def main_sagemaker(args, cfg):
"""
Main training entry point for jobs launched via SageMaker
"""
instance_name = cfg.sagemaker_job['job_name']
s3_path = cfg.sagemaker_job['s3_path']
decompress_data() # setup data dirs based on SM CHANNELS
num_gpus = len(gpus)
# update configs according to CLI args
if args.work_dir is not None:
cfg.work_dir = args.work_dir
if args.resume_from is not None:
cfg.resume_from = args.resume_from
if args.autoscale_lr:
# apply the linear scaling rule (https://arxiv.org/abs/1706.02677)
total_bs = get_dist_info()[2] * cfg.data.imgs_per_gpu
cfg.optimizer[
'learning_rate'] = cfg.optimizer['learning_rate'] * total_bs / 8
# init distributed env first, since logger depends on the dist info.
init_dist()
if not gpus:
distributed = False # single node single gpu
else:
distributed = True
# create work_dir
mkdir_or_exist(osp.abspath(cfg.work_dir))
# init the logger before other steps
timestamp = time.strftime('%Y%m%d_%H%M%S', time.localtime())
log_file = osp.join(cfg.work_dir, '{}.log'.format(timestamp))
logger = get_root_logger(log_file=log_file, log_level=cfg.log_level)
# log some basic info
logger.info('Distributed training: {}'.format(distributed))
logger.info('TF MMDetection Version: {}'.format(__version__))
logger.info('Config:\n{}'.format(cfg.text))
logger.info('Tensorflow version: {}'.format(tf.version.VERSION))
# set random seeds
if args.seed is not None:
logger.info('Set random seed to {}, deterministic: {}'.format(
args.seed, args.deterministic))
set_random_seed(args.seed + get_dist_info()[0],
deterministic=args.deterministic)
model = build_detector(cfg.model,
train_cfg=cfg.train_cfg,
test_cfg=cfg.test_cfg)
# dummy data to init network
padded_img_side = max(cfg.data.train['scale'])
img = tf.random.uniform(shape=[padded_img_side, padded_img_side, 3],
dtype=tf.float32)
img_meta = tf.constant([
465., 640., 3., 800., 1101., 3.,
float(padded_img_side),
float(padded_img_side), 3., 1.7204301, 0.
],
dtype=tf.float32)
# bboxes = tf.constant([[1.0, 1.0, 10.0, 10.0]], dtype=tf.float32)
# labels = tf.constant([1], dtype=tf.int32)
_ = model((tf.expand_dims(img, axis=0), tf.expand_dims(img_meta, axis=0)),
training=False)
# print('BEFORE:', model.layers[0].layers[0].get_weights()[0][0,0,0,:])
# sagemaker specific path resolution
import os, pathlib
data_root = pathlib.Path(
os.getenv('SM_CHANNEL_COCO')).joinpath('coco').as_posix()
cfg.data.train['dataset_dir'] = data_root
cfg.data.val['dataset_dir'] = data_root
weights_file = cfg.model['backbone']['weights_path']
weights_path = pathlib.Path(
os.getenv('SM_CHANNEL_WEIGHTS')).joinpath(weights_file).as_posix()
logger.info('Loading weights from: {}'.format(weights_path))
if osp.splitext(weights_file
)[1] == '.h5': # older keras format from Keras model zoo
model.layers[0].layers[0].load_weights(weights_path,
by_name=True,
skip_mismatch=True)
else: # SavedModel format assumed - extract weights
backbone_model = tf.keras.models.load_model(weights_path)
# load weights if layers match
for layer_idx, layer in enumerate(backbone_model.layers):
if layer_idx < len(model.layers[0].layers[0].layers):
model.layers[0].layers[0].layers[layer_idx].set_weights(
layer.get_weights())
print('Loaded weights for:', layer.name)
del backbone_model
# print('AFTER:',model.layers[0].layers[0].get_weights()[0][0,0,0,:])
print_model_info(model, logger)
datasets = [build_dataset(cfg.data.train)]
if len(cfg.workflow) == 2:
datasets.append(build_dataset(cfg.data.val))
datasets = [build_dataset(cfg.data.train)]
if len(cfg.workflow) > 1:
raise NotImplementedError
train_detector(model,
datasets,
cfg,
num_gpus=num_gpus,
distributed=distributed,
mixed_precision=args.amp,
validate=args.validate,
timestamp=timestamp)
def main():
args = parse_args()
num_gpus = len(gpus)
cfg = Config.fromfile(args.config)
# update configs according to CLI args
if args.work_dir is not None:
cfg.work_dir = args.work_dir
if args.resume_from is not None:
cfg.resume_from = args.resume_from
if args.autoscale_lr:
# apply the linear scaling rule (https://arxiv.org/abs/1706.02677)
total_bs = len(gpus) * cfg.data.imgs_per_gpu
cfg.optimizer['learning_rate'] = \
cfg.optimizer['learning_rate'] * total_bs / 8
# init distributed env first, since logger depends on the dist info.
init_dist()
if not gpus:
distributed = False # single node single gpu
else:
distributed = True
# create work_dir
mkdir_or_exist(osp.abspath(cfg.work_dir))
# init the logger before other steps
timestamp = time.strftime('%Y%m%d_%H%M%S', time.localtime())
log_file = osp.join(cfg.work_dir, '{}.log'.format(timestamp))
logger = get_root_logger(log_file=log_file, log_level=cfg.log_level)
# log some basic info
logger.info('Distributed training: {}'.format(distributed))
logger.info('TF MMDetection Version: {}'.format(__version__))
logger.info('Config:\n{}'.format(cfg.text))
logger.info('Tensorflow version: {}'.format(tf.version.VERSION))
# set random seeds
if args.seed is not None:
logger.info('Set random seed to {}, deterministic: {}'.format(
args.seed, args.deterministic))
set_random_seed(args.seed + get_dist_info()[0], deterministic=args.deterministic)
model = build_detector(cfg.model,
train_cfg=cfg.train_cfg,
test_cfg=cfg.test_cfg)
# dummy data to init network
padded_img_side = max(cfg.data.train['scale'])
img = tf.random.uniform(shape=[padded_img_side, padded_img_side, 3], dtype=tf.float32)
img_meta = tf.constant(
[465., 640., 3., 800., 1101., 3., float(padded_img_side), float(padded_img_side), 3., 1.7204301, 0.],
dtype=tf.float32)
# bboxes = tf.constant([[1.0, 1.0, 10.0, 10.0]], dtype=tf.float32)
# labels = tf.constant([1], dtype=tf.int32)
_ = model((tf.expand_dims(img, axis=0), tf.expand_dims(img_meta, axis=0)),
training=False)
# print('BEFORE:', model.layers[0].layers[0].get_weights()[0][0,0,0,:])
weights_path = cfg.model['backbone']['weights_path']
logger.info('Loading weights from: {}'.format(weights_path))
model.layers[0].layers[0].load_weights(weights_path, by_name=True, skip_mismatch=True) #by_name=False)
# print('AFTER:',model.layers[0].layers[0].get_weights()[0][0,0,0,:])
print_model_info(model, logger)
datasets = [build_dataset(cfg.data.train)]
if len(cfg.workflow) == 2:
datasets.append(build_dataset(cfg.data.val))
datasets = [build_dataset(cfg.data.train)]
if len(cfg.workflow) > 1:
raise NotImplementedError
train_detector(model,
datasets,
cfg,
num_gpus=num_gpus,
distributed=distributed,
mixed_precision=args.amp,
validate=args.validate,
timestamp=timestamp)
)
from awsdet.datasets import build_dataset, build_dataloader
from awsdet.models import build_detector
gpus = tf.config.experimental.list_physical_devices('GPU')
##### TENSORFLOW RUNTIME OPTIONS #####
# tf.config.experimental_run_functions_eagerly(True)
os.environ['TF_CUDNN_USE_AUTOTUNE'] = str(0)
os.environ['TF_DETERMINISTIC_OPS'] = str(1)
os.environ['PYTHONHASHSEED'] = str(17)
os.environ['HOROVOD_FUSION_THRESHOLD'] = str(0)
# init distributed env first
init_dist()
# avoid large pool of Eigen threads
tf.config.threading.set_intra_op_parallelism_threads(5)
tf.config.threading.set_inter_op_parallelism_threads(
max(2, 40 // get_dist_info()[2]))
# reduce TF warning verbosity
os.environ['TF_CPP_MIN_LOG_LEVEL'] = str(2)
import logging
logging.getLogger('tensorflow').setLevel(logging.ERROR)
def parse_args():
def str2bool(v):
if isinstance(v, bool):
return v