def init_classifier_compression_arg_parser():
'''Common classifier-compression application command-line arguments.
'''
SUMMARY_CHOICES = ['sparsity', 'compute', 'model', 'modules', 'png', 'png_w_params']
parser = argparse.ArgumentParser(description='Distiller image classification model compression')
parser.add_argument('data', metavar='DIR', help='path to dataset')
parser.add_argument('--arch', '-a', metavar='ARCH', default='resnet18', type=lambda s: s.lower(),
choices=models.ALL_MODEL_NAMES,
help='model architecture: ' +
' | '.join(models.ALL_MODEL_NAMES) +
' (default: resnet18)')
parser.add_argument('-j', '--workers', default=4, type=int, metavar='N',
help='number of data loading workers (default: 4)')
parser.add_argument('--epochs', type=int, metavar='N', default=90,
help='number of total epochs to run (default: 90')
parser.add_argument('-b', '--batch-size', default=256, type=int,
metavar='N', help='mini-batch size (default: 256)')
optimizer_args = parser.add_argument_group('Optimizer arguments')
optimizer_args.add_argument('--lr', '--learning-rate', default=0.1,
type=float, metavar='LR', help='initial learning rate')
optimizer_args.add_argument('--momentum', default=0.9, type=float,
metavar='M', help='momentum')
optimizer_args.add_argument('--weight-decay', '--wd', default=1e-4, type=float,
metavar='W', help='weight decay (default: 1e-4)')
parser.add_argument('--print-freq', '-p', default=10, type=int,
metavar='N', help='print frequency (default: 10)')
parser.add_argument('--verbose', '-v', action='store_true', help='Emit debug log messages')
load_checkpoint_group = parser.add_argument_group('Resuming arguments')
load_checkpoint_group_exc = load_checkpoint_group.add_mutually_exclusive_group()
# TODO(barrh): args.deprecated_resume is deprecated since v0.3.1
load_checkpoint_group_exc.add_argument('--resume', dest='deprecated_resume', default='', type=str,
metavar='PATH', help=argparse.SUPPRESS)
load_checkpoint_group_exc.add_argument('--resume-from', dest='resumed_checkpoint_path', default='',
type=str, metavar='PATH',
help='path to latest checkpoint. Use to resume paused training session.')
load_checkpoint_group_exc.add_argument('--exp-load-weights-from', dest='load_model_path',
default='', type=str, metavar='PATH',
help='path to checkpoint to load weights from (excluding other fields) (experimental)')
load_checkpoint_group.add_argument('--pretrained', dest='pretrained', action='store_true',
help='use pre-trained model')
load_checkpoint_group.add_argument('--reset-optimizer', action='store_true',
help='Flag to override optimizer if resumed from checkpoint. This will reset epochs count.')
parser.add_argument('-e', '--evaluate', dest='evaluate', action='store_true',
help='evaluate model on test set')
parser.add_argument('--activation-stats', '--act-stats', nargs='+', metavar='PHASE', default=list(),
help='collect activation statistics on phases: train, valid, and/or test'
' (WARNING: this slows down training)')
parser.add_argument('--activation-histograms', '--act-hist',
type=distiller.utils.float_range_argparse_checker(exc_min=True),
metavar='PORTION_OF_TEST_SET',
help='Run the model in evaluation mode on the specified portion of the test dataset and '
'generate activation histograms. NOTE: This slows down evaluation significantly')
parser.add_argument('--masks-sparsity', dest='masks_sparsity', action='store_true', default=False,
help='print masks sparsity table at end of each epoch')
parser.add_argument('--param-hist', dest='log_params_histograms', action='store_true', default=False,
help='log the parameter tensors histograms to file '
'(WARNING: this can use significant disk space)')
parser.add_argument('--summary', type=lambda s: s.lower(), choices=SUMMARY_CHOICES, action='append',
help='print a summary of the model, and exit - options: | '.join(SUMMARY_CHOICES))
parser.add_argument('--export-onnx', action='store', nargs='?', type=str, const='model.onnx', default=None,
help='export model to ONNX format')
parser.add_argument('--compress', dest='compress', type=str, nargs='?', action='store',
help='configuration file for pruning the model (default is to use hard-coded schedule)')
parser.add_argument('--sense', dest='sensitivity', choices=['element', 'filter', 'channel'],
type=lambda s: s.lower(), help='test the sensitivity of layers to pruning')
parser.add_argument('--sense-range', dest='sensitivity_range', type=float, nargs=3, default=[0.0, 0.95, 0.05],
help='an optional parameter for sensitivity testing '
'providing the range of sparsities to test.\n'
'This is equivalent to creating sensitivities = np.arange(start, stop, step)')
parser.add_argument('--extras', default=None, type=str,
help='file with extra configuration information')
parser.add_argument('--deterministic', '--det', action='store_true',
help='Ensure deterministic execution for re-producible results.')
parser.add_argument('--seed', type=int, default=None,
help='seed the PRNG for CPU, CUDA, numpy, and Python')
parser.add_argument('--gpus', metavar='DEV_ID', default=None,
help='Comma-separated list of GPU device IDs to be used '
'(default is to use all available devices)')
parser.add_argument('--cpu', action='store_true', default=False,
help='Use CPU only. \n'
'Flag not set => uses GPUs according to the --gpus flag value.'
'Flag set => overrides the --gpus flag')
parser.add_argument('--name', '-n', metavar='NAME', default=None, help='Experiment name')
parser.add_argument('--out-dir', '-o', dest='output_dir', default='logs', help='Path to dump logs and checkpoints')
parser.add_argument('--validation-split', '--valid-size', '--vs', dest='validation_split',
type=float_range(exc_max=True), default=0.1,
help='Portion of training dataset to set aside for validation')
parser.add_argument('--effective-train-size', '--etrs', type=float_range(exc_min=True), default=1.,
help='Portion of training dataset to be used in each epoch. '
'NOTE: If --validation-split is set, then the value of this argument is applied '
'AFTER the train-validation split according to that argument')
parser.add_argument('--effective-valid-size', '--evs', type=float_range(exc_min=True), default=1.,
help='Portion of validation dataset to be used in each epoch. '
'NOTE: If --validation-split is set, then the value of this argument is applied '
'AFTER the train-validation split according to that argument')
parser.add_argument('--effective-test-size', '--etes', type=float_range(exc_min=True), default=1.,
help='Portion of test dataset to be used in each epoch')
parser.add_argument('--confusion', dest='display_confusion', default=False, action='store_true',
help='Display the confusion matrix')
parser.add_argument('--num-best-scores', dest='num_best_scores', default=1, type=int,
help='number of best scores to track and report (default: 1)')
parser.add_argument('--load-serialized', dest='load_serialized', action='store_true', default=False,
help='Load a model without DataParallel wrapping it')
parser.add_argument('--thinnify', dest='thinnify', action='store_true', default=False,
help='physically remove zero-filters and create a smaller model')
distiller.quantization.add_post_train_quant_args(parser)
return parser
def get_parser():
parser = argparse.ArgumentParser(
description='Distiller image classification model compression')
parser.add_argument('--arch',
'-a',
metavar='ARCH',
default='resnet152',
type=lambda s: s.lower(),
choices=models.ALL_MODEL_NAMES,
help='model architecture: ' +
' | '.join(models.ALL_MODEL_NAMES) +
' (default: resnet152)')
parser.add_argument('-j',
'--workers',
default=44,
type=int,
metavar='N',
help='number of data loading workers (default: 44)')
parser.add_argument('--epochs',
default=20,
type=int,
metavar='N',
help='number of total epochs to run')
parser.add_argument('-b',
'--batch-size',
default=16,
type=int,
metavar='N',
help='mini-batch size (default: 16)')
parser.add_argument('--lr',
'--learning-rate',
default=0.001,
type=float,
metavar='LR',
help='initial learning rate')
parser.add_argument('--momentum',
default=0.9,
type=float,
metavar='M',
help='momentum')
parser.add_argument('--weight-decay',
'--wd',
default=0e-5,
type=float,
metavar='W',
help='weight decay (default: 1e-4)')
parser.add_argument('--print-freq',
'-p',
default=500,
type=int,
metavar='N',
help='print frequency (default: 50)')
parser.add_argument('--resume',
default='',
type=str,
metavar='PATH',
help='path to latest checkpoint (default: none)')
parser.add_argument('-e',
'--evaluate',
dest='evaluate',
action='store_true',
help='evaluate model on validation set')
parser.add_argument('--pretrained',
dest='pretrained',
action='store_true',
help='use pre-trained model')
parser.add_argument(
'--activation-stats',
'--act-stats',
nargs='+',
metavar='PHASE',
default=list(),
help='collect activation statistics on phases: train, valid, and/or test'
' (WARNING: this slows down training)'
) # 收集阶段的激活统计信息：列车、有效和/或测试（警告：这会减慢训练速度）
parser.add_argument('--masks-sparsity',
dest='masks_sparsity',
action='store_true',
default=False,
help='print masks sparsity table at end of each epoch'
) # 打印掩盖稀疏表 在end of each epoch
parser.add_argument(
'--param-hist',
dest='log_params_histograms',
action='store_true',
default=False,
help=
'log the parameter tensors histograms to file (WARNING: this can use significant disk space)'
) # 将参数张量柱状图记录到文件中（警告：这可能会占用大量磁盘空间）
parser.add_argument(
'--summary',
type=lambda s: s.lower(),
choices=SUMMARY_CHOICES,
help='print a summary of the model, and exit - options: ' +
' | '.join(SUMMARY_CHOICES)) # 打印模型摘要，并退出-选项
parser.add_argument(
'--compress',
dest='compress',
type=str,
nargs='?',
action='store',
help=
'configuration file for pruning the model (default is to use hard-coded schedule)'
) # 用于修剪模型的配置文件（默认为使用硬编码计划）
parser.add_argument(
'--sense',
dest='sensitivity',
choices=['element', 'filter', 'channel'],
type=lambda s: s.lower(),
help='test the thinnify of layers to pruning') # 测试修剪层的敏感性
parser.add_argument(
'--sense-range',
dest='sensitivity_range',
type=float,
nargs=3,
default=[0.4, 0.9, 0.1],
help=
'an optional parameter for sensitivity testing providing the range of sparsities to test.\n'
'This is equivalent to creating sensitivities = np.arange(start, stop, step)'
) # 灵敏度测试的可选参数，提供要测试的稀疏度范围。这相当于创建灵敏度=np.arange（开始、停止、步骤）
parser.add_argument(
'--extras',
default=None,
type=str,
help='file with extra configuration information') # 包含额外配置信息的文件
parser.add_argument(
'--deterministic',
'--det',
action='store_true',
help='Ensure deterministic execution for re-producible results.'
) # 确保可重复生产结果的确定性执行。
parser.add_argument(
'--gpus',
metavar='DEV_ID',
default=None,
help=
'Comma-separated list of GPU device IDs to be used (default is to use all available devices)'
) # 要使用的GPU设备ID的逗号分隔列表（默认为使用所有可用设备）
parser.add_argument(
'--cpu',
action='store_true',
default=False,
help='Use CPU only. \n'
'Flag not set => uses GPUs according to the --gpus flag value.'
'Flag set => overrides the --gpus flag')
parser.add_argument('--name',
'-n',
metavar='NAME',
default=None,
help='Experiment name') # 实验名
parser.add_argument(
'--out-dir',
'-o',
dest='output_dir',
default='logs',
help='Path to dump logs and checkpoints') # 转储日志和检查点的路径
parser.add_argument(
'--validation-split',
'--valid-size',
'--vs',
dest='validation_split',
type=float_range(exc_max=True),
default=0.1,
help='Portion of training dataset to set aside for validation'
) # 用于验证的培训数据集的一部分
parser.add_argument(
'--effective-train-size',
'--etrs',
type=float_range(exc_min=True),
default=1.,
help='Portion of training dataset to be used in each epoch. '
'NOTE: If --validation-split is set, then the value of this argument is applied '
'AFTER the train-validation split according to that argument')
# 培训数据集的一部分将在每个时期使用。
# '注意：如果设置了--validation split，则应用此参数的值'
# '在根据该参数划分列车验证之后'
parser.add_argument(
'--effective-valid-size',
'--evs',
type=float_range(exc_min=True),
default=1.,
help='Portion of validation dataset to be used in each epoch. '
'NOTE: If --validation-split is set, then the value of this argument is applied '
'AFTER the train-validation split according to that argument')
# 在每个时期使用的验证数据集的一部分。
# '注意：如果设置了--validation split，则应用此参数的值'
# '根据该参数，在列车验证拆分后
parser.add_argument('--effective-test-size',
'--etes',
type=float_range(exc_min=True),
default=1.,
help='Portion of test dataset to be used in each epoch'
) # 每个时期使用的测试数据集的一部分
parser.add_argument('--confusion',
dest='display_confusion',
default=False,
action='store_true',
help='Display the confusion matrix') # 显示混淆矩阵
parser.add_argument(
'--earlyexit_lossweights',
type=float,
nargs='*',
dest='earlyexit_lossweights',
default=None,
help=
'List of loss weights for early exits (e.g. --earlyexit_lossweights 0.1 0.3)'
) # 提前退出的的损失权重列表（例如0.1 0.3）
parser.add_argument(
'--earlyexit_thresholds',
type=float,
nargs='*',
dest='earlyexit_thresholds',
default=None,
help='List of EarlyExit thresholds (e.g. --earlyexit_thresholds 1.2 0.9)'
) # 提前退出阈值列表（例如1.2 0.9）
parser.add_argument(
'--num-best-scores',
dest='num_best_scores',
default=1,
type=int,
help='number of best scores to track and report (default: 1)'
) # 要跟踪和报告的最佳分数数（默认值：1）
parser.add_argument(
'--load-serialized',
dest='load_serialized',
action='store_true',
default=False,
help='Load a model without DataParallel wrapping it') # 加载模型时不使用数据并行包装
parser.add_argument(
'--thinnify',
dest='thinnify',
action='store_true',
default=False,
help='physically remove zero-filters and create a smaller model'
) # 物理删除零过滤器并创建较小的模型
distiller.knowledge_distillation.add_distillation_args(
parser, models.ALL_MODEL_NAMES, True)
distiller.quantization.add_post_train_quant_args(parser)
distiller.pruning.greedy_filter_pruning.add_greedy_pruner_args(parser)
adc.automl_args.add_automl_args(parser)
return parser
def get_parser():
parser = argparse.ArgumentParser(
description='Distiller image classification model compression')
parser.add_argument('data', metavar='DIR', help='path to dataset')
parser.add_argument('--arch',
'-a',
metavar='ARCH',
default='resnet18',
type=lambda s: s.lower(),
choices=models.ALL_MODEL_NAMES,
help='model architecture: ' +
' | '.join(models.ALL_MODEL_NAMES) +
' (default: resnet18)')
parser.add_argument(
'--loaders',
type=int,
metavar='N',
help=
'number of data loading workers (default: max({}, {} per GPU). 1 if deterministic is set.)'
.format(DEFAULT_LOADERS_COUNT, DEFAULT_LOADERS_COUNT))
parser.add_argument('--epochs',
default=90,
type=int,
metavar='N',
help='number of total epochs to run')
parser.add_argument('-b',
'--batch-size',
default=256,
type=int,
metavar='N',
help='mini-batch size (default: 256)')
optimizer_args = parser.add_argument_group('optimizer_arguments')
optimizer_args.add_argument('--lr',
'--learning-rate',
default=0.1,
type=float,
metavar='LR',
help='initial learning rate')
optimizer_args.add_argument('--momentum',
default=0.9,
type=float,
metavar='M',
help='momentum')
optimizer_args.add_argument('--weight-decay',
'--wd',
default=1e-4,
type=float,
metavar='W',
help='weight decay (default: 1e-4)')
parser.add_argument(
'--reset-optimizer',
'--reset-lr',
action='store_true',
help='Flag to override optimizer if resumed from checkpoint')
print_freq_group = parser.add_mutually_exclusive_group()
print_freq_group.add_argument(
'--print-frequency',
type=int,
metavar='N',
help='print frequency (default: {} prints per epoch)'.format(
DEFAULT_PRINT_FREQUENCY))
print_freq_group.add_argument('--print-period',
type=int,
metavar='N',
help='print every N mini-batches')
load_checkpoint_group = parser.add_mutually_exclusive_group()
load_checkpoint_group.add_argument(
'--resume',
default='',
type=str,
metavar='PATH',
help='path to latest checkpoint (default: none)')
load_checkpoint_group.add_argument(
'--load-state-dict',
default='',
type=str,
metavar='PATH',
help='load only state dict field from checkpoint at given path')
parser.add_argument('-e',
'--evaluate',
dest='evaluate',
action='store_true',
help='evaluate model on validation set')
parser.add_argument('--pretrained',
dest='pretrained',
action='store_true',
help='use pre-trained model')
parser.add_argument(
'--activation-stats',
'--act-stats',
nargs='+',
metavar='PHASE',
default=list(),
help='collect activation statistics on phases: train, valid, and/or test'
' (WARNING: this slows down training)')
parser.add_argument('--masks-sparsity',
dest='masks_sparsity',
action='store_true',
default=False,
help='print masks sparsity table at end of each epoch')
parser.add_argument(
'--param-hist',
dest='log_params_histograms',
action='store_true',
default=False,
help=
'log the parameter tensors histograms to file (WARNING: this can use significant disk space)'
)
parser.add_argument(
'--summary',
type=lambda s: s.lower(),
choices=SUMMARY_CHOICES,
help='print a summary of the model, and exit - options: ' +
' | '.join(SUMMARY_CHOICES))
parser.add_argument(
'--compress',
dest='compress',
type=str,
nargs='?',
action='store',
help=
'configuration file for pruning the model (default is to use hard-coded schedule)'
)
parser.add_argument('--sense',
dest='sensitivity',
choices=['element', 'filter', 'channel'],
type=lambda s: s.lower(),
help='test the sensitivity of layers to pruning')
parser.add_argument(
'--sense-range',
dest='sensitivity_range',
type=float,
nargs=3,
default=[0.0, 0.95, 0.05],
help=
'an optional parameter for sensitivity testing providing the range of sparsities to test.\n'
'This is equivalent to creating sensitivities = np.arange(start, stop, step)'
)
parser.add_argument('--extras',
default=None,
type=str,
help='file with extra configuration information')
parser.add_argument(
'--deterministic',
'--det',
action='store_true',
help='Ensure deterministic execution for re-producible results.')
device_group = parser.add_mutually_exclusive_group()
device_group.add_argument(
'--gpus',
metavar='DEV_ID',
default=None,
help=
'Comma-separated list of GPU device IDs to be used (default: use all available devices)'
)
device_group.add_argument('--use-cpu',
action='store_true',
default=False,
help='Force use of CPU only')
parser.add_argument('--name',
'-n',
metavar='NAME',
default=None,
help='Experiment name')
parser.add_argument('--out-dir',
'-o',
dest='output_dir',
default='logs',
help='Path to dump logs and checkpoints')
parser.add_argument(
'--validation-split',
'--vs',
type=float_range(exc_max=True),
default=0,
metavar='FRACTION',
help='Portion of training dataset to set aside for validation')
parser.add_argument(
'--effective-train-size',
'--etrs',
type=float_range(exc_min=True),
default=1.,
help='Portion of training dataset to be used in each epoch. '
'NOTE: If --validation-split is set, then the value of this argument is applied '
'AFTER the train-validation split according to that argument')
parser.add_argument(
'--effective-valid-size',
'--evs',
type=float_range(exc_min=True),
default=1.,
help='Portion of validation dataset to be used in each epoch. '
'NOTE: If --validation-split is set, then the value of this argument is applied '
'AFTER the train-validation split according to that argument')
parser.add_argument(
'--effective-test-size',
'--etes',
type=float_range(exc_min=True),
default=1.,
help='Portion of test dataset to be used in each epoch')
parser.add_argument('--confusion',
dest='display_confusion',
default=False,
action='store_true',
help='Display the confusion matrix')
parser.add_argument(
'--earlyexit_lossweights',
type=float,
nargs='*',
dest='earlyexit_lossweights',
default=None,
help=
'List of loss weights for early exits (e.g. --earlyexit_lossweights 0.1 0.3)'
)
parser.add_argument(
'--earlyexit_thresholds',
type=float,
nargs='*',
dest='earlyexit_thresholds',
default=None,
help=
'List of EarlyExit thresholds (e.g. --earlyexit_thresholds 1.2 0.9)')
parser.add_argument(
'--num-best-scores',
dest='num_best_scores',
default=1,
type=int,
help='number of best scores to track and report (default: 1)')
parser.add_argument('--load-serialized',
dest='load_serialized',
action='store_true',
default=False,
help='Load a model without DataParallel wrapping it')
parser.add_argument(
'--thinnify',
dest='thinnify',
action='store_true',
default=False,
help='physically remove zero-filters and create a smaller model')
# deprecations
def deprecation_warning(*args, old_keys=None, new_keys=None, **kwargs):
if old_keys and new_keys:
msglogger.warning(
'{okey} have been deprecated. Try {nkey} instead.'.format(
okey=old_keys, nkey=new_keys))
elif old_keys:
msglogger.warning(
'{okey} have been deprecated.'.format(okey=old_keys))
else:
msglogger.warning(
'Some arguments have been deprecated and ignored.')
parser.add_argument('--valid-size',
'--validation-size',
type=functools.partial(
deprecation_warning,
old_keys=['--valid-size', '--validation-size'],
new_keys=['--validation-split', '--vs']),
help=argparse.SUPPRESS)
parser.add_argument('--print-freq',
'-p',
type=functools.partial(deprecation_warning,
old_keys=['--print-freq', '-p'],
new_keys=['--print-period']),
help=argparse.SUPPRESS)
parser.add_argument('-j',
'--workers',
type=functools.partial(deprecation_warning,
old_keys=['-j', '--workers'],
new_keys=['--loaders']),
help=argparse.SUPPRESS)
distiller.knowledge_distillation.add_distillation_args(
parser, models.ALL_MODEL_NAMES, True)
distiller.quantization.add_post_train_quant_args(parser)
distiller.pruning.greedy_filter_pruning.add_greedy_pruner_args(parser)
adc.automl_args.add_automl_args(parser)
return parser
def get_parser(model_names, dataset_names):
"""
Return the argument parser
"""
parser = argparse.ArgumentParser(description='Image classification model')
parser.add_argument('--device',
type=device,
default=84,
help='set device (default: AI84)')
parser.add_argument('--8-bit-mode',
'-8',
dest='act_mode_8bit',
action='store_true',
default=False,
help='simluate device operation (8-bit data)')
parser.add_argument('--arch',
'-a',
'--model',
metavar='ARCH',
required=True,
type=lambda s: s.lower(),
dest='cnn',
choices=model_names,
help='model architecture: ' + ' | '.join(model_names))
parser.add_argument('--dataset',
metavar='S',
required=True,
choices=dataset_names,
help="dataset(s) (" + ', '.join(dataset_names) + ")")
parser.add_argument('--truncate-testset',
action='store_true',
default=False,
help='get only the first image from the test set')
parser.add_argument('--data',
metavar='DIR',
default='data',
help='path to dataset')
parser.add_argument('-j',
'--workers',
default=4,
type=int,
metavar='N',
help='number of data loading workers (default: 4)')
parser.add_argument('--epochs',
type=int,
metavar='N',
help='number of total epochs to run (default: 90)')
parser.add_argument('-b',
'--batch-size',
default=256,
type=int,
metavar='N',
help='mini-batch size (default: 256)')
parser.add_argument('--kernel-stats',
action='store_true',
default=False,
help='compute kernel statistics')
parser.add_argument(
'--use-bias',
action='store_true',
default=False,
help='for models that support both bias and no bias, set the '
'`use bias` flag to true')
parser.add_argument(
'--avg-pool-rounding',
action='store_true',
default=False,
help='when simulating, use "round()" in AvgPool operations '
'(default: use "floor()")')
qat_args = parser.add_argument_group('Quantization Arguments')
qat_args.add_argument('--qat-policy',
dest='qat_policy',
default='qat_policy.yaml',
help='path to YAML file that defines the '
'QAT (quantization-aware training) policy')
optimizer_args = parser.add_argument_group('Optimizer Arguments')
optimizer_args.add_argument('--optimizer',
default='SGD',
help='optimizer for training (default: SGD)')
optimizer_args.add_argument('--lr',
'--learning-rate',
default=0.1,
type=float,
metavar='LR',
help='initial learning rate')
optimizer_args.add_argument('--momentum',
default=0.9,
type=float,
metavar='M',
help='momentum')
optimizer_args.add_argument('--weight-decay',
'--wd',
default=1e-4,
type=float,
metavar='W',
help='weight decay (default: 1e-4)')
parser.add_argument('--print-freq',
'-p',
default=10,
type=int,
metavar='N',
help='print frequency (default: 10)')
load_checkpoint_group = parser.add_argument_group('Resuming Arguments')
load_checkpoint_group_exc = load_checkpoint_group.add_mutually_exclusive_group(
)
load_checkpoint_group_exc.add_argument(
'--resume-from',
dest='resumed_checkpoint_path',
default='',
type=str,
metavar='PATH',
help='path to latest checkpoint. Use to resume paused '
'training session.')
load_checkpoint_group_exc.add_argument(
'--exp-load-weights-from',
dest='load_model_path',
default='',
type=str,
metavar='PATH',
help='path to checkpoint to load weights from '
'(excluding other fields) (experimental)')
load_checkpoint_group.add_argument('--pretrained',
dest='pretrained',
action='store_true',
help='use pre-trained model')
load_checkpoint_group.add_argument(
'--reset-optimizer',
action='store_true',
help='Flag to override optimizer if resumed from '
'checkpoint. This will reset epochs count.')
parser.add_argument('-e',
'--evaluate',
dest='evaluate',
action='store_true',
help='evaluate model on test set')
mgroup = parser.add_mutually_exclusive_group()
mgroup.add_argument(
'--save-csv',
dest='csv_prefix',
default=None,
type=str,
help='save as CSVs with the given prefix during evaluation')
mgroup.add_argument(
'--save-sample',
dest='generate_sample',
type=int,
help='save the sample at given index as NumPy sample data')
parser.add_argument(
'--shap',
default=0,
type=int,
help=
'select # of images from the test set and plot SHAP after evaluation')
parser.add_argument(
'--activation-stats',
'--act-stats',
nargs='+',
metavar='PHASE',
default=list(),
help='collect activation statistics on phases: train, valid, and/or test'
' (WARNING: this slows down training)')
parser.add_argument('--masks-sparsity',
dest='masks_sparsity',
action='store_true',
default=False,
help='print masks sparsity table at end of each epoch')
parser.add_argument(
'--param-hist',
dest='log_params_histograms',
action='store_true',
default=False,
help=
'log the parameter tensors histograms to file (WARNING: this can use '
'significant disk space)')
parser.add_argument(
'--summary',
type=lambda s: s.lower(),
choices=SUMMARY_CHOICES,
help='print a summary of the model, and exit - options: ' +
' | '.join(SUMMARY_CHOICES))
parser.add_argument(
'--summary-filename',
default='model',
help='file name (w/o extension) for the model summary (default: "model"'
)
parser.add_argument('--compress',
dest='compress',
type=str,
nargs='?',
action='store',
help='configuration file for pruning the model '
'(default is to use hard-coded schedule)')
parser.add_argument('--sense',
dest='sensitivity',
choices=['element', 'filter', 'channel'],
type=lambda s: s.lower(),
help='test the sensitivity of layers to pruning')
parser.add_argument(
'--sense-range',
dest='sensitivity_range',
type=float,
nargs=3,
default=[0.0, 0.95, 0.05],
help='an optional parameter for sensitivity testing providing the range '
'of sparsities to test.\n'
'This is equivalent to creating sensitivities = np.arange(start, '
'stop, step)')
parser.add_argument('--extras',
default=None,
type=str,
help='file with extra configuration information')
parser.add_argument(
'--deterministic',
'--det',
action='store_true',
help='Ensure deterministic execution for re-producible results.')
parser.add_argument('--seed',
type=int,
default=None,
help='seed the PRNG for CPU, CUDA, numpy, and Python')
parser.add_argument(
'--gpus',
metavar='DEV_ID',
default=None,
help='Comma-separated list of GPU device IDs to be used (default is to '
'use all available devices)')
parser.add_argument(
'--cpu',
action='store_true',
default=False,
help='Use CPU only. \n'
'Flag not set => uses GPUs according to the --gpus flag value.'
'Flag set => overrides the --gpus flag')
parser.add_argument('--name',
'-n',
metavar='NAME',
default=None,
help='Experiment name')
parser.add_argument('--out-dir',
'-o',
dest='output_dir',
default='logs',
help='Path to dump '
'logs and checkpoints')
parser.add_argument(
'--validation-split',
'--valid-size',
'--vs',
dest='validation_split',
type=float_range(exc_max=True),
default=0.1,
help='Portion of training dataset to set aside for validation')
parser.add_argument(
'--effective-train-size',
'--etrs',
type=float_range(exc_min=True),
default=1.,
help='Portion of training dataset to be used in each epoch. '
'NOTE: If --validation-split is set, then the value of this argument '
'is applied AFTER the train-validation split according to that '
'argument')
parser.add_argument(
'--effective-valid-size',
'--evs',
type=float_range(exc_min=True),
default=1.,
help='Portion of validation dataset to be used in each epoch. '
'NOTE: If --validation-split is set, then the value of this argument '
'is applied AFTER the train-validation split according to that '
'argument')
parser.add_argument(
'--effective-test-size',
'--etes',
type=float_range(exc_min=True),
default=1.,
help='Portion of test dataset to be used in each epoch')
parser.add_argument('--confusion',
dest='display_confusion',
default=False,
action='store_true',
help='Display the confusion matrix')
parser.add_argument('--embedding',
dest='display_embedding',
default=False,
action='store_true',
help='Display embedding (using projector)')
parser.add_argument('--pr-curves',
dest='display_prcurves',
default=False,
action='store_true',
help='Display the precision-recall curves')
parser.add_argument('--no-tensorboard',
dest='tblog',
default=True,
action='store_false',
help='Disable TensorBoard')
parser.add_argument('--regression',
dest='regression',
default=False,
action='store_true',
help='Force regression output')
parser.add_argument('--earlyexit_lossweights',
type=float,
nargs='*',
dest='earlyexit_lossweights',
default=None,
help='List of loss weights for early exits '
'(e.g. --earlyexit_lossweights 0.1 0.3)')
parser.add_argument(
'--earlyexit_thresholds',
type=float,
nargs='*',
dest='earlyexit_thresholds',
default=None,
help=
'List of EarlyExit thresholds (e.g. --earlyexit_thresholds 1.2 0.9)')
parser.add_argument(
'--num-best-scores',
dest='num_best_scores',
default=1,
type=int,
help='number of best scores to track and report (default: 1)')
parser.add_argument('--load-serialized',
dest='load_serialized',
action='store_true',
default=False,
help='Load a model without DataParallel wrapping it')
parser.add_argument(
'--thinnify',
dest='thinnify',
action='store_true',
default=False,
help='physically remove zero-filters and create a smaller model')
parser.add_argument(
'--sparsity-perf',
action='store_true',
default=False,
help='when determining best epoch, use sparsity as primary key')
distiller.knowledge_distillation.add_distillation_args(
parser, model_names, True)
distiller.quantization.add_post_train_quant_args(parser)
distiller.pruning.greedy_filter_pruning.add_greedy_pruner_args(parser)
adc.add_automl_args(parser)
return parser
def get_parser():
parser = argparse.ArgumentParser(
description='Distiller image classification model compression')
parser.add_argument('data', metavar='DIR', help='path to dataset')
parser.add_argument('--arch',
'-a',
metavar='ARCH',
default='resnet18',
type=lambda s: s.lower(),
choices=models.ALL_MODEL_NAMES,
help='model architecture: ' +
' | '.join(models.ALL_MODEL_NAMES) +
' (default: resnet18)')
parser.add_argument('-j',
'--workers',
default=4,
type=int,
metavar='N',
help='number of data loading workers (default: 4)')
parser.add_argument('--epochs',
default=90,
type=int,
metavar='N',
help='number of total epochs to run')
parser.add_argument('-b',
'--batch-size',
default=256,
type=int,
metavar='N',
help='mini-batch size (default: 256)')
parser.add_argument('--lr',
'--learning-rate',
default=0.1,
type=float,
metavar='LR',
help='initial learning rate')
parser.add_argument('--momentum',
default=0.9,
type=float,
metavar='M',
help='momentum')
parser.add_argument('--weight-decay',
'--wd',
default=1e-4,
type=float,
metavar='W',
help='weight decay (default: 1e-4)')
parser.add_argument('--print-freq',
'-p',
default=10,
type=int,
metavar='N',
help='print frequency (default: 10)')
parser.add_argument('--resume',
default='',
type=str,
metavar='PATH',
help='path to latest checkpoint (default: none)')
parser.add_argument('-e',
'--evaluate',
dest='evaluate',
action='store_true',
help='evaluate model on validation set')
parser.add_argument('--pretrained',
dest='pretrained',
action='store_true',
help='use pre-trained model')
parser.add_argument(
'--activation-stats',
'--act-stats',
nargs='+',
metavar='PHASE',
default=list(),
help='collect activation statistics on phases: train, valid, and/or test'
' (WARNING: this slows down training)')
parser.add_argument('--masks-sparsity',
dest='masks_sparsity',
action='store_true',
default=False,
help='print masks sparsity table at end of each epoch')
parser.add_argument(
'--param-hist',
dest='log_params_histograms',
action='store_true',
default=False,
help=
'log the parameter tensors histograms to file (WARNING: this can use significant disk space)'
)
parser.add_argument(
'--summary',
type=lambda s: s.lower(),
choices=SUMMARY_CHOICES,
help='print a summary of the model, and exit - options: ' +
' | '.join(SUMMARY_CHOICES))
parser.add_argument(
'--compress',
dest='compress',
type=str,
nargs='?',
action='store',
help=
'configuration file for pruning the model (default is to use hard-coded schedule)'
)
parser.add_argument('--sense',
dest='sensitivity',
choices=['element', 'filter', 'channel'],
type=lambda s: s.lower(),
help='test the sensitivity of layers to pruning')
parser.add_argument(
'--sense-range',
dest='sensitivity_range',
type=float,
nargs=3,
default=[0.0, 0.95, 0.05],
help=
'an optional parameter for sensitivity testing providing the range of sparsities to test.\n'
'This is equivalent to creating sensitivities = np.arange(start, stop, step)'
)
parser.add_argument('--extras',
default=None,
type=str,
help='file with extra configuration information')
parser.add_argument(
'--deterministic',
'--det',
action='store_true',
help='Ensure deterministic execution for re-producible results.')
parser.add_argument(
'--gpus',
metavar='DEV_ID',
default=None,
help=
'Comma-separated list of GPU device IDs to be used (default is to use all available devices)'
)
parser.add_argument(
'--cpu',
action='store_true',
default=False,
help='Use CPU only. \n'
'Flag not set => uses GPUs according to the --gpus flag value.'
'Flag set => overrides the --gpus flag')
parser.add_argument('--name',
'-n',
metavar='NAME',
default=None,
help='Experiment name')
parser.add_argument('--out-dir',
'-o',
dest='output_dir',
default='logs',
help='Path to dump logs and checkpoints')
parser.add_argument(
'--validation-split',
'--valid-size',
'--vs',
dest='validation_split',
type=float_range(exc_max=True),
default=0.1,
help='Portion of training dataset to set aside for validation')
parser.add_argument(
'--effective-train-size',
'--etrs',
type=float_range(exc_min=True),
default=1.,
help='Portion of training dataset to be used in each epoch. '
'NOTE: If --validation-split is set, then the value of this argument is applied '
'AFTER the train-validation split according to that argument')
parser.add_argument(
'--effective-valid-size',
'--evs',
type=float_range(exc_min=True),
default=1.,
help='Portion of validation dataset to be used in each epoch. '
'NOTE: If --validation-split is set, then the value of this argument is applied '
'AFTER the train-validation split according to that argument')
parser.add_argument(
'--effective-test-size',
'--etes',
type=float_range(exc_min=True),
default=1.,
help='Portion of test dataset to be used in each epoch')
parser.add_argument('--adc',
dest='ADC',
action='store_true',
help='temp HACK')
parser.add_argument('--adc-params',
dest='ADC_params',
default=None,
help='temp HACK')
parser.add_argument('--confusion',
dest='display_confusion',
default=False,
action='store_true',
help='Display the confusion matrix')
parser.add_argument(
'--earlyexit_lossweights',
type=float,
nargs='*',
dest='earlyexit_lossweights',
default=None,
help=
'List of loss weights for early exits (e.g. --earlyexit_lossweights 0.1 0.3)'
)
parser.add_argument(
'--earlyexit_thresholds',
type=float,
nargs='*',
dest='earlyexit_thresholds',
default=None,
help=
'List of EarlyExit thresholds (e.g. --earlyexit_thresholds 1.2 0.9)')
parser.add_argument(
'--num-best-scores',
dest='num_best_scores',
default=1,
type=int,
help='number of best scores to track and report (default: 1)')
parser.add_argument('--load-serialized',
dest='load_serialized',
action='store_true',
default=False,
help='Load a model without DataParallel wrapping it')
parser.add_argument(
'--thinnify',
dest='thinnify',
action='store_true',
default=False,
help='physically remove zero-filters and create a smaller model')
distiller.knowledge_distillation.add_distillation_args(
parser, models.ALL_MODEL_NAMES, True)
distiller.quantization.add_post_train_quant_args(parser)
return parser
