 - time1) / measure_time)
print("output_dict:", output_dict)
frozen pb tf-trt优化
import tensorflow as tf
from tensorflow.python.compiler.tensorrt import trt_convert as trt
def write_graph_def(graph_def, dump_graph_name):
if dump_graph_name.endswith("pb"):
with tf.io.gfile.GFile(dump_graph_name, "wb") as f:
f.write(graph_def.SerializeToString())
elif dump_graph_name.endswith("pbtxt"):
tf.io.write_graph(graph_def, os.path.dirname(dump_graph_name), os.path.basename(dump_graph_name))
else:
raise ValueError("dump_graph_name {} is invalid".format(dump_graph_name))
model_path = "./frozen_graph.pb"
with tf.gfile.GFile(model_path, "rb") as f:
frozen_graph = tf.GraphDef()
frozen_graph.ParseFromString(f.read())
output_names = [
"head/Tile",
"head/predictions/probabilities",
]
converter = trt.TrtGraphConverter(
input_graph_def=frozen_graph,
nodes_blacklist=output_names,
precision_mode="FP32", # fp16 mode
is_dynamic_op=True)
trt_graph = converter.convert()
write_graph_def(trt_graph, "trt_graph.pb")
参考
https://tensorflow.google.cn/tutorials/load_data/tfrecord
https://www.tensorflow.org/api_docs/python/tf/io/parse_example
https://www.tensorflow.org/api_docs/cc/class/tensorflow/ops/parse-example
tensorflow中tfrecords格式数据读写小记 tensorflow中tfrecords格式数据读写小记 - 知乎
tensorflow 模型导出总结 tensorflow 模型导出总结 - 知乎
一套灵活的Tensorflow Train/Serving方案 一套灵活的Tensorflow Train/Serving方案 - 知乎
https://docs.nvidia.com/deeplearning/frameworks/tf-trt-user-guide/index.html