def __init__(self, configFn, ctx, outFolder, threshold):
os.environ["MXNET_CUDNN_AUTOTUNE_DEFAULT"] = "0"
config = importlib.import_module(
configFn.replace('.py', '').replace('/', '.'))
_, _, _, _, _, _, self.__pModel, _, self.__pTest, self.transform, _, _, _ = config.get_config(
is_train=False)
self.__pModel = patch_config_as_nothrow(self.__pModel)
self.__pTest = patch_config_as_nothrow(self.__pTest)
self.resizeParam = (800, 1200)
if callable(self.__pTest.nms.type):
self.__nms = self.__pTest.nms.type(self.__pTest.nms.thr)
else:
from operator_py.nms import py_nms_wrapper
self.__nms = py_nms_wrapper(self.__pTest.nms.thr)
arg_params, aux_params = load_checkpoint(self.__pTest.model.prefix,
self.__pTest.model.epoch)
sym = self.__pModel.test_symbol
from utils.graph_optimize import merge_bn
sym, arg_params, aux_params = merge_bn(sym, arg_params, aux_params)
self.__mod = DetModule(
sym,
data_names=['data', 'im_info', 'im_id', 'rec_id'],
context=ctx)
self.__mod.bind(data_shapes=[('data', (1, 3, self.resizeParam[0],
self.resizeParam[1])),
('im_info', (1, 3)), ('im_id', (1, )),
('rec_id', (1, ))],
for_training=False)
self.__mod.set_params(arg_params, aux_params, allow_extra=False)
self.__saveSymbol(sym, outFolder,
self.__pTest.model.prefix.split('/')[-1])
self.__threshold = threshold
self.outFolder = outFolder
collector_queue_depth=2,
kv=None)
print("total number of images: {}".format(loader.total_record))
data_names = [k[0] for k in loader.provide_data]
if index_split == 0:
arg_params, aux_params = load_checkpoint(
pTest.model.prefix, args.epoch or pTest.model.epoch)
if pModel.process_weight is not None:
pModel.process_weight(sym, arg_params, aux_params)
# merge batch normalization to speedup test
from utils.graph_optimize import merge_bn
sym, arg_params, aux_params = merge_bn(sym, arg_params, aux_params)
sym.save(pTest.model.prefix + "_test.json")
# infer shape
worker_data_shape = dict(loader.provide_data +
loader.provide_label)
for key in worker_data_shape:
worker_data_shape[key] = (
pKv.batch_image, ) + worker_data_shape[key][1:]
arg_shape, _, aux_shape = sym.infer_shape(**worker_data_shape)
_, out_shape, _ = sym.get_internals().infer_shape(
**worker_data_shape)
out_shape_dict = list(
zip(sym.get_internals().list_outputs(), out_shape))
_, out_shape, _ = sym.infer_shape(**worker_data_shape)
terminal_out_shape_dict = zip(sym.list_outputs(), out_shape)
def train_net(config):
pGen, pKv, pRpn, pRoi, pBbox, pDataset, pModel, pOpt, pTest, \
transform, data_name, label_name, metric_list = config.get_config(is_train=True)
pGen = patch_config_as_nothrow(pGen)
pKv = patch_config_as_nothrow(pKv)
pRpn = patch_config_as_nothrow(pRpn)
pRoi = patch_config_as_nothrow(pRoi)
pBbox = patch_config_as_nothrow(pBbox)
pDataset = patch_config_as_nothrow(pDataset)
pModel = patch_config_as_nothrow(pModel)
pOpt = patch_config_as_nothrow(pOpt)
pTest = patch_config_as_nothrow(pTest)
ctx = [mx.gpu(int(i)) for i in pKv.gpus]
pretrain_prefix = pModel.pretrain.prefix
pretrain_epoch = pModel.pretrain.epoch
prefix = pGen.name
save_path = os.path.join("experiments", prefix)
begin_epoch = pOpt.schedule.begin_epoch
end_epoch = pOpt.schedule.end_epoch
lr_iter = pOpt.schedule.lr_iter
# only rank==0 print all debug infos
kvstore_type = "dist_sync" if os.environ.get(
"DMLC_ROLE") == "worker" else pKv.kvstore
kv = mx.kvstore.create(kvstore_type)
rank = kv.rank
# for distributed training using shared file system
os.makedirs(save_path, exist_ok=True)
from utils.logger import config_logger
config_logger(os.path.join(save_path, "log.txt"))
model_prefix = os.path.join(save_path, "checkpoint")
# set up logger
logger = logging.getLogger()
sym = pModel.train_symbol
# setup multi-gpu
input_batch_size = pKv.batch_image * len(ctx)
# print config
# if rank == 0:
# logger.info(pprint.pformat(config))
# load dataset and prepare imdb for training
image_sets = pDataset.image_set
roidbs = [
pkl.load(open("data/cache/{}.roidb".format(i), "rb"),
encoding="latin1") for i in image_sets
]
roidb = reduce(lambda x, y: x + y, roidbs)
# filter empty image
roidb = [rec for rec in roidb if rec["gt_bbox"].shape[0] > 0]
# add flip roi record
flipped_roidb = []
for rec in roidb:
new_rec = rec.copy()
new_rec["flipped"] = True
flipped_roidb.append(new_rec)
roidb = roidb + flipped_roidb
from core.detection_input import AnchorLoader
train_data = AnchorLoader(roidb=roidb,
transform=transform,
data_name=data_name,
label_name=label_name,
batch_size=input_batch_size,
shuffle=True,
kv=kv,
num_worker=pGen.loader_worker or 12,
num_collector=pGen.loader_collector or 1,
worker_queue_depth=2,
collector_queue_depth=2)
# infer shape
worker_data_shape = dict(train_data.provide_data +
train_data.provide_label)
for key in worker_data_shape:
worker_data_shape[key] = (
pKv.batch_image, ) + worker_data_shape[key][1:]
arg_shape, _, aux_shape = sym.infer_shape(**worker_data_shape)
_, out_shape, _ = sym.get_internals().infer_shape(**worker_data_shape)
out_shape_dict = list(zip(sym.get_internals().list_outputs(), out_shape))
_, out_shape, _ = sym.infer_shape(**worker_data_shape)
terminal_out_shape_dict = zip(sym.list_outputs(), out_shape)
if rank == 0:
logger.info('parameter shape')
logger.info(
pprint.pformat(
[i for i in out_shape_dict if not i[0].endswith('output')]))
logger.info('intermediate output shape')
logger.info(
pprint.pformat(
[i for i in out_shape_dict if i[0].endswith('output')]))
logger.info('terminal output shape')
logger.info(pprint.pformat([i for i in terminal_out_shape_dict]))
# memonger
if pModel.memonger:
last_block = pModel.memonger_until or ""
if rank == 0:
logger.info("do memonger up to {}".format(last_block))
type_dict = {k: np.float32 for k in worker_data_shape}
sym = search_plan_to_layer(sym,
last_block,
1000,
type_dict=type_dict,
**worker_data_shape)
# load and initialize params
if pOpt.schedule.begin_epoch != 0:
arg_params, aux_params = load_checkpoint(model_prefix, begin_epoch)
elif pModel.from_scratch:
arg_params, aux_params = dict(), dict()
else:
arg_params, aux_params = load_checkpoint(pretrain_prefix,
pretrain_epoch)
if pModel.process_weight is not None:
pModel.process_weight(sym, arg_params, aux_params)
'''
there are some conflicts between `mergebn` and `attach_quantized_node` in graph_optimize.py
when mergebn ahead of attach_quantized_node
such as `Symbol.ComposeKeyword`
'''
if pModel.QuantizeTrainingParam is not None and pModel.QuantizeTrainingParam.quantize_flag:
pQuant = pModel.QuantizeTrainingParam
assert pGen.fp16 == False, "current quantize training only support fp32 mode."
from utils.graph_optimize import attach_quantize_node
_, out_shape, _ = sym.get_internals().infer_shape(**worker_data_shape)
out_shape_dictoinary = dict(
zip(sym.get_internals().list_outputs(), out_shape))
sym = attach_quantize_node(sym, out_shape_dictoinary,
pQuant.WeightQuantizeParam,
pQuant.ActQuantizeParam,
pQuant.quantized_op)
# merge batch normalization to save memory in fix bn training
from utils.graph_optimize import merge_bn
sym, arg_params, aux_params = merge_bn(sym, arg_params, aux_params)
if pModel.random:
import time
mx.random.seed(int(time.time()))
np.random.seed(int(time.time()))
init = mx.init.Xavier(factor_type="in", rnd_type='gaussian', magnitude=2)
init.set_verbosity(verbose=True)
# create solver
fixed_param = pModel.pretrain.fixed_param
excluded_param = pModel.pretrain.excluded_param
data_names = [k[0] for k in train_data.provide_data]
label_names = [k[0] for k in train_data.provide_label]
if pModel.teacher_param:
from models.KD.utils import create_teacher_module
from models.KD.detection_module import KDDetModule
t_mod, t_label_name, t_label_shape = create_teacher_module(
pModel.teacher_param, worker_data_shape, input_batch_size, ctx,
rank, logger)
mod = KDDetModule(sym,
teacher_module=t_mod,
teacher_label_names=t_label_name,
teacher_label_shapes=t_label_shape,
data_names=data_names,
label_names=label_names,
logger=logger,
context=ctx,
fixed_param=fixed_param,
excluded_param=excluded_param)
else:
mod = DetModule(sym,
data_names=data_names,
label_names=label_names,
logger=logger,
context=ctx,
fixed_param=fixed_param,
excluded_param=excluded_param)
eval_metrics = mx.metric.CompositeEvalMetric(metric_list)
# callback
batch_end_callback = [
callback.Speedometer(train_data.batch_size,
frequent=pGen.log_frequency)
]
batch_end_callback += pModel.batch_end_callbacks or []
epoch_end_callback = callback.do_checkpoint(model_prefix)
sym.save(model_prefix + ".json")
# decide learning rate
lr_mode = pOpt.optimizer.lr_mode or 'step'
base_lr = pOpt.optimizer.lr * kv.num_workers
lr_factor = pOpt.schedule.lr_factor or 0.1
iter_per_epoch = len(train_data) // input_batch_size
total_iter = iter_per_epoch * (end_epoch - begin_epoch)
lr_iter = [total_iter + it if it < 0 else it for it in lr_iter]
lr_iter = [it // kv.num_workers for it in lr_iter]
lr_iter = [it - iter_per_epoch * begin_epoch for it in lr_iter]
lr_iter_discount = [it for it in lr_iter if it > 0]
current_lr = base_lr * (lr_factor**(len(lr_iter) - len(lr_iter_discount)))
if rank == 0:
logging.info('total iter {}'.format(total_iter))
logging.info('lr {}, lr_iters {}'.format(current_lr, lr_iter_discount))
logging.info('lr mode: {}'.format(lr_mode))
if pOpt.warmup and pOpt.schedule.begin_epoch == 0:
if rank == 0:
logging.info('warmup lr {}, warmup step {}'.format(
pOpt.warmup.lr, pOpt.warmup.iter))
if lr_mode == 'step':
lr_scheduler = WarmupMultiFactorScheduler(
step=lr_iter_discount,
factor=lr_factor,
warmup=True,
warmup_type=pOpt.warmup.type,
warmup_lr=pOpt.warmup.lr,
warmup_step=pOpt.warmup.iter)
elif lr_mode == 'cosine':
warmup_lr_scheduler = AdvancedLRScheduler(mode='linear',
base_lr=pOpt.warmup.lr,
target_lr=base_lr,
niters=pOpt.warmup.iter)
cosine_lr_scheduler = AdvancedLRScheduler(
mode='cosine',
base_lr=base_lr,
target_lr=0,
niters=(iter_per_epoch *
(end_epoch - begin_epoch)) - pOpt.warmup.iter)
lr_scheduler = LRSequential(
[warmup_lr_scheduler, cosine_lr_scheduler])
else:
raise NotImplementedError
else:
if lr_mode == 'step':
lr_scheduler = WarmupMultiFactorScheduler(step=lr_iter_discount,
factor=lr_factor)
elif lr_mode == 'cosine':
lr_scheduler = AdvancedLRScheduler(mode='cosine',
base_lr=base_lr,
target_lr=0,
niters=iter_per_epoch *
(end_epoch - begin_epoch))
else:
lr_scheduler = None
# optimizer
optimizer_params = dict(momentum=pOpt.optimizer.momentum,
wd=pOpt.optimizer.wd,
learning_rate=current_lr,
lr_scheduler=lr_scheduler,
rescale_grad=1.0 / (len(ctx) * kv.num_workers),
clip_gradient=pOpt.optimizer.clip_gradient)
if pKv.fp16:
optimizer_params['multi_precision'] = True
optimizer_params['rescale_grad'] /= 128.0
profile = pGen.profile or False
if profile:
mx.profiler.set_config(profile_all=True,
filename=os.path.join(save_path,
"profile.json"))
# train
mod.fit(train_data=train_data,
eval_metric=eval_metrics,
epoch_end_callback=epoch_end_callback,
batch_end_callback=batch_end_callback,
kvstore=kv,
optimizer=pOpt.optimizer.type,
optimizer_params=optimizer_params,
initializer=init,
allow_missing=True,
arg_params=arg_params,
aux_params=aux_params,
begin_epoch=begin_epoch,
num_epoch=end_epoch,
profile=profile)
logging.info("Training has done")
time.sleep(10)
logging.info("Exiting")
