def __init__(self, resnet_size_choices=None):
super(ResnetArgParser, self).__init__(parents=[
parsers.BaseParser(),
parsers.PerformanceParser(),
parsers.ImageModelParser(),
parsers.ExportParser(),
parsers.BenchmarkParser(),
])
self.add_argument(
'--version',
'-v',
type=int,
choices=[1, 2],
default=resnet_model.DEFAULT_VERSION,
help='Version of ResNet. (1 or 2) See README.md for details.')
self.add_argument(
'--resnet_size',
'-rs',
type=int,
default=50,
choices=resnet_size_choices,
help='[default: %(default)s] The size of the ResNet model to use.',
metavar='<RS>' if resnet_size_choices is None else None)
def __init__(self):
super(TestParser, self).__init__(parents=[
parsers.BaseParser(multi_gpu=True, num_gpu=False),
parsers.PerformanceParser(num_parallel_calls=True, inter_op=True,
intra_op=True, use_synthetic_data=True),
parsers.ImageModelParser(data_format=True),
parsers.BenchmarkParser(benchmark_log_dir=True, bigquery_uploader=True)
])
def __init__(self):
super(MNISTArgParser, self).__init__(parents=[
parsers.BaseParser(multi_gpu=True, num_gpu=False),
parsers.ImageModelParser(),
parsers.ExportParser(),
])
self.set_defaults(data_dir=DATA_DIR,
model_dir=MODEL_DIR,
export_dir=EXPORT_DIR,
batch_size=100,
train_epochs=1)
def __init__(self):
super(ArgParser, self).__init__(
parents=[parsers.BaseParser(),
parsers.ImageModelParser()])
# self.add_argument(
# '--export_dir',
# type=str)
self.add_argument('--data_root', type=str)
self.add_argument('--image_size', type=int)
self.set_defaults(
# data_dir='./data',
model_dir=os.path.join(sys.path[0], 'fire', 'model'),
export_dir=os.path.join(sys.path[0], 'fire', 'export'),
data_root=os.path.join(sys.path[0], 'fire', 'data', 'tfrecords'),
image_size=28,
batch_size=5,
train_epochs=1,
data_format='channels_first')
def __init__(self, resnet_size_choices=None):
super(ResnetArgParser, self).__init__(parents=[
parsers.BaseParser(multi_gpu=False),
parsers.PerformanceParser(num_parallel_calls=False),
parsers.ImageModelParser(),
parsers.ExportParser(),
parsers.BenchmarkParser(),
])
self.add_argument(
'--version',
'-v',
type=int,
choices=[1, 2, 14, 24, 34],
default=resnet_model.DEFAULT_VERSION,
help=
'Version of ResNet. (1, 2, 14 or 24, 34) See README.md for details.'
)
self.add_argument(
'--version_t',
'-vt',
type=int,
choices=[1, 2, 14],
default=resnet_model.DEFAULT_VERSION,
help=
'Version of ResNet Teacher. (1, 2 or 14) See README.md for details.'
)
self.add_argument(
'--resnet_size',
'-rs',
type=int,
default=50,
choices=resnet_size_choices,
help='[default: %(default)s] The size of the ResNet model to use.',
metavar='<RS>' if resnet_size_choices is None else None)
self.add_argument(
'--final_size',
'-fs',
type=int,
default=2048,
help='[default: %(default)s] The final size for dense layer.',
metavar='<FS>')
self.add_argument(
"--pickle_model",
"-pm",
default="./gap_save/gap_pruned.pkl",
help=
"[default: %(default)s] The location of the pruned model param file "
"files.",
metavar="<PM>",
)
self.add_argument(
"--random_init",
"-ri",
action='store_true',
help=
"[default: %(default)s] random_init: If True the gap fine-tune is from scratch."
)
self.add_argument(
"--enable_kd",
"-ek",
action='store_true',
help=
"[default: %(default)s] enable_kd: If True knowledge distillation is enabled."
)
self.add_argument(
'--kd_size',
'-ks',
type=int,
default=50,
choices=resnet_size_choices,
help=
'[default: %(default)s] The size of the ResNet model (teach) to use.',
metavar='<KS>' if resnet_size_choices is None else None)
self.add_argument(
"--temp_dst",
"-tdst",
type=float,
default=2.,
help=
'[default: %(default)s] temp_dst: temperature for knowledge distillation.',
metavar="<TDST>")
self.add_argument(
"--w_dst",
"-wdst",
type=float,
default=2.,
help=
'[default: %(default)s] w_dst: balance factor for knowledge distillation.',
metavar="<WDST>")
self.add_argument(
"--mix_up",
"-mu",
action='store_true',
help="[default: %(default)s] mix_up: If True mix_up is enabled.")
self.add_argument(
'--mx_mode',
'-mmd',
type=int,
default=0,
help=
'[default: %(default)s] the mode of mixup: randome or reverse.',
metavar='<MMD>')
self.add_argument(
"--enable_quantize",
"-eqz",
action='store_true',
help=
"[default: %(default)s] enable_quantize: If True quantization-aware training is enabled."
)
self.add_argument(
"--online_quantize",
"-oqz",
action='store_true',
help=
"[default: %(default)s] enable_quantize: If True online quantization-aware training is enabled."
)
self.add_argument('--q_bits',
'-qbt',
type=int,
default=8,
help='[default: %(default)s] quantization bits.',
metavar='<QBT>')
self.add_argument('--copy_num',
'-cnum',
type=int,
default=10,
help='[default: %(default)s] quantization copies.',
metavar='<CNUM>')
self.add_argument(
'--q_mode',
'-qme',
type=int,
default=1,
help='[default: %(default)s] quantization method (KL, MAX, Perc.).',
metavar='<QME>')
self.add_argument(
"--enable_at",
"-eat",
action='store_true',
help=
"[default: %(default)s] enable_at: If True attention transfer is enabled."
)
self.add_argument(
"--w_at",
"-wat",
type=float,
default=8.,
help=
'[default: %(default)s] w_at: balance factor for attention transfer.',
metavar="<WDST>")
self.add_argument(
"--oss_load",
"-osl",
action='store_true',
help=
"[default: %(default)s] oss_load: If True dataset is loaded from oss."
)
def __init__(self, resnet_size_choices=None):
super(ResnetArgParser, self).__init__(parents=[
parsers.BaseParser(),
parsers.PerformanceParser(),
parsers.ImageModelParser(),
])
self.add_argument(
'--version', '-v', type=int, choices=[1, 2],
default=resnet_model.DEFAULT_VERSION,
help="Version of ResNet. (1 or 2) See README.md for details."
)
self.add_argument(
'--resnet_size', '-rs', type=int, default=50,
choices=resnet_size_choices,
help="[default: %(default)s] The size of the ResNet model to use.",
metavar='<RS>' if resnet_size_choices is None else None
)
self.add_argument(
'--method', '-mt',
default=resnet_model.DEFAULT_METHOD,
help = "[default: %(default)s] The type of tensor decomposition used to compress the reference model."
)
self.add_argument(
'--scope', '-sc',
default=resnet_model.DEFAULT_SCOPE,
help = "[default: %(default)s] The variable scope used with the tensor decomposition method."
)
self.add_argument(
'--reference_model_dir', '-rd',
help = "[default: %(default)s] The location of the reference model.",
metavar = '<RD>'
)
self.add_argument(
'--reference_model_checkpoint_name', '-rdcn',
help = "[default: %(default)s] The name of the reference model checkpoint.",
metavar = '<RDC>'
)
self.add_argument(
'--output_model_dir', '-omd', default = '/tmp',
help = "[default: %(default)s] The location of the tensorized model.",
metavar = '<MD>'
)
self.add_argument(
'--rate', '-cr', default = 1, type=float,
help = "[default: %(default)d] The targeted compression rate of the tensorized model.",
metavar = '<CR>'
)
self.add_argument(
'--rate_decay', '-rdy', default = 'flat',
help = "[default: %(default)d] The rate decay function used for dynamic rate change.",
metavar = '<RDY>'
)