def parser_(desc): parser = parser_def_mgpu(desc) remove_options(parser, ['--nccl', '--enqueue', '--syncopt', '--rdma']) parser.add_argument( '--checkpt', action='store', nargs='?', const=checkptfile, default=SUPPRESS, help='S|Save (overwrites) and load the model weights if available.' '\nOptionally specify a file/filepath if the default name is ' 'undesired.\n(default: {})'.format(checkptfile)) parser.add_argument('--aug', action='store_true', default=False, help='S|Perform data augmentation on cifar10 set.\n') parser.add_argument('--logdevp', action='store_true', default=False, help='S|Log device placement in Tensorflow.\n') parser.add_argument('--datadir', default=SUPPRESS, help='Data directory with Cifar10 dataset.') args = parser.parse_args() return args
def parser_(desc): parser = parser_def_mgpu(desc) checkptfile = 'cifar10_cnn_mgpu.weights.best.hdf5' parser.add_argument( '--checkpt', action='store', nargs='?', const=checkptfile, default=SUPPRESS, help='S|Save (overwrites) and load the model weights if available.' '\nOptionally specify a file/filepath if the default name is ' 'undesired.\n(default: {})'.format(checkptfile)) parser.add_argument('--aug', action='store_true', default=False, help='S|Perform data augmentation on cifar10 set.\n') parser.add_argument('--logdevp', action='store_true', default=False, help='S|Log device placement in Tensorflow.\n') args = parser.parse_args() return args
def parser_(desc): '''Parser for Cifar10 CNN Horovod training script.''' parser = parser_def_mgpu(desc) remove_options( parser, ['--nccl', '--enqueue', '--syncopt', '--rdma', '--mgpu', '--network']) parser.add_argument('--batch_size', type=int, default=CifarTrainDefaults.batch_size, help='S|Batch size. Default: %(default)s') parser.add_argument( '--nranks_per_gpu', type=int, default=1, help='S|Number of ranks to run on each GPUs. Use this parameter to\n' 'oversubscribe a GPU. When oversubscribing a GPU use in combination\n' 'with MPS (multi-process service). Default: %(default)s') checkptfile = 'cifar10_cnn_hvd.weights.best.hdf5' parser.add_argument( '--checkpt', action='store', nargs='?', const=checkptfile, help='S|Save (overwrites) and load the model weights if available.' '\nOptionally specify a file/filepath if the default name is ' 'undesired.\n(default: {})'.format(checkptfile)) parser.add_argument('--aug', action='store_true', default=False, help='S|Perform data augmentation on cifar10 set.\n') parser.add_argument('--logdevp', action='store_true', default=False, help='S|Log device placement in Tensorflow.\n') parser.add_argument( '--datadir', help='S|Data directory with Cifar10 dataset. Otherwise Keras\n' 'downloads to "<HOME>/.keras/datasets" directory by default.') parser.add_argument( '--use-dataset-api', action='store_true', default=False, help='S|Use Tensorflow Dataset API for Keras model training.') args = parser.parse_args() return args
def parser_(desc): '''CLI parser for Cifar10 multigpu example.''' parser = parser_def_mgpu(desc) remove_options(parser, ['--rdma', '--network']) parser.add_argument( '--batch_size', type=int, default=CifarTrainDefaults.batch_size, help='S|Batch size. Default: %(default)s') checkptfile = 'cifar10_cnn_mgpu.weights.best.hdf5' parser.add_argument( '--checkpt', action='store', nargs='?', const=checkptfile, help='S|Save (overwrites) and load the model weights if available.' '\nOptionally specify a file/filepath if the default name is ' 'undesired.\n(default: {})'.format(checkptfile)) parser.add_argument( '--mgpu-type', action='store', nargs='?', type=str.lower, const='expmgpu', default='expmgpu', choices=['expmgpu', 'kerasmgpu'], help='S|Use experimental or Keras multigpu conversion. For\n' 'experimental uses ModelMGPU and for Keras uses ModelKerasMGPU\n' 'which is a wrapper around multi_gpu_model function.\n' 'Default: expmgpu') parser.add_argument( '--syncopt', action='store_true', default=False, help='S|Use gradient synchronization in Optimizer. Not sure if this\n' 'feature is working correctly. Default: %(default)s') parser.add_argument( '--aug', action='store_true', default=False, help='S|Perform data augmentation on cifar10 set.\n') parser.add_argument( '--logdevp', action='store_true', default=False, help='S|Log device placement in Tensorflow.\n') parser.add_argument( '--datadir', help='Data directory with Cifar10 dataset.') parser.add_argument( '--use-dataset-api', action='store_true', default=False, help='S|Use Tensorflow Dataset API for Keras model training.') args = parser.parse_args() return args
def parser_(desc): parser = parser_def_mgpu(desc) remove_options(parser, ['--nccl', '--enqueue', '--syncopt', '--rdma', '--mgpu']) parser.add_argument('--batch_size', type=int, default=32, help='S|Batch size. Default: %(default)s') parser.add_argument( '--nranks_per_gpu', type=int, default=1, help='S|Number of ranks to run on each GPUs. Use this parameter to\n' 'oversubscribe a GPU. When oversubscribing a GPU use in combination\n' 'with MPS (multi-process service). Default: %(default)s') parser.add_argument( '--checkpt', action='store', nargs='?', const=checkptfile, default=SUPPRESS, help='S|Save (overwrites) and load the model weights if available.' '\nOptionally specify a file/filepath if the default name is ' 'undesired.\n(default: {})'.format(checkptfile)) parser.add_argument('--aug', action='store_true', default=False, help='S|Perform data augmentation on cifar10 set.\n') parser.add_argument('--logdevp', action='store_true', default=False, help='S|Log device placement in Tensorflow.\n') parser.add_argument('--datadir', default=SUPPRESS, help='Data directory with Cifar10 dataset.') args = parser.parse_args() return args
def parser_(desc): parser = parser_def_mgpu(desc) remove_options(parser, ['--nccl', '--enqueue', '--syncopt', '--rdma', '--mgpu', '--network']) parser.add_argument( '--batch_size', type=int, default=32, help='S|Batch size. Default: %(default)s') # parser.add_argument( # '--ngpus_per_node', type=int, default=-1, # help='S|Number of GPUs per node. Default: Horovod local_size()') parser.add_argument( '--nranks_per_gpu', type=int, default=1, help='S|Number of ranks to run on each GPUs. Use this parameter to\n' 'oversubscribe a GPU. When oversubscribing a GPU use in combination\n' 'with MPS (multi-process service). Default: %(default)s') checkptfile = 'cifar10_cnn_hvd.weights.best.hdf5' parser.add_argument( '--checkpt', action='store', nargs='?', const=checkptfile, default=SUPPRESS, help='S|Save (overwrites) and load the model weights if available.' '\nOptionally specify a file/filepath if the default name is ' 'undesired.\n(default: {})'.format(checkptfile)) parser.add_argument('--aug', action='store_true', default=False, help='S|Perform data augmentation on cifar10 set.\n') parser.add_argument('--logdevp', action='store_true', default=False, help='S|Log device placement in Tensorflow.\n') parser.add_argument('--datadir', default=SUPPRESS, help='Data directory with Cifar10 dataset.') args = parser.parse_args() return args