Skip to content

Mace
Mace

孤小逸 / Mace
与 Fork 源项目一致

Fork自 Xiaomi / Mace

通知 1
Star 1
Fork 0
Mace
Mace

前往新版Gitcode，体验更适合开发者的 AI 搜索 >>

提交 8af28527 编写于 作者: L liuqi
浏览文件
操作

Update docs for latest usage.

上级 ced4a49d
隐藏空白更改
内联 并排
Showing with 263 addition and 216 deletion
docs/getting_started/create_a_model_deployment.rst
浏览文件 @ 8af28527
......@@ -30,22 +30,31 @@ Configurations
:header-rows: 1
:align: left
* - Configuration key
- Description
* - library_name
- library name
* - target_abis
- The target ABI to build, can be one or more of 'host', 'armeabi-v7a' or 'arm64-v8a'
* - target_socs
- build for specified socs if you just want use the model for that socs.
* - embed_model_data
- Whether embedding model weights as the code, default to 1
- Whether embedding model weights as the code, default to 0
* - build_type
- model build type, can be ['proto', 'code']. 'proto' for converting model to ProtoBuf file and 'code' for converting model to c++ code.
* - model_name
- model name. should be unique if there are multiple models.
**LIMIT: if build_type is code, model_name will used in c++ code so that model_name must fulfill c++ name specification.**
* - platform
- The source framework, tensorflow or caffe
- The source framework, one of [tensorflow, caffe]
* - model_file_path
- The path of the model file, can be local or remote
* - weight_file_path
- The path of the model weights file, used by Caffe model
* - model_sha256_checksum
- The SHA256 checksum of the model file
* - weight_file_path
- The path of the model weights file, used by Caffe model
* - weight_sha256_checksum
- The SHA256 checksum of the weight file, used by Caffe model
* - subgraphs
- subgraphs key. ** DO NOT EDIT **
* - input_tensors
- The input tensor names (tensorflow), top name of inputs' layer (caffe). one or more strings
* - output_tensors
......@@ -55,14 +64,16 @@ Configurations
* - output_shapes
- The shapes of the output tensors, in NHWC order
* - runtime
- The running device, one of CPU, GPU or DSP
- The running device, one of [cpu, gpu, dsp, cpu_gpu]. cpu_gpu contains cpu and gpu model definition so you can run the model on both cpu and gpu.
* - data_type
- [optional] The data type used for specified runtime. [fp16_fp32, fp32_fp32] for gpu, default is fp16_fp32. [fp32] for cpu. [uint8] for dsp.
* - limit_opencl_kernel_time
- Whether splitting the OpenCL kernel within 1 ms to keep UI responsiveness, default to 0
* - dsp_mode
- Control the DSP precision and performance, default to 0 usually works for most cases
- [optional] Whether splitting the OpenCL kernel within 1 ms to keep UI responsiveness, default to 0
* - nnlib_graph_mode
- [optional] Control the DSP precision and performance, default to 0 usually works for most cases
* - obfuscate
- Whether to obfuscate the model operator name, default to 0
* - fast_conv
- Whether to enable Winograd convolution, **will increase memory consumption**
- [optional] Whether to obfuscate the model operator name, default to 0
* - winograd
- [optional] Whether to enable Winograd convolution, **will increase memory consumption**
* - input_files
- Specify Numpy validation inputs. When not provided, [-1, 1] random values will be used
- [optional] Specify Numpy validation inputs. When not provided, [-1, 1] random values will be used
docs/getting_started/how_to_build.rst
浏览文件 @ 8af28527
......@@ -166,11 +166,11 @@ Caffe目前只支持最新版本，旧版本请使用Caffe的工具进行升级
模型静态库的生成需要使用目标机型，\ ***并且要求必须在目标SOC的机型上编译生成静态库。***
我们提供了\ ``mace_tools.py``\ 工具，可以将模型文件转换成静态库。\ ``tools/mace_tools.py``\ 使用步骤：
我们提供了\ ``converter.py``\ 工具，可以将模型文件转换成静态库。\ ``tools/converter.py``\ 使用步骤：
3.2 运行\ ``tools/mace_tools.py``\ 脚本
3.2 运行\ ``tools/converter.py``\ 脚本
**Commands**
......@@ -181,8 +181,8 @@ Caffe目前只支持最新版本，旧版本请使用Caffe的工具进行升级
build模型静态库以及测试工具。
* *--config* (type=str, default="", required)：模型配置yaml文件路径.
* *--tuning* (optional)：是否为特定SOC调制GPU参数.
* *--enable_openmp* (optional)：是否启用openmp.
* *--tuning* (default=false, optional)：是否为特定SOC调制GPU参数.
* *--enable_openmp* (default=true, optional)：是否启用openmp.
**run**
......@@ -192,10 +192,10 @@ Caffe目前只支持最新版本，旧版本请使用Caffe的工具进行升级
* *--config* (type=str, default="", required)：模型配置yaml文件路径.
* *--round* (type=int, default=1, optional)：模型运行次数。
* *--validate* (optional): 是否需要验证运行结果与框架运行结果是否一致。
* *--validate* (default=false, optional): 是否需要验证运行结果与框架运行结果是否一致。
* *--caffe_env* (type=local/docker, default=docker, optional)：当vaildate时，可以选择指定caffe环境,local表示本地，docker表示使用docker容器.
* *--restart_round* (type=int, default=1, optional)：模型重启次数。
* *--check_gpu_out_of_memory* (optional): 是否需要检查gpu内存越界。
* *--check_gpu_out_of_memory* (default=false, optional): 是否需要检查gpu内存越界。
* *--vlog_level* (type=int[0-5], default=0, optional)：详细日志级别.
.. warning::
......@@ -256,26 +256,26 @@ Caffe目前只支持最新版本，旧版本请使用Caffe的工具进行升级
.. code:: sh
# print help message
python tools/mace_tools.py -h
python tools/mace_tools.py build -h
python tools/mace_tools.py run -h
python tools/mace_tools.py benchmark -h
python tools/converter.py -h
python tools/converter.py build -h
python tools/converter.py run -h
python tools/converter.py benchmark -h
# 仅编译模型和生成静态库
python tools/mace_tools.py build --config=models/config.yaml
python tools/converter.py build --config=models/config.yaml
# 测试模型的运行时间
python tools/mace_tools.py run --config=models/config.yaml --round=100
python tools/converter.py run --config=models/config.yaml --round=100
# 对比编译好的模型在mace上与直接使用tensorflow或者caffe运行的结果，相似度使用`余弦距离表示`
# 其中使用OpenCL设备，默认相似度大于等于`0.995`为通过；DSP设备下，相似度需要达到`0.930`。
python tools/mace_tools.py run --config=models/config.yaml --validate
python tools/converter.py run --config=models/config.yaml --validate
# 模型Benchmark：查看每个Op的运行时间
python tools/mace_tools.py benchmark --config=models/config.yaml
python tools/converter.py benchmark --config=models/config.yaml
# 查看模型运行时占用内存（如果有多个模型，可能需要注释掉一部分配置，只剩一个模型的配置）
python tools/mace_tools.py run --config=models/config.yaml --round=10000 &
python tools/converter.py run --config=models/config.yaml --round=10000 &
adb shell dumpsys meminfo | grep mace_run
sleep 10
kill %1
......
docs/getting_started/models/demo_app_models.yaml
浏览文件 @ 8af28527
......@@ -7,7 +7,7 @@ target_socs: [msm8998]
embed_model_data: 1
build_type: code # 模型build类型。code表示将模型转为代码，proto表示将模型转为protobuf文件
models: # 一个配置文件可以包含多个模型的配置信息，最终生成的库中包含多个模型
first_net: # 模型的标签，在调度模型的时候，会用这个变量，必须唯一
model_name: # 模型的标签，在调度模型的时候，会用这个变量，必须唯一
platform: tensorflow
model_file_path: path/to/model64.pb # also support http:// and https://
model_sha256_checksum: 7f7462333406e7dea87222737590ebb7d94490194d2f21a7d72bafa87e64e9f9
......
mace/benchmark/statistics.cc
浏览文件 @ 8af28527
......@@ -238,7 +238,7 @@ std::string OpStat::StatByNodeType() const {
std::string OpStat::Summary() const {
std::stringstream stream;
if (!records_.empty()) {
stream << total_time_.ToString("Summary") << std::endl;
stream << total_time_.ToString("Summary of Ops' Stat") << std::endl;
}
stream << records_.size() << " ops total." << std::endl;
......
tools/bazel_adb_run.py
浏览文件 @ 8af28527
......@@ -136,11 +136,6 @@ def main(unused_args):
sh_commands.gen_mace_version()
sh_commands.gen_tuning_param_code([])
strip = "always"
debug = False
if FLAGS.address_sanitizer:
strip = "never"
debug = True
for target_abi in target_abis:
sh_commands.bazel_build(target, abi=target_abi,
enable_neon=FLAGS.enable_neon,
......@@ -158,10 +153,10 @@ def main(unused_args):
host_bin_path,
bin_name,
args=FLAGS.args,
opencl_profiling=1,
opencl_profiling=True,
vlog_level=0,
device_bin_path="/data/local/tmp/mace",
out_of_range_check=1,
out_of_range_check=True,
address_sanitizer=FLAGS.address_sanitizer)
device_properties = sh_commands.adb_getprop_by_serialno(
serialno)
......
tools/mace_tools.py → tools/converter.py
浏览文件 @ 8af28527
......@@ -12,21 +12,17 @@
# See the License for the specific language governing permissions and
# limitations under the License.
# python tools/mace_tools.py \
# --config=tools/example.yaml \
# --round=100 \
# --mode=all
import argparse
import filelock
import hashlib
import os
import re
import sh
import subprocess
import sys
import urllib
import yaml
import re
from enum import Enum
import sh_commands
......@@ -46,13 +42,23 @@ MODEL_OUTPUT_DIR_NAME = 'model'
BUILD_TMP_DIR_NAME = '_tmp'
BUILD_TMP_GENERAL_OUTPUT_DIR_NAME = 'general'
OUTPUT_LIBRARY_DIR_NAME = 'library'
CL_BUILT_KERNEL_FILE_NAME = "mace_cl_compiled_program.bin"
CL_PLATFORM_INFO_FILE_NAME = "mace_cl_platform_info.txt"
CODEGEN_BASE_DIR = 'mace/codegen'
MODEL_CODEGEN_DIR = CODEGEN_BASE_DIR + '/models'
MACE_RUN_TARGET = "//mace/tools/validation:mace_run"
ABITypeStrs = [
"armeabi-v7a",
"arm64-v8a",
"host",
'armeabi-v7a',
'arm64-v8a',
'host',
]
ABIType = Enum('ABIType', [(ele, ele) for ele in ABITypeStrs], type=str)
class ABIType(object):
armeabi_v7a = 'armeabi-v7a'
arm64_v8a = 'arm64-v8a'
host = 'host'
PlatformTypeStrs = [
......@@ -92,6 +98,13 @@ GPUDataTypeStrs = [
GPUDataType = Enum('GPUDataType', [(ele, ele) for ele in GPUDataTypeStrs],
type=str)
DSPDataTypeStrs = [
"uint8",
]
DSPDataType = Enum('DSPDataType', [(ele, ele) for ele in DSPDataTypeStrs],
type=str)
class DefaultValues(object):
omp_num_threads = -1,
......@@ -129,6 +142,8 @@ class YAMLKeyword(object):
class ModuleName(object):
YAML_CONFIG = 'YAML CONFIG'
MODEL_CONVERTER = 'Model Converter'
RUN = 'RUN'
BENCHMARK = 'Benchmark'
CPP_KEYWORDS = [
......@@ -171,11 +186,13 @@ def parse_device_type(runtime):
def get_hexagon_mode(configs):
runtime_list = []
for model_name in configs["models"]:
model_runtime = configs["models"][model_name].get("runtime", "")
for model_name in configs[YAMLKeyword.models]:
model_runtime =\
configs[YAMLKeyword.models][model_name].get(
YAMLKeyword.runtime, "")
runtime_list.append(model_runtime.lower())
if "dsp" in runtime_list:
if RuntimeType.dsp in runtime_list:
return True
return False
......@@ -200,7 +217,7 @@ def format_model_config(config_file_path):
library_name = configs.get(YAMLKeyword.library_name, "")
mace_check(len(library_name) > 0,
ModuleName.YAML_CONFIG, "library name shuold not be empty")
ModuleName.YAML_CONFIG, "library name should not be empty")
target_abis = configs.get(YAMLKeyword.target_abis, [])
mace_check((isinstance(target_abis, list) and len(target_abis) > 0),
......@@ -216,7 +233,8 @@ def format_model_config(config_file_path):
elif not isinstance(target_socs, list):
configs[YAMLKeyword.target_socs] = [target_socs]
if ABIType.host not in target_abis:
if ABIType.armeabi_v7a in target_abis \
or ABIType.arm64_v8a in target_abis:
available_socs = sh_commands.adb_get_all_socs()
if YAMLKeyword.target_socs in configs:
target_socs = set(configs[YAMLKeyword.target_socs])
......@@ -261,7 +279,7 @@ def format_model_config(config_file_path):
mace_check((model_name[0] == '_' or model_name[0].isalpha())
and bool(model_name_reg.match(model_name)),
ModuleName.YAML_CONFIG,
"model name shuold Meet the c++ naming convention"
"model name should Meet the c++ naming convention"
" which start with '_' or alpha"
" and only contain alpha, number and '_'")
......@@ -318,6 +336,15 @@ def format_model_config(config_file_path):
else:
model_config[YAMLKeyword.data_type] =\
GPUDataType.fp16_fp32.value
elif runtime == RuntimeType.dsp:
if len(data_type) > 0:
mace_check(data_type in DSPDataTypeStrs,
ModuleName.YAML_CONFIG,
"'data_type' must be in " + str(DSPDataTypeStrs)
+ " for dsp runtime")
else:
model_config[YAMLKeyword.data_type] = \
DSPDataType.uint8.value
subgraphs = model_config.get(YAMLKeyword.subgraphs, "")
mace_check(len(subgraphs) > 0, ModuleName.YAML_CONFIG,
......@@ -342,15 +369,17 @@ def format_model_config(config_file_path):
if value == "":
model_config[key] = 0
validation_inputs_data = model_config.get("validation_inputs_data",
[])
model_config["validation_inputs_data"] = validation_inputs_data
validation_inputs_data = model_config.get(
YAMLKeyword.validation_inputs_data, [])
if not isinstance(validation_inputs_data, list):
model_config["validation_inputs_data"] = [
model_config[YAMLKeyword.validation_inputs_data] = [
validation_inputs_data]
else:
model_config[YAMLKeyword.validation_inputs_data] = \
validation_inputs_data
weight_file_path = model_config.get("weight_file_path", "")
model_config["weight_file_path"] = weight_file_path
weight_file_path = model_config.get(YAMLKeyword.weight_file_path, "")
model_config[YAMLKeyword.weight_file_path] = weight_file_path
return configs
......@@ -359,9 +388,9 @@ def get_build_binary_dir(library_name, target_abi, target_soc,
serial_num):
if not target_soc or not serial_num:
binary_path_digest = md5sum(target_abi)
binary_path_digest = "%s_%s" % (target_abi, binary_path_digest)
else:
device_name = sh_commands.adb_get_device_name_by_serialno(serial_num)\
.replace(' ', '')
device_name = sh_commands.adb_get_device_name_by_serialno(serial_num)
binary_path_digest = md5sum(target_abi + target_soc + serial_num)
binary_path_digest = "%s_%s_%s" % \
(device_name, target_soc, binary_path_digest)
......@@ -386,7 +415,7 @@ def get_build_model_dirs(library_name, model_name, target_abi, target_soc,
device_name = \
sh_commands.adb_get_device_name_by_serialno(serial_num)
model_output_dir = "%s/%s_%s/%s" % (
model_output_base_dir, device_name.replace(' ', ''),
model_output_base_dir, device_name,
target_soc, target_abi)
mace_model_dir = \
......@@ -401,21 +430,16 @@ def get_build_model_dirs(library_name, model_name, target_abi, target_soc,
def pull_opencl_binary_and_tuning_param(target_abi,
serialno,
model_output_dirs):
cl_built_kernel_file_name = "mace_cl_compiled_program.bin"
cl_platform_info_file_name = "mace_cl_platform_info.txt"
sh_commands.pull_binaries(target_abi, serialno, model_output_dirs,
cl_built_kernel_file_name,
cl_platform_info_file_name)
CL_BUILT_KERNEL_FILE_NAME,
CL_PLATFORM_INFO_FILE_NAME)
def gen_opencl_and_tuning_code(model_output_dirs):
cl_built_kernel_file_name = "mace_cl_compiled_program.bin"
cl_platform_info_file_name = "mace_cl_platform_info.txt"
# generate opencl binary code
sh_commands.gen_opencl_binary_code(model_output_dirs,
cl_built_kernel_file_name,
cl_platform_info_file_name)
CL_BUILT_KERNEL_FILE_NAME,
CL_PLATFORM_INFO_FILE_NAME)
sh_commands.gen_tuning_param_code(model_output_dirs)
......@@ -475,8 +499,9 @@ def convert_model(configs):
library_name = configs[YAMLKeyword.library_name]
if not os.path.exists(BUILD_OUTPUT_DIR):
os.makedirs(BUILD_OUTPUT_DIR)
elif not os.path.exists(os.path.join(BUILD_OUTPUT_DIR, library_name)):
os.makedirs(os.path.join(BUILD_OUTPUT_DIR, library_name))
elif os.path.exists(os.path.join(BUILD_OUTPUT_DIR, library_name)):
sh.rm("-rf", os.path.join(BUILD_OUTPUT_DIR, library_name))
os.makedirs(os.path.join(BUILD_OUTPUT_DIR, library_name))
model_output_dir = \
'%s/%s/%s' % (BUILD_OUTPUT_DIR, library_name, MODEL_OUTPUT_DIR_NAME)
......@@ -527,12 +552,10 @@ def convert_model(configs):
"weight file sha256checksum not match")
data_type = model_config[YAMLKeyword.data_type]
if ABIType.host.value in configs[YAMLKeyword.target_abis]:
data_type = CPUDataType.fp32.value
# TODO(liuqi): support multiple subgraphs
subgraphs = model_config[YAMLKeyword.subgraphs]
model_codegen_dir = "mace/codegen/models/%s" % model_name
model_codegen_dir = "%s/%s" % (MODEL_CODEGEN_DIR, model_name)
sh_commands.gen_model_code(
model_codegen_dir,
model_config[YAMLKeyword.platform],
......@@ -561,14 +584,13 @@ def convert_model(configs):
output_dir=model_output_dir
)
MaceLogger.header(
MaceLogger.summary(
StringFormatter.block("Model %s converted" % model_name))
def build_specific_lib(target_abi, target_soc, serial_num,
configs, tuning, enable_openmp,
address_sanitizer):
mace_run_target = "//mace/tools/validation:mace_run"
library_name = configs[YAMLKeyword.library_name]
build_type = configs[YAMLKeyword.build_type]
embed_model_data = configs[YAMLKeyword.embed_model_data]
......@@ -583,7 +605,7 @@ def build_specific_lib(target_abi, target_soc, serial_num,
gen_opencl_and_tuning_code([])
sh_commands.bazel_build(
mace_run_target,
MACE_RUN_TARGET,
abi=target_abi,
hexagon_mode=hexagon_mode,
enable_openmp=enable_openmp,
......@@ -608,7 +630,7 @@ def build_specific_lib(target_abi, target_soc, serial_num,
os.makedirs(model_output_dir)
# build for specified soc
if not address_sanitizer and target_abi != ABIType.host \
if not address_sanitizer and tuning and target_abi != ABIType.host \
and target_soc is not None and \
model_runtime in [RuntimeType.gpu, RuntimeType.cpu_gpu]:
sh_commands.clear_phone_data_dir(serial_num, PHONE_DATA_DIR)
......@@ -653,7 +675,7 @@ def build_specific_lib(target_abi, target_soc, serial_num,
if binary_changed:
gen_opencl_and_tuning_code(model_output_dirs)
sh_commands.bazel_build(
mace_run_target,
MACE_RUN_TARGET,
abi=target_abi,
hexagon_mode=hexagon_mode,
enable_openmp=enable_openmp,
......@@ -671,6 +693,7 @@ def build_specific_lib(target_abi, target_soc, serial_num,
# generate library
sh_commands.merge_libs(target_soc,
serial_num,
target_abi,
library_name,
BUILD_OUTPUT_DIR,
......@@ -705,17 +728,18 @@ def generate_library(configs, tuning, enable_openmp, address_sanitizer):
target_socs = configs[YAMLKeyword.target_socs]
for target_abi in configs[YAMLKeyword.target_abis]:
if not target_socs or target_abi == ABIType.host.value:
if not target_socs or target_abi == ABIType.host:
build_specific_lib(target_abi, None, None, configs,
tuning, enable_openmp, address_sanitizer)
else:
for target_soc in target_socs:
serial_num = sh_commands.get_target_soc_serial_number(
target_soc)
with sh_commands.device_lock(serial_num):
build_specific_lib(target_abi, target_soc, serial_num,
configs, tuning, enable_openmp,
address_sanitizer)
serial_nums = \
sh_commands.get_target_socs_serialnos([target_soc])
for serial_num in serial_nums:
with sh_commands.device_lock(serial_num):
build_specific_lib(target_abi, target_soc, serial_num,
configs, tuning, enable_openmp,
address_sanitizer)
# package library
sh_commands.packaging_lib(BUILD_OUTPUT_DIR,
......@@ -804,8 +828,19 @@ def run_specific_target(flags, configs, target_abi,
else:
build_tmp_binary_dir = get_build_binary_dir(library_name, target_abi,
target_soc, serial_num)
mace_check(os.path.exists(build_tmp_binary_dir),
ModuleName.RUN,
'You should build before run.')
for model_name in configs[YAMLKeyword.models]:
if target_abi == ABIType.host:
device_name = ABIType.host
else:
device_name =\
sh_commands.adb_get_device_name_by_serialno(serial_num)
MaceLogger.header(
StringFormatter.block(
"Run model %s on %s" % (model_name, device_name)))
model_config = configs[YAMLKeyword.models][model_name]
model_runtime = model_config[YAMLKeyword.runtime]
subgraphs = model_config[YAMLKeyword.subgraphs]
......@@ -820,6 +855,10 @@ def run_specific_target(flags, configs, target_abi,
get_build_model_dirs(library_name, model_name, target_abi,
target_soc, serial_num,
model_config[YAMLKeyword.model_file_path])
mace_check(os.path.exists(model_output_dir)
and os.path.exists(mace_model_dir),
ModuleName.RUN,
'You should build before run.')
if target_abi != ABIType.host:
sh_commands.clear_phone_data_dir(serial_num, PHONE_DATA_DIR)
......@@ -869,9 +908,9 @@ def run_specific_target(flags, configs, target_abi,
)
if flags.validate:
model_file_path, weight_file_path = get_model_files_path(
model_config["model_file_path"],
model_config[YAMLKeyword.model_file_path],
model_output_base_dir,
model_config["weight_file_path"])
model_config[YAMLKeyword.weight_file_path])
sh_commands.validate_model(
abi=target_abi,
......@@ -899,20 +938,21 @@ def run_mace(flags):
target_socs = configs[YAMLKeyword.target_socs]
if not target_socs:
target_socs = sh_commands.adb_get_all_socs()
if ABIType.host not in configs[YAMLKeyword.target_abis] \
and not target_socs:
MaceLogger.warning('There is no device plugin the computer.')
for target_abi in configs[YAMLKeyword.target_abis]:
if target_abi == ABIType.host:
run_specific_target(flags, configs, target_abi, None, None)
else:
for target_soc in target_socs:
serial_num = sh_commands.get_target_soc_serial_number(
target_soc)
with sh_commands.device_lock(serial_num):
run_specific_target(flags, configs, target_abi,
target_soc, serial_num)
serial_nums = \
sh_commands.get_target_socs_serialnos([target_soc])
mace_check(serial_nums,
ModuleName.RUN,
'There is no device with soc: ' + target_soc)
for serial_num in serial_nums:
with sh_commands.device_lock(serial_num):
run_specific_target(flags, configs, target_abi,
target_soc, serial_num)
################################
......@@ -928,8 +968,19 @@ def bm_specific_target(flags, configs, target_abi, target_soc, serial_num):
else:
build_tmp_binary_dir = get_build_binary_dir(library_name, target_abi,
target_soc, serial_num)
mace_check(os.path.exists(build_tmp_binary_dir),
ModuleName.BENCHMARK,
'You should build before benchmark.')
for model_name in configs[YAMLKeyword.models]:
if target_abi == ABIType.host:
device_name = ABIType.host
else:
device_name = \
sh_commands.adb_get_device_name_by_serialno(serial_num)
MaceLogger.header(
StringFormatter.block(
"Benchmark model %s on %s" % (model_name, device_name)))
model_config = configs[YAMLKeyword.models][model_name]
model_runtime = model_config[YAMLKeyword.runtime]
subgraphs = model_config[YAMLKeyword.subgraphs]
......@@ -944,6 +995,10 @@ def bm_specific_target(flags, configs, target_abi, target_soc, serial_num):
get_build_model_dirs(library_name, model_name, target_abi,
target_soc, serial_num,
model_config[YAMLKeyword.model_file_path])
mace_check(os.path.exists(model_output_dir)
and os.path.exists(mace_model_dir),
ModuleName.BENCHMARK,
'You should build before benchmark.')
if target_abi != ABIType.host:
sh_commands.clear_phone_data_dir(serial_num, PHONE_DATA_DIR)
......@@ -990,20 +1045,21 @@ def benchmark_model(flags):
target_socs = configs[YAMLKeyword.target_socs]
if not target_socs:
target_socs = sh_commands.adb_get_all_socs()
if ABIType.host.value not in configs[YAMLKeyword.target_abis] \
and not target_socs:
MaceLogger.warning('There is no device plugin the computer.')
for target_abi in configs[YAMLKeyword.target_abis]:
if target_abi == ABIType.host.value:
if target_abi == ABIType.host:
bm_specific_target(flags, configs, target_abi, None, None)
else:
for target_soc in target_socs:
serial_num = sh_commands.get_target_soc_serial_number(
target_soc)
with sh_commands.device_lock(serial_num):
bm_specific_target(flags, configs, target_abi,
target_soc, serial_num)
serial_nums = \
sh_commands.get_target_socs_serialnos([target_soc])
mace_check(serial_nums,
ModuleName.BENCHMARK,
'There is no device with soc: ' + target_soc)
for serial_num in serial_nums:
with sh_commands.device_lock(serial_num):
bm_specific_target(flags, configs, target_abi,
target_soc, serial_num)
################################
......
tools/sh_commands.py
浏览文件 @ 8af28527
......@@ -165,7 +165,7 @@ def adb_getprop_by_serialno(serialno):
def adb_get_device_name_by_serialno(serialno):
props = adb_getprop_by_serialno(serialno)
return props.get("ro.product.model", "")
return props.get("ro.product.model", "").replace(' ', '')
def adb_supported_abis(serialno):
......@@ -201,10 +201,10 @@ def adb_run(abi,
host_bin_path,
bin_name,
args="",
opencl_profiling=1,
opencl_profiling=True,
vlog_level=0,
device_bin_path="/data/local/tmp/mace",
out_of_range_check=1,
out_of_range_check=True,
address_sanitizer=False):
host_bin_full_path = "%s/%s" % (host_bin_path, bin_name)
device_bin_full_path = "%s/%s" % (device_bin_path, bin_name)
......@@ -225,11 +225,13 @@ def adb_run(abi,
adb_push(find_asan_rt_library(abi), device_bin_path, serialno)
ld_preload = "LD_PRELOAD=%s/%s" % (device_bin_path,
asan_rt_library_names(abi)),
opencl_profiling = 1 if opencl_profiling else 0
out_of_range_check = 1 if out_of_range_check else 0
print("Run %s" % device_bin_full_path)
stdout_buff = []
process_output = make_output_processor(stdout_buff)
p = sh.adb(
sh.adb(
"-s",
serialno,
"shell",
......@@ -239,9 +241,7 @@ def adb_run(abi,
"MACE_CPP_MIN_VLOG_LEVEL=%d" % vlog_level,
"%s %s" % (device_bin_full_path, args),
_out=process_output,
_bg=True,
_err_to_out=True)
p.wait()
_fg=True)
return "".join(stdout_buff)
......@@ -309,12 +309,10 @@ def bazel_build(target,
bazel_args += ("--config", "asan")
else:
bazel_args += ("--config", "optimization")
p = sh.bazel(
sh.bazel(
_out=process_output,
_bg=True,
_err_to_out=True,
_fg=True,
*bazel_args)
p.wait()
print("Build done!\n")
return "".join(stdout_buff)
......@@ -322,13 +320,11 @@ def bazel_build(target,
def bazel_build_common(target, build_args=""):
stdout_buff = []
process_output = make_output_processor(stdout_buff)
p = sh.bazel(
sh.bazel(
"build",
target + build_args,
_out=process_output,
_bg=True,
_err_to_out=True)
p.wait()
_fg=True)
return "".join(stdout_buff)
......@@ -462,30 +458,28 @@ def gen_model_code(model_codegen_dir,
stdout_buff = []
process_output = make_output_processor(stdout_buff)
p = sh.python("bazel-bin/mace/python/tools/converter",
"-u",
"--platform=%s" % platform,
"--model_file=%s" % model_file_path,
"--weight_file=%s" % weight_file_path,
"--model_checksum=%s" % model_sha256_checksum,
"--weight_checksum=%s" % weight_sha256_checksum,
"--input_node=%s" % input_nodes,
"--output_node=%s" % output_nodes,
"--runtime=%s" % runtime,
"--template=%s" % "mace/python/tools",
"--model_tag=%s" % model_tag,
"--input_shape=%s" % input_shapes,
"--dsp_mode=%s" % dsp_mode,
"--embed_model_data=%s" % embed_model_data,
"--winograd=%s" % fast_conv,
"--obfuscate=%s" % obfuscate,
"--output_dir=%s" % model_codegen_dir,
"--model_build_type=%s" % model_build_type,
"--data_type=%s" % data_type,
_out=process_output,
_bg=True,
_err_to_out=True)
p.wait()
sh.python("bazel-bin/mace/python/tools/converter",
"-u",
"--platform=%s" % platform,
"--model_file=%s" % model_file_path,
"--weight_file=%s" % weight_file_path,
"--model_checksum=%s" % model_sha256_checksum,
"--weight_checksum=%s" % weight_sha256_checksum,
"--input_node=%s" % input_nodes,
"--output_node=%s" % output_nodes,
"--runtime=%s" % runtime,
"--template=%s" % "mace/python/tools",
"--model_tag=%s" % model_tag,
"--input_shape=%s" % input_shapes,
"--dsp_mode=%s" % dsp_mode,
"--embed_model_data=%s" % embed_model_data,
"--winograd=%s" % fast_conv,
"--obfuscate=%s" % obfuscate,
"--output_dir=%s" % model_codegen_dir,
"--model_build_type=%s" % model_build_type,
"--data_type=%s" % data_type,
_out=process_output,
_fg=True)
def gen_random_input(model_output_dir,
......@@ -692,15 +686,13 @@ def tuning_run(abi,
"--model_file=%s" % mace_model_phone_path,
])
adb_cmd = ' '.join(adb_cmd)
p = sh.adb(
sh.adb(
"-s",
serialno,
"shell",
adb_cmd,
_out=process_output,
_bg=True,
_err_to_out=True)
p.wait()
_fg=True)
print("Running finished!\n")
return "".join(stdout_buff)
......@@ -802,26 +794,24 @@ def validate_model(abi,
stdout_buff = []
process_output = make_output_processor(stdout_buff)
p = sh.docker(
"exec",
container_name,
"python",
"-u",
"/mace/validate.py",
"--platform=caffe",
"--model_file=/mace/%s" % model_file_name,
"--weight_file=/mace/%s" % weight_file_name,
"--input_file=/mace/%s" % input_file_name,
"--mace_out_file=/mace/%s" % output_file_name,
"--device_type=%s" % device_type,
"--input_node=%s" % ",".join(input_nodes),
"--output_node=%s" % ",".join(output_nodes),
"--input_shape=%s" % ":".join(input_shapes),
"--output_shape=%s" % ":".join(output_shapes),
_out=process_output,
_bg=True,
_err_to_out=True)
p.wait()
sh.docker(
"exec",
container_name,
"python",
"-u",
"/mace/validate.py",
"--platform=caffe",
"--model_file=/mace/%s" % model_file_name,
"--weight_file=/mace/%s" % weight_file_name,
"--input_file=/mace/%s" % input_file_name,
"--mace_out_file=/mace/%s" % output_file_name,
"--device_type=%s" % device_type,
"--input_node=%s" % ",".join(input_nodes),
"--output_node=%s" % ",".join(output_nodes),
"--input_shape=%s" % ":".join(input_shapes),
"--output_shape=%s" % ":".join(output_shapes),
_out=process_output,
_fg=True)
print("Validation done!\n")
......@@ -843,6 +833,7 @@ def build_host_libraries(model_build_type, abi):
def merge_libs(target_soc,
serial_num,
abi,
project_name,
build_output_dir,
......@@ -911,8 +902,10 @@ def merge_libs(target_soc,
mri_stream += "create %s/libmace_%s.a\n" % \
(model_bin_dir, project_name)
else:
mri_stream += "create %s/libmace_%s.%s.a\n" % \
(model_bin_dir, project_name, target_soc)
device_name = adb_get_device_name_by_serialno(serial_num)
mri_stream += "create %s/libmace_%s.%s.%s.a\n" % \
(model_bin_dir, project_name,
device_name, target_soc)
if model_build_type == BuildType.code:
mri_stream += (
"addlib "
......@@ -969,16 +962,14 @@ def packaging_lib(libmace_output_dir, project_name):
tar_package_path))
stdout_buff = []
process_output = make_output_processor(stdout_buff)
p = sh.tar(
"cvzf",
"%s" % tar_package_path,
glob.glob("%s/*" % project_dir),
"--exclude",
"%s/_tmp" % project_dir,
_out=process_output,
_bg=True,
_err_to_out=True)
p.wait()
sh.tar(
"cvzf",
"%s" % tar_package_path,
glob.glob("%s/*" % project_dir),
"--exclude",
"%s/_tmp" % project_dir,
_out=process_output,
_fg=True)
print("Packaging Done!\n")
......@@ -1068,7 +1059,7 @@ def benchmark_model(abi,
adb_push("%s/benchmark_model" % benchmark_binary_dir, phone_data_dir,
serialno)
p = sh.adb(
sh.adb(
"-s",
serialno,
"shell",
......@@ -1093,9 +1084,7 @@ def benchmark_model(abi,
"--gpu_priority_hint=%s" % gpu_priority_hint,
"--model_file=%s" % mace_model_phone_path,
_out=process_output,
_bg=True,
_err_to_out=True)
p.wait()
_fg=True)
print("Benchmark done!\n")
return "".join(stdout_buff)
......@@ -1134,7 +1123,7 @@ def build_run_throughput_test(abi,
sh.cp("-f", merged_lib_file, "mace/benchmark/libmace_merged.a")
stdout_buff = []
process_output = make_output_processor(stdout_buff)
p = sh.bazel(
sh.bazel(
"build",
"-c",
"opt",
......@@ -1155,9 +1144,7 @@ def build_run_throughput_test(abi,
"openmp=true",
model_tag_build_flag,
_out=process_output,
_bg=True,
_err_to_out=True)
p.wait()
_fg=True)
sh.rm("mace/benchmark/libmace_merged.a")
sh.adb("-s",
......@@ -1187,31 +1174,29 @@ def build_run_throughput_test(abi,
phone_data_dir,
serialno)
p = sh.adb(
"-s",
serialno,
"shell",
"LD_LIBRARY_PATH=%s" % phone_data_dir,
"MACE_CPP_MIN_VLOG_LEVEL=%s" % vlog_level,
"MACE_RUN_PARAMETER_PATH=%s/mace_run.config" %
phone_data_dir,
"%s/model_throughput_test" % phone_data_dir,
"--input_node=%s" % ",".join(input_nodes),
"--output_node=%s" % ",".join(output_nodes),
"--input_shape=%s" % ":".join(input_shapes),
"--output_shape=%s" % ":".join(output_shapes),
"--input_file=%s/%s" % (phone_data_dir, input_file_name),
"--cpu_model_data_file=%s/%s.data" % (phone_data_dir,
cpu_model_tag),
"--gpu_model_data_file=%s/%s.data" % (phone_data_dir,
gpu_model_tag),
"--dsp_model_data_file=%s/%s.data" % (phone_data_dir,
dsp_model_tag),
"--run_seconds=%s" % run_seconds,
_out=process_output,
_bg=True,
_err_to_out=True)
p.wait()
sh.adb(
"-s",
serialno,
"shell",
"LD_LIBRARY_PATH=%s" % phone_data_dir,
"MACE_CPP_MIN_VLOG_LEVEL=%s" % vlog_level,
"MACE_RUN_PARAMETER_PATH=%s/mace_run.config" %
phone_data_dir,
"%s/model_throughput_test" % phone_data_dir,
"--input_node=%s" % ",".join(input_nodes),
"--output_node=%s" % ",".join(output_nodes),
"--input_shape=%s" % ":".join(input_shapes),
"--output_shape=%s" % ":".join(output_shapes),
"--input_file=%s/%s" % (phone_data_dir, input_file_name),
"--cpu_model_data_file=%s/%s.data" % (phone_data_dir,
cpu_model_tag),
"--gpu_model_data_file=%s/%s.data" % (phone_data_dir,
gpu_model_tag),
"--dsp_model_data_file=%s/%s.data" % (phone_data_dir,
dsp_model_tag),
"--run_seconds=%s" % run_seconds,
_out=process_output,
_fg=True)
print("throughput_test done!\n")
......
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或