tensorlayer3/tensorlayer/cli/train.py

172 lines
5.8 KiB
Python

#! /usr/bin/python
# -*- coding: utf-8 -*-
"""
tl train
========
(Alpha release - usage might change later)
The tensorlayer.cli.train module provides the ``tl train`` subcommand.
It helps the user bootstrap a TensorFlow/TensorLayer program for distributed training
using multiple GPU cards or CPUs on a computer.
You need to first setup the `CUDA_VISIBLE_DEVICES <http://acceleware.com/blog/cudavisibledevices-masking-gpus>`_
to tell ``tl train`` which GPUs are available. If the CUDA_VISIBLE_DEVICES is not given,
``tl train`` would try best to discover all available GPUs.
In distribute training, each TensorFlow program needs a TF_CONFIG environment variable to describe
the cluster. It also needs a master daemon to
monitor all trainers. ``tl train`` is responsible
for automatically managing these two tasks.
Usage
-----
tl train [-h] [-p NUM_PSS] [-c CPU_TRAINERS] <file> [args [args ...]]
.. code-block:: bash
# example of using GPU 0 and 1 for training mnist
CUDA_VISIBLE_DEVICES="0,1"
tl train example/tutorial_mnist_distributed.py
# example of using CPU trainers for inception v3
tl train -c 16 example/tutorial_imagenet_inceptionV3_distributed.py
# example of using GPU trainers for inception v3 with customized arguments
# as CUDA_VISIBLE_DEVICES is not given, tl would try to discover all available GPUs
tl train example/tutorial_imagenet_inceptionV3_distributed.py -- --batch_size 16
Command-line Arguments
----------------------
- ``file``: python file path.
- ``NUM_PSS`` : The number of parameter servers.
- ``CPU_TRAINERS``: The number of CPU trainers.
It is recommended that ``NUM_PSS + CPU_TRAINERS <= cpu count``
- ``args``: Any parameter after ``--`` would be passed to the python program.
Notes
-----
A parallel training program would require multiple parameter servers
to help parallel trainers to exchange intermediate gradients.
The best number of parameter servers is often proportional to the
size of your model as well as the number of CPUs available.
You can control the number of parameter servers using the ``-p`` parameter.
If you have a single computer with massive CPUs, you can use the ``-c`` parameter
to enable CPU-only parallel training.
The reason we are not supporting GPU-CPU co-training is because GPU and
CPU are running at different speeds. Using them together in training would
incur stragglers.
"""
import argparse
import json
import multiprocessing
import os
import platform
import re
import subprocess
import sys
PORT_BASE = 10000
def _get_gpu_ids():
if 'CUDA_VISIBLE_DEVICES' in os.environ:
return [int(x) for x in os.environ.get('CUDA_VISIBLE_DEVICES', '').split(',')]
if platform.system() in ['Darwin', 'Linux']:
return [int(d.replace('nvidia', '')) for d in os.listdir('/dev') if re.match('^nvidia\d+$', d)]
else:
print('Please set CUDA_VISIBLE_DEVICES (see http://acceleware.com/blog/cudavisibledevices-masking-gpus)')
return []
GPU_IDS = _get_gpu_ids()
def create_tf_config(cluster_spec, task_type, task_index):
return {
'cluster': cluster_spec,
'task': {
'type': task_type,
'index': task_index
},
}
def create_tf_jobs(cluster_spec, prog, args):
gpu_assignment = dict((('worker', idx), gpu_idx) for (idx, gpu_idx) in enumerate(GPU_IDS))
for job_type in cluster_spec:
for task_index in range(len(cluster_spec[job_type])):
new_env = os.environ.copy()
new_env.update(
{
'CUDA_VISIBLE_DEVICES': str(gpu_assignment.get((job_type, task_index), '')),
'TF_CONFIG': json.dumps(create_tf_config(cluster_spec, job_type, task_index)),
}
)
yield subprocess.Popen(['python3', prog] + args, env=new_env)
def validate_arguments(args):
if args.num_pss < 1:
print('Value error: must have ore than one parameter servers.')
exit(1)
if not GPU_IDS:
num_cpus = multiprocessing.cpu_count()
if args.cpu_trainers > num_cpus:
print('Value error: there are %s available CPUs but you are requiring %s.' % (num_cpus, args.cpu_trainers))
exit(1)
if not os.path.isfile(args.file):
print('Value error: model trainning file does not exist')
exit(1)
def main(args):
validate_arguments(args)
num_workers = len(GPU_IDS) if GPU_IDS else args.cpu_trainers
print('Using program %s with args %s' % (args.file, ' '.join(args.args)))
print('Using %d workers, %d parameter servers, %d GPUs.' % (num_workers, args.num_pss, len(GPU_IDS)))
cluster_spec = {
'ps': ['localhost: %d' % (PORT_BASE + i) for i in range(args.num_pss)],
'worker': ['localhost: %d' % (PORT_BASE + args.num_pss + i) for i in range(num_workers)]
}
processes = list(create_tf_jobs(cluster_spec, args.file, args.args))
try:
print('Press ENTER to exit the training ...')
sys.stdin.readline()
except KeyboardInterrupt: # https://docs.python.org/3/library/exceptions.html#KeyboardInterrupt
print('Keyboard interrupt received')
finally:
print('stopping all subprocesses ...')
for p in processes:
p.kill()
for p in processes:
p.wait()
print('END')
def build_arg_parser(parser):
parser.add_argument('-p', '--pss', dest='num_pss', type=int, default=1, help='number of parameter servers')
parser.add_argument('-c', '--cpu_trainers', dest='cpu_trainers', type=int, default=1, help='number of CPU trainers')
parser.add_argument('file', help='model trainning file path')
parser.add_argument('args', nargs='*', type=str, help='arguments to <file>')
if __name__ == "__main__":
parser = argparse.ArgumentParser()
build_arg_parser(parser)
args = parser.parse_args()
main(args)