NVIDIA GPU卷积网络的自动调谐
针对特定设备和工作负载的自动调整对于获得最佳性能至关重要。这是关于如何为NVIDIA GPU调整整个卷积网络。
NVIDIA GPU在TVM中的操作实现是以模板形式编写的。模板有许多可调旋钮(平铺系数、展开等)。将调整神经网络中的所有卷积和深度卷积算子。在调优之后,生成一个日志文件,其中存储了所有所需操作符的最佳旋钮值。当TVM编译器编译这些运算符时,它将查询此日志文件以获得最佳的旋钮值。
还发布了一些NVIDIA GPU的预调参数。可以去NVIDIA GPU基准测试看看结果。
本文不会在Windows或最新版本的macOS上运行。要让它运行,需要将主体包装在if __name__ == "__main__": 块中。
安装依赖项
要在tvm中使用autotvm包,需要安装一些额外的依赖项。(如果使用python2,请将“3”更改为“2”):
pip3 install --user psutil xgboost tornado
为了使TVM在调谐过程中运行更快,建议使用cython作为TVM的FFI。在tvm的根目录下,执行:
pip3 install --user cython
sudo make cython3
现在回到python代码。导入包。
import os
import numpy as np
import tvm
from tvm import relay, autotvm
import tvm.relay.testing
from tvm.autotvm.tuner import XGBTuner, GATuner, RandomTuner, GridSearchTuner
import tvm.contrib.graph_runtime as runtime
Define Network
首先需要在中继前端API中定义网络。可以从tvm中转测试. 也可以从MXNet、ONNX和TensorFlow加载模型。
def get_network(name, batch_size):
"""Get the symbol definition and random weight of a network"""
input_shape = (batch_size, 3, 224, 224)
output_shape = (batch_size, 1000)
if "resnet" in name:
n_layer = int(name.split("-")[1])
mod, params = relay.testing.resnet.get_workload(
num_layers=n_layer, batch_size=batch_size, dtype=dtype
)
elif "vgg" in name:
n_layer = int(name.split("-")[1])
mod, params = relay.testing.vgg.get_workload(
num_layers=n_layer, batch_size=batch_size, dtype=dtype
)
elif name == "mobilenet":
mod, params = relay.testing.mobilenet.get_workload(batch_size=batch_size, dtype=dtype)
elif name == "squeezenet_v1.1":
mod, params = relay.testing.squeezenet.get_workload(
batch_size=batch_size, version="1.1", dtype=dtype
)
elif name == "inception_v3":
input_shape = (batch_size, 3, 299, 299)
mod, params = relay.testing.inception_v3.get_workload(batch_size=batch_size, dtype=dtype)
elif name == "mxnet":
# an example for mxnet model
from mxnet.gluon.model_zoo.vision import get_model
block = get_model("resnet18_v1", pretrained=True)
mod, params = relay.frontend.from_mxnet(block, shape={"data": input_shape}, dtype=dtype)
net = mod["main"]
net = relay.Function(
net.params, relay.nn.softmax(net.body), None, net.type_params, net.attrs
)
mod = tvm.IRModule.from_expr(net)
else:
raise ValueError("Unsupported network: " + name)
return mod, params, input_shape, output_shape
Set Tuning Options
在调整之前,应用一些配置。
#### DEVICE CONFIG ####
target = tvm.target.cuda()
#### TUNING OPTION ####
network = "resnet-18"
log_file = "%s.log" % network
dtype = "float32"
tuning_option = {
"log_filename": log_file,
"tuner": "xgb",
"n_trial": 2000,
"early_stopping": 600,
"measure_option": autotvm.measure_option(
builder=autotvm.LocalBuilder(timeout=10),
runner=autotvm.LocalRunner(number=20, repeat=3, timeout=4, min_repeat_ms=150),
),
}
注意
如何设置调整选项
一般来说,这里提供的默认值工作正常。
如果有大量的时间预算,可以设置n_trial, early_stopping,这使调整运行更长时间。
如果有多个设备,则可以使用所有设备进行测量,以加快调整过程。(请参阅下面的“放大测量”部分)。
开始调谐
现在可以从网络中提取调优任务并开始调优。这里,提供了一个简单的实用函数来优化任务列表。这个函数只是一个按顺序调整它们的初始实现。将在将来引入更复杂的调优调度程序。
# You can skip the implementation of this function for this tutorial.
def tune_tasks(
tasks,
measure_option,
tuner="xgb",
n_trial=1000,
early_stopping=None,
log_filename="tuning.log",
use_transfer_learning=True,
):
# create tmp log file
tmp_log_file = log_filename + ".tmp"
if os.path.exists(tmp_log_file):
os.remove(tmp_log_file)
for i, tsk in enumerate(reversed(tasks)):
prefix = "[Task %2d/%2d] " % (i + 1, len(tasks))
# create tuner
if tuner == "xgb" or tuner == "xgb-rank":
tuner_obj = XGBTuner(tsk, loss_type="rank")
elif tuner == "ga":
tuner_obj = GATuner(tsk, pop_size=100)
elif tuner == "random":
tuner_obj = RandomTuner(tsk)
elif tuner == "gridsearch":
tuner_obj = GridSearchTuner(tsk)
else:
raise ValueError("Invalid tuner: " + tuner)
if use_transfer_learning:
if os.path.isfile(tmp_log_file):
tuner_obj.load_history(autotvm.record.load_from_file(tmp_log_file))
# do tuning
tsk_trial = min(n_trial, len(tsk.config_space))
tuner_obj.tune(
n_trial=tsk_trial,
early_stopping=early_stopping,
measure_option=measure_option,
callbacks=[
autotvm.callback.progress_bar(tsk_trial, prefix=prefix),
autotvm.callback.log_to_file(tmp_log_file),
],
)
# pick best records to a cache file
autotvm.record.pick_best(tmp_log_file, log_filename)
os.remove(tmp_log_file)
最后,启动优化作业并评估端到端性能。
def tune_and_evaluate(tuning_opt):
# extract workloads from relay program
print("Extract tasks...")
mod, params, input_shape, out_shape = get_network(network, batch_size=1)
tasks = autotvm.task.extract_from_program(
mod["main"], target=target, params=params, ops=(relay.op.get("nn.conv2d"),)
)
# run tuning tasks
print("Tuning...")
tune_tasks(tasks, **tuning_opt)
# compile kernels with history best records
with autotvm.apply_history_best(log_file):
print("Compile...")
with tvm.transform.PassContext(opt_level=3):
lib = relay.build_module.build(mod, target=target, params=params)
# load parameters
ctx = tvm.context(str(target), 0)
module = runtime.GraphModule(lib["default"](ctx))
data_tvm = tvm.nd.array((np.random.uniform(size=input_shape)).astype(dtype))
module.set_input("data", data_tvm)
# evaluate
print("Evaluate inference time cost...")
ftimer = module.module.time_evaluator("run", ctx, number=1, repeat=600)
prof_res = np.array(ftimer().results) * 1000 # convert to millisecond
print(
"Mean inference time (std dev): %.2f ms (%.2f ms)"
% (np.mean(prof_res), np.std(prof_res))
)
# We do not run the tuning in our webpage server since it takes too long.
# Uncomment the following line to run it by yourself.
# tune_and_evaluate(tuning_option)
Sample Output
调整需要编译许多程序并从中提取特性。因此建议使用高性能CPU。下面列出了一个示例输出。它需要大约4个小时来获得以下输出在一个32T的AMD Ryzen Threadripper。调谐目标是NVIDIA 1080TI。(在编译过程中可以看到一些错误。如果调谐没有卡住,就可以了。)
Extract tasks...
Tuning...
[Task 1/12] Current/Best: 541.83/3570.66 GFLOPS | Progress: (960/2000) | 1001.31 s Done.
[Task 2/12] Current/Best: 0.56/ 803.33 GFLOPS | Progress: (704/2000) | 608.08 s Done.
[Task 3/12] Current/Best: 103.69/1141.25 GFLOPS | Progress: (768/2000) | 702.13 s Done.
[Task 4/12] Current/Best: 2905.03/3925.15 GFLOPS | Progress: (864/2000) | 745.94 sterminate called without an active exception
[Task 4/12] Current/Best: 2789.36/3925.15 GFLOPS | Progress: (1056/2000) | 929.40 s Done.
[Task 5/12] Current/Best: 89.06/1076.24 GFLOPS | Progress: (704/2000) | 601.73 s Done.
[Task 6/12] Current/Best: 40.39/2129.02 GFLOPS | Progress: (1088/2000) | 1125.76 s Done.
[Task 7/12] Current/Best: 4090.53/5007.02 GFLOPS | Progress: (800/2000) | 903.90 s Done.
[Task 8/12] Current/Best: 4.78/1272.28 GFLOPS | Progress: (768/2000) | 749.14 s Done.
[Task 9/12] Current/Best: 1391.45/2325.08 GFLOPS | Progress: (992/2000) | 1084.87 s Done.
[Task 10/12] Current/Best: 1995.44/2383.59 GFLOPS | Progress: (864/2000) | 862.60 s Done.
[Task 11/12] Current/Best: 4093.94/4899.80 GFLOPS | Progress: (224/2000) | 240.92 sterminate called without an active exception
[Task 11/12] Current/Best: 3487.98/4909.91 GFLOPS | Progress: (480/2000) | 534.96 sterminate called without an active exception
[Task 11/12] Current/Best: 4636.84/4912.17 GFLOPS | Progress: (1184/2000) | 1381.16 sterminate called without an active exception
[Task 11/12] Current/Best: 50.12/4912.17 GFLOPS | Progress: (1344/2000) | 1602.81 s Done.
[Task 12/12] Current/Best: 3581.31/4286.30 GFLOPS | Progress: (736/2000) | 943.52 s Done.
Compile...
Evaluate inference time cost...
Mean inference time (std dev): 1.07 ms (0.05 ms)
作为参考基线,MXNet+TensorRT在resnet-18上的时间开销是1.30ms,所以要快一点。
注意
遇到困难?
自动调谐模块容易出错。如果总是看到“0.00/0.00 GFLOPS”,那么一定是出了什么问题。
首先,确保设置了正确的设备配置。然后,可以通过在脚本开头添加这些行来打印调试信息。它将打印每个测量结果,可以在其中找到有用的错误消息。
导入日志记录
logging.getLogger('autotvm').setLevel(logging.DEBUG)
最后,请随时向寻求帮助https://discus.tvm.apache.org
使用多个设备放大测量
如果有多个设备,可以使用所有设备进行测量。TVM使用RPC跟踪器来管理分布式设备。RPC跟踪器是一个集中的控制器节点。可以把所有设备注册到跟踪器上。例如,如果有10个GPU卡,可以将它们全部注册到跟踪器中,并并行运行10个测量,从而加快调谐过程。
要启动RPC跟踪器,在主机上运行此命令。在整个整定过程中需要跟踪器,所以需要为这个命令打开一个新的终端:
python -m tvm.exec.rpc_tracker --host=0.0.0.0 --port=9190
The expected output is
INFO:RPCTracker:bind to 0.0.0.0:9190
然后为RPC服务器打开另一个新终端。需要为每个专用设备启动一个服务器。使用字符串键来区分设备的类型。可以选一个你喜欢的名字。(注意:对于rocm后端,编译器有一些内部错误,需要在参数列表中add –no-fork。)
python -m tvm.exec.rpc_server --tracker=0.0.0.0:9190 --key=1080ti
After registering devices, we can confirm it by querying rpc_tracker
python -m tvm.exec.query_rpc_tracker --host=0.0.0.0 --port=9190
For example, if we have four 1080ti, two titanx and one gfx900, the output can be
Queue Status
----------------------------------
key total free pending
----------------------------------
1080ti 4 4 0
titanx 2 2 0
gfx900 1 1 0
----------------------------------
最后,需要将调优选项更改为使用RPCRunner。使用下面的代码替换上面相应的部件。
tuning_option = {
"log_filename": log_file,
"tuner": "xgb",
"n_trial": 2000,
"early_stopping": 600,
"measure_option": autotvm.measure_option(
builder=autotvm.LocalBuilder(timeout=10),
runner=autotvm.RPCRunner(
"1080ti", # change the device key to your key
"0.0.0.0",
9190,
number=20,
repeat=3,
timeout=4,
min_repeat_ms=150,
),
),
}
https://tvm.apache.org/docs/tutorials/autotvm/tune_relay_cuda.html
下载Python源代码:tune_relay_cuda.py
下载Jupyter笔记本:tune_relay_cuda.ipynbDownload
Python
source
code:
tune_relay_cuda.py
Download
Jupyter
notebook:
tune_relay_cuda.ipynb