def prepare_model(self):
"""load 网络模型，（更新输出层）
"""
# 需要更新模型输出层种类数目的场景，用户自行决定模型文件的路径，非必要场景
# 用户可以直接使用训练框架来读取模型
origin_model_path = self.train_ins.get_model_base_path() + "/Inception-BN-symbol.json"
weight_file_path = self.train_ins.get_model_base_path() + "/Inception-BN-0126.params"
fixed_model_path = self.train_ins.get_model_base_path() + "/fixed_train.symbol.json"
# AVA-SDK 获取数据集类型数 && 更新网络模型输出层
output_layer_num = utils.get_sampleset_class_num()
# old_output_layer_name = model_utils.update_model_output_num(
# origin_model_path, fixed_model_path, output_layer_num)
sym,arg_params,aux_params=mx.model.load_checkpoint(prefix='/workspace/model/Inception-BN',epoch=126)
layer_name='flatten'
all_layers=sym.get_internals()
net = all_layers[layer_name+'_output']
net = mx.symbol.FullyConnected(data=net, num_hidden=output_layer_num, name='fc1_new')
net = mx.symbol.SoftmaxOutput(data=net, name='softmax')
sym = net
# sym = mx.symbol.load(fixed_model_path)
gpu_count = self.solver_config.get('gpu_counts', 0)
ctx = [mx.cpu()] if gpu_count == 0 else [
mx.gpu(i) for i in xrange(gpu_count)
]
mod = mx.mod.Module(symbol=sym, context=ctx)
mod.bind(data_shapes=self.train_data.provide_data,
label_shapes=self.train_data.provide_label)
# 默认权值初始化方式
mod.init_params(initializer=mx.init.Xavier(rnd_type='gaussian',
factor_type="in",
magnitude=2))
# AVA-SDK 在替换网络输出层的场景下读取权重参数
# arg_params, aux_params = model_utils.load_model_params(
# weight_file_path, old_output_layer_name)
arg_params = dict({k:arg_params[k] for k in arg_params if 'fc1' not in k})
if arg_params:
logger.info("set pretrained weights")
mod.set_params(arg_params, aux_params, allow_missing=True)
self.mod = mod
def prepare_model(self):
'''load 网络模型，（更新输出层）
'''
sym, arg_params, aux_params = mx.model.load_checkpoint(
'/workspace/model/resnet-50', 0)
num_hidden = 10
(new_sym, new_args) = self.get_fine_tune_model(sym, arg_params,
num_hidden)
# get information of ctx
gpu_count = self.solver_config.get('gpu_counts', 0)
ctx = [mx.cpu()
] if gpu_count == 0 else [mx.gpu(i) for i in xrange(gpu_count)]
mod = mx.mod.Module(symbol=new_sym, context=ctx)
mod.bind(data_shapes=self.train_data.provide_data,
label_shapes=self.train_data.provide_label)
logger.info("set pretrained weights")
mod.set_params(new_args, aux_params, allow_missing=True)
self.mod = mod
def prepare_model(self):
"""load 网络模型，（更新输出层）
"""
# 需要更新模型输出层种类数目的场景，用户自行决定模型文件的路径，非必要场景
# 用户可以直接使用训练框架来读取模型
# 替换成自己需要的模型名以及参数名 Riheng 2018/02/23
origin_model_path = self.train_ins.get_model_base_path(
) + "/resnet-50-symbol.json"
weight_file_path = self.train_ins.get_model_base_path(
) + "/resnet-50-0000.params"
fixed_model_path = self.train_ins.get_model_base_path(
) + "/fixed_resnet-50-symbol.json"
# AVA-SDK 获取数据集类型数 && 更新网络模型输出层
output_layer_num = utils.get_sampleset_class_num()
old_output_layer_name = model_utils.update_model_output_num(
origin_model_path, fixed_model_path, output_layer_num)
sym = mx.symbol.load(fixed_model_path)
gpu_count = self.solver_config.get('gpu_counts', 0)
ctx = [mx.cpu()
] if gpu_count == 0 else [mx.gpu(i) for i in xrange(gpu_count)]
mod = mx.mod.Module(symbol=sym, context=ctx)
mod.bind(data_shapes=self.train_data.provide_data,
label_shapes=self.train_data.provide_label)
# 默认权值初始化方式
mod.init_params(initializer=mx.init.Xavier(
rnd_type='gaussian', factor_type="in", magnitude=2))
# AVA-SDK 在替换网络输出层的场景下读取权重参数
arg_params, aux_params = model_utils.load_model_params(
weight_file_path, old_output_layer_name)
if arg_params:
logger.info("set pretrained weights")
mod.set_params(arg_params, aux_params, allow_missing=True)
self.mod = mod
def signal_handler(self, signum, stack):
logger.info("received signal: %s, do clean_up", signum)
self.clean_up()
sys.exit()
def start_new_training(self):
try:
# 绑定信号，如果是接收到信号，表示用户自己选择退出训练实例
# 训练实例状态为正常结束
SUPPORTED_SIGNALS = (
signal.SIGINT,
signal.SIGTERM,
)
for signum in SUPPORTED_SIGNALS:
try:
signal.signal(signum, self.signal_handler)
logger.info("Bind signal '%s' success to %s", signum,
self.signal_handler)
except Exception as identifier:
logger.warning("Bind signal '%s' failed, err: %s", signum,
identifier)
# AVA-SDK 初始化一个训练实例
self.train_ins = train.TrainInstance()
logger.info("start new tarining, training_ins_id: %s",
self.train_ins.get_training_ins_id())
logger.info("prepare_train_config")
self.prepare_train_config()
logger.info("prepare_solver_config")
self.prepare_solver_config()
logger.info("prepare_sampleset_config")
self.prepare_sampleset_data()
logger.info("prepare_model")
self.prepare_model()
opts = self.train_config
opts.update(self.solver_config)
fit_args = {k: opts.get(k) for k in FIT_KWARGS_KEYS}
logger.info("fit args: %s" % fit_args)
self.mod.fit(self.train_data, eval_data=self.val_data, **fit_args)
logger.info("training finish")
err_msg = ""
except Exception as err:
err_msg = "training failed, err: %s" % (err)
logger.info(err_msg)
traceback.print_exc(file=sys.stderr)
self.clean_up(err_msg=err_msg)
def start_new_training():
# binding signals
SUPPORTED_SIGNALS = (
signal.SIGINT,
signal.SIGTERM,
)
for signum in SUPPORTED_SIGNALS:
try:
signal.signal(signum, signal_handler)
logger.info("Bind signal '%s' success to %s", signum,
signal_handler)
except Exception as identifier:
logger.warning("Bind signal '%s' failed, err: %s", signum,
identifier)
try:
# parse args
parser = argparse.ArgumentParser(
description="train imagenet-1k",
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
fit.add_fit_args(parser)
data.add_data_args(parser)
data.add_data_aug_args(parser)
# use a large aug level
data.set_data_aug_level(parser, 3)
parser.set_defaults(
# network
network='resnet',
num_layers=50,
# data
num_classes=10,
num_examples=60000,
image_shape='3,28,28',
min_random_scale=1, # if input image has min size k, suggest to use
# 256.0/x, e.g. 0.533 for 480
# train
num_epochs=80,
lr_step_epochs='30,60',
dtype='float32',
batch_size=32)
args = parser.parse_args()
# AVA-SDK new an Instance
train_ins = train.TrainInstance()
# add CALLBACK
batch_end_cb = train_ins.get_monitor_callback(
"mxnet", batch_size=args.batch_size, batch_freq=10)
args.batch_end_callback = batch_end_cb
# load network
from importlib import import_module
net = import_module('symbols.' + args.network)
sym = net.get_symbol(**vars(args))
# train
fit.fit(args, sym, data.get_rec_iter)
logger.info("training finish")
err_msg = ""
if train_ins == None:
return
train_ins.done(err_msg=err_msg)
except Exception as err:
err_msg = "training failed, err: %s" % (err)
logger.info(err_msg)
traceback.print_exc(file=sys.stderr)
if train_ins == None:
return
train_ins.done(err_msg=err_msg)
def signal_handler(signum, frame):
logger.info("received signal: %s, do clean_up", signum)
train_ins = frame.f_locals['train_ins']
clean_up(train_ins)
exit()
def start_new_training():
# AVA-SDK training Instance
SUPPORTED_SIGNALS = (signal.SIGINT, signal.SIGTERM,)
for signum in SUPPORTED_SIGNALS:
try:
signal.signal(signum, signal_handler)
logger.info("Bind signal '%s' success to %s",
signum, signal_handler)
except Exception as identifier:
logger.warning(
"Bind signal '%s' failed, err: %s", signum, identifier)
try:
# parse args
train_ins = train.TrainInstance()
err_msg=''
# add CALLBACK
solver_param = caffe_pb2.SolverParameter()
with open('/workspace/model/lenet_solver.prototxt','r') as f:
text_format.Merge(f.read(),solver_param)
solver_param.snapshot_prefix = train_ins.get_snapshot_base_path()
logger.info("saving to %s", solver_param.snapshot_prefix)
fixed_solver = train_ins.get_base_path() + "/solver.prototxt"
with open(fixed_solver, 'w') as f:
f.write(str(solver_param))
logger.info("write fixed solver to %s", fixed_solver)
# AVA-SDK start caffe process
training_cmd = ['caffe','train','-solver',fixed_solver,'-gpu','0']
proc = cmd.startproc(training_cmd)
logger.info("Started %s", proc)
# AVA-SDK add caffe callback
cmd.logproc(proc, [train_ins.get_monitor_callback("caffe")])
exit_code = proc.wait()
logger.info("Finished proc with code %s", exit_code)
logger.info("Gracefully shutdown after 5s, wait cleaner ...")
time.sleep(5)
logger.info("Done.")
if exit_code != 0:
logger.error(
"training exit code [%d] != 0, raise Exception", exit_code)
raise Exception("training exit code [%d] != 0" % (exit_code))
train_ins.done(err_msg=err_msg)
except Exception as err:
err_msg = "training failed, err: %s" % (err)
logger.info(err_msg)
traceback.print_exc(file=sys.stderr)
if train_ins == None:
return
train_ins.done(err_msg=err_msg)
def prepare_train_config(self):
"""配置训练参数"""
# AVA-SDK 获取训练参数
"""
1) 获取所有配置 example
param_dict = params.get_all()
value1 = param_dict["key1"]
2) 获取某项value
value1 = params.get_value("key1", default=1)
"""
snapshot_interval_epochs = params.get_value("snapshot_interval_epochs",
default=1)
max_epochs = params.get_value("max_epochs", default=3)
rand_crop = params.get_value("rand_crop", default=True)
rand_mirror = params.get_value("rand_mirror", default=True)
batch_size, actual_batch_size, val_batch_size = utils.get_batch_size()
batch_size = params.get_value("batchSize", default=8)
val_batch_size = params.get_value("valBatchSize", default=batch_size)
use_gpu, cores = utils.get_cores()
logger.info("Cores GPU=%s, count=%d", use_gpu, cores)
actual_batch_size = batch_size if not use_gpu else batch_size * cores
if use_gpu:
val_batch_size *= cores
# USING the trainning batch size as valadition batch size
val_batch_size = actual_batch_size
# crop_w, crop_h = utils.get_crop_size()
# 根据模型的输入要求选择 crop_size
crop_w = params.get_value("crop_w")
crop_h = params.get_value("crop_h")
# 添加监控
snapshot_prefix = self.train_ins.get_snapshot_base_path() + "/snapshot"
kv_store = "device"
kv = mx.kvstore.create(kv_store)
'''
rank = int(kv.rank)
if rank > 0:
snapshot_prefix += "-%s" % rank
'''
batch_freq = 10 # 打印/上报指标的 batch 粒度
batch_of_epoch = utils.ceil_by_level(
float(utils.get_sampleset_num() / actual_batch_size))
# AVA-SDK mxnet monitor callback 初始化
batch_end_cb = self.train_ins.get_monitor_callback(
"mxnet", batch_size=actual_batch_size, batch_freq=batch_freq)
epoch_end_cb = [
# mxnet default epoch callback
mx.callback.do_checkpoint(snapshot_prefix,
snapshot_interval_epochs),
self.train_ins.get_epoch_end_callback(
"mxnet",
batch_of_epoch=batch_of_epoch,
epoch_interval=snapshot_interval_epochs,
other_files=[])
]
# 训练参数，用户可以自行配置
self.train_config = {
"input_data_shape": (CROP_CHANNELS, crop_h, crop_w),
"rand_crop": rand_crop,
"rand_mirror": rand_mirror,
"batch_size": batch_size,
"actual_batch_size": actual_batch_size,
"val_batch_size": val_batch_size,
# fit_args
"eval_metric":
mxnet_monitor.full_mxnet_metrics(), # AVA-SDK 获取mxnet metric 列表
"epoch_end_callback": epoch_end_cb,
"batch_end_callback": batch_end_cb,
"kvstore": kv,
"num_epoch": max_epochs,
}
def start_new_training():
# AVA-SDK training Instance
SUPPORTED_SIGNALS = (signal.SIGINT, signal.SIGTERM,)
for signum in SUPPORTED_SIGNALS:
try:
signal.signal(signum, signal_handler)
logger.info("Bind signal '%s' success to %s",
signum, signal_handler)
except Exception as identifier:
logger.warning(
"Bind signal '%s' failed, err: %s", signum, identifier)
try:
# parse args
train_ins = train.TrainInstance()
err_msg=''
# add CALLBACK
# AVA-SDK start caffe process
out_dir = train_ins.get_snapshot_base_path()
roidb_path = train_ins.get_trainset_base_path() + "/cache/gt_roidb.pkl"
training_cmd = ['python', 'detect_py_faster_rcnn.py', '--solver', 'vgg_solver.prototxt', '--gpu', '0',
'--output_path', out_dir, '--ava_roidb_path', roidb_path,
'--train_base_path', train_ins.get_trainset_base_path()+'/cache']
proc = cmd.startproc(training_cmd)
logger.info("Started %s", proc)
# AVA-SDK add caffe callback
cmd.logproc(proc, [train_ins.get_monitor_callback("caffe")])
exit_code = proc.wait()
logger.info("Finished proc with code %s", exit_code)
logger.info("Gracefully shutdown after 5s, wait cleaner ...")
time.sleep(5)
logger.info("Done.")
if exit_code != 0:
logger.error(
"training exit code [%d] != 0, raise Exception", exit_code)
raise Exception("training exit code [%d] != 0" % (exit_code))
train_ins.done(err_msg=err_msg)
except Exception as err:
err_msg = "training failed, err: %s" % (err)
logger.info(err_msg)
traceback.print_exc(file=sys.stderr)
if train_ins == None:
return
train_ins.done(err_msg=err_msg)