in banding_removal/fastmri/args.py [0:0]
def __init__(self, **overrides):
super().__init__()
self.overrides = overrides
self.add_argument('--run_name', default="run", type=str)
self.add_argument('--seed', default=42, type=int)
self.add_argument('--debug', default=False, type=bool,
help="Additional debug logging/checks. Very slow.")
self.add_argument('--strace', default=False, type=bool,
help="Monitor stdout using strace, for detailed logging of distributed runs")
self.add_argument("--trainer_class", default="fastmri.trainer.Trainer", type=str)
self.add_argument("--architecture", default="public_unet.unet", type=str)
self.add_argument('--gan', dest='gan', action='store_true')
self.add_argument('--eval', dest='eval', action='store_true')
self.add_argument('--copy_data_to_scratch', dest='copy_data_to_scratch', action='store_true')
self.add_argument('--override_with_runinfo_args', type=Path, default=None,
help='If specified, override the args with the ones given in the provided runinfo file')
self.add_argument('--resize_type', default="crop", choices=["none", "crop", "pad", 'singlebatch'],
help='How to scale the images to the given resolution')
self.add_argument("--data_transform", default="rss.RSSDataTransform", type=str)
self.add_argument('--resolution_width', default=320, type=int, help='Resolution width of images')
self.add_argument('--resolution_height', default=320, type=int, help='Resolution height of images')
self.add_argument('--grappa_input', type=bool, default=False, help='Should you use Grappa for input')
self.add_argument('--grappa_input_path', type=Path, default=None, help='Path to grappa kernels')
self.add_argument('--grappa_path', type=Path, default=None, help='Path to grappa kernels')
self.add_argument('--grappa_target', type=bool, default=False, help='Should you use 2x grappa for the ground truth')
self.add_argument('--grappa_target_path', type=Path, default=None, help='Path to grappa kernels')
self.add_argument('--apply_grappa', type=bool, default=True, help='Should you use Grappa in the kspace model')
self.add_argument('--display_ifft', dest="display_ifft", action='store_true')
self.add_argument('--challenge', choices=['singlecoil', 'multicoil'], default="multicoil",
help='Which challenge')
self.add_argument('--magnet', type=int, default=None)
self.add_argument('--data_path', type=Path, default="/datasets01_101/fastMRI/112718",
help='Path to challenge dataset')
self.add_argument('--sample_rate', type=float, default=1.,
help='Fraction of total volumes to include')
self.add_argument('--calculate_offsets_directly', type=bool, default=False,
help="Ignore acq start/end info in metadata and just calculate directly")
self.add_argument('--min_kspace_width', type=int, default=None)
self.add_argument('--max_kspace_width', type=int, default=None)
self.add_argument('--min_kspace_height', type=int, default=None)
self.add_argument('--max_kspace_height', type=int, default=None)
self.add_argument('--min_target_width', type=int, default=None)
self.add_argument('--min_target_height', type=int, default=None)
self.add_argument('--max_target_width', type=int, default=None)
self.add_argument('--max_target_height', type=int, default=None)
self.add_argument('--start_slice', type=int, default=None)
self.add_argument('--end_slice', type=int, default=None)
self.add_argument('--only_square_targets', type=bool, default=False)
self.add_argument('--filter_acceleration', type=int, default=None, help="Filter input data that has accelerated by the given amount")
self.add_argument('--max_num_coils', type=int, default=None)
self.add_argument('--min_num_coils', type=int, default=None)
self.add_argument('--scale_inputs', type=bool, default=True)
self.add_argument('--scale_type', type=str, default='mean')
self.add_argument('--acquisition_types', nargs='*', default=None)
self.add_argument('--acquisition_systems', nargs='*', default=None, help="Choices: ['Avanto', 'TrioTim', 'Skyra', 'Aera', 'Biograph_mMR', 'Prisma_fit']")
self.add_argument('--coil_compress_coils', type=int, default=None)
# var net args
self.add_argument('--method-str', help='The main model is build from this string in var_net.var_net')
self.add_argument('--sens-method-str', help='The sensitivty maps model is build from this string in var_net.var_net')
self.add_argument('--norm', choices=('layer', 'instance', 'group'), default='group', help='Normalization of the unets in var_net.var_net')
self.add_argument('--norm-type', choices=('layer', 'instance', 'group'), default='group')
self.add_argument('--norm-mean', type=int, default=1)
self.add_argument('--norm-std', type=int, default=1)
self.add_argument('--kernel-size', type=int, default=3, help='Kernel size in convolutions in var_net.var_net')
self.add_argument('--sqrt-eps', type=float, default=0, help='Number to add before calling .sqrt() in var_net.var_net')
self.add_argument('--var_net_model', type=str, default='unet', help='The UNet model used in var net')
self.add_argument('--compute_sensitivities', type=bool, default=False)
# Mask parameters
self.add_argument('--accelerations', nargs='+', default=[4], type=int,
help='Ratio of k-space columns to be sampled. If multiple values are '
'provided, then one of those is chosen uniformly at random for '
'each volume.')
self.add_argument('--num_low_frequencies', nargs='+', default=[28], type=int,
help='Number of low-frequency k-space columns to be sampled. Should '
'have the same length as accelerations')
self.add_argument('--train_accelerations', nargs='+', default=[4], type=int,
help='Equivalent to --accelerations but for train set')
self.add_argument('--train_num_low_frequencies', nargs='+', default=[28], type=int,
help='Equivalent to --num_low_frequencies but for train st')
self.add_argument('--mask_type', default="equispaced", choices=["equispaced", "random", "random_fraction", "magic"],
help='The strategy used to mask k-space inputs')
#### Data augmentations
self.add_argument('--transforms_on_gpu', default=False, type=bool,
help="Do data transforms on the gpu when possible")
self.add_argument('--padding_augmentation', default=0, type=int,
help="Number of levels of padding augmentation to randomly sample from")
self.add_argument('--rotation_augmentation', default=False, type=bool)
self.add_argument('--elastic_augmentation', default=False, type=bool)
self.add_argument('--orientation_augmentation', default=False, type=bool)
self.add_argument('--orientation_augmentation_dev', default=False, type=bool)
self.add_argument('--debug_phase_direction', default=False, type=bool, help="For debugging only")
self.add_argument('--add_gibbs_artifacts_augmentation', default=False, type=bool,
help="Add false Gibbs artifacts to the image")
self.add_argument('--resize_min_width', default=None, type=int,
help="Input is resized in image space to if less wide than this")
self.add_argument('--resize_max_width', default=None, type=int,
help="Input is resized in image space to if wider than this")
self.add_argument('--batch_size', default=20, type=int, help="Per-gpu batch-size")
self.add_argument('--eval_batch_size', default=-1, type=int,
help="Larger batches can be used during eval as less memory is needed")
self.add_argument('--workers', default=8, type=int, help="Data loader worker count per GPU")
self.add_argument('--pin_memory', default=False, type=bool,
help="Pin tensors in dataloader. Can cause issues with distributed training")
self.add_argument('--is_distributed', default=False, type=bool, help="Distributed training flag (set automatically)")
self.add_argument('--use_barriers', default=True, type=bool,
help="During distributed training, keep processes in sync using barriers")
self.add_argument('--sync_params', default=False, type=bool,
help="Sync paramters between machines each epoch to reduce drift, and diagnose initialization bugs")
self.add_argument('--rank', default=0, type=int,
help="Distributed rank (0 for single). Set by environment variable automatically")
self.add_argument('--world_size', default=1, type=int,
help="Distributed world_size (1 for single). Set by environment variable automatically")
self.add_argument('--apex_distributed', default=False, type=bool, help="Use Apex for distributed training")
self.add_argument('--apex', default=False, type=bool, help="NVIDIA Apex half-prec training")
self.add_argument('--apex_amp', default=False, type=bool)
self.add_argument('--apex_loss_scale', default=1e4, type=float,
help="Prevent underflow by scaling the loss internally up by this amount")
self.add_argument('--nan_detection', default=False, type=bool,
help="Use pytorch's NaN detection mode which is slower but extremely useful for debugging")
self.add_argument('--log_interval', default=10, type=int)
self.add_argument('--save_info', default=True, type=bool)
self.add_argument('--save_model', default=True, type=bool)
self.add_argument('--display_count', default=16, type=int,
help="How many images to save out every epoch for display")
self.add_argument('--visual_first_epoch', default=True, type=bool,
help="save image grid visual of untrained model before training begins")
self.add_argument('--method', default='rmsprop', type=str)
self.add_argument('--lr', default=0.001, type=float)
self.add_argument('--momentum', default=0.9, type=float)
self.add_argument('--beta2', default=None, type=float)
self.add_argument('--adam_eps', default=1e-8, type=float)
self.add_argument('--decay', default=0.0, type=float)
self.add_argument('--lr_reduction', default="every40", type=str)
self.add_argument('--parameter_groups', default=False, type=bool,
help="Split scalar and vector parameters into separate parameter groups")
self.add_argument('--bias_lr_scale', default=0.1, type=float,
help="Scale the learning rate of all scalar/vector model parameters by this amount if --parameter_groups is set")
self.add_argument('--ramp_lr_by', default=1, type=int,
help="Start learning rate x times smaller and ramp up") # Not implemented yet
self.add_argument('--epochs', default=50, type=int)
self.add_argument('--eval_at_start', default=False, type=bool,
help="Perform DEV set evaluation at the beginning of training")
self.add_argument('--debug_epoch', default=False, type=bool,
help="Only process one batch each epoch for debugging purposes")
self.add_argument('--debug_epoch_stats', default=False, type=bool,
help="Only process one batch each stats (eval) epoch for debugging purposes")
self.add_argument('--break_early', default=None, type=float,
help="Percentage to break each epoch at")
self.add_argument('--debug_memory', default=False, type=bool, help='Output memory diagnostics then quit')
self.add_argument('--channels', default=2, type=int)
# UNET Settings
self.add_argument('--num_chans', type=int, default=128, help='Number of U-Net channels')
self.add_argument('--res_chans', type=int, default=128, help='Number of U-Net channels')
self.add_argument('--num_pools', type=int, default=4, help='Number of U-Net pooling layers')
self.add_argument('--drop_prob', type=float, default=0.0, help='Dropout probability')
self.add_argument('--num_models', type=int, default=4, help='Number of models for cascaded models')
self.add_argument('--dilation', type=int, default=1, help='Conv dilation')
self.add_argument('--num_layers', type=int, default=4, help='Num of conv layers')
self.add_argument('--groups', type=int, default=1, help='Num of connections between inputs and outputs')
# Initialization
self.add_argument('--smart_initialization', default=False, type=bool)
self.add_argument('--initialization', default="fan_out", type=str)
self.add_argument('--dropout', default=False, type=bool)
self.add_argument('--exp_dir', type=Path, default=Path.cwd() / "logs" / "run",
help='Path where model and results should be saved')
# Dicom
self.add_argument('--dicom', default=False, type=bool,
help='Use dicom dataset rather than challenge')
self.add_argument('--dicom_root', type=Path, default="/checkpoint/jzb/data/mmap",
help='Path to dicom dataset')
self.add_argument('--dicom_normalization', default="volume", choices=["volume", "instance"],
help='Normalize dicom images by volume or instance mean and std.')
self.add_argument('--filter_dicom_scan_type', default=False, type=bool,
help='Filter dicom dataset type by the "cor" scan type. These images are closer '
'to the challenge dataset.')
# Perecptual loss
self.add_argument('--perceptual', type=bool, default=False,
help='Use perceptual loss while training')
self.add_argument('--perceptual_loss_normalize', type=bool, default=False,
help="Prenormalize inputs to the perceptual loss to be mean 0 variance 1")
self.add_argument('--perceptual_loss_architecture', type=str,
default='vgg_bw_perceptual.vgg19bw_features',
help='Base architecture of the perceptual loss net')
self.add_argument('--perceptual_loss_kwargs', type=dict, default={},
help='kwargs passed to the perceptual loss architecture')
self.add_argument('--perceptual_loss_checkpoint', type=Path,
default='/checkpoint/mikerabbat/fast_mri/perceptual/pretrained_models/vgg19bw/checkpoint.pt')
self.add_argument('--perceptual_loss_cutoff_layer', type=str,
default='relu2_2',
help='Layer of the base architecture to use as '
'features for perceptual loss')
self.add_argument('--perceptual_l1_coefficient', type=float,
default=0.1,
help='Weight multiplying L1 loss added to perceptual')
self.add_argument('--ssim_loss', type=bool, default=False,
help='Use SSIM loss while training')
self.add_argument('--ssim_l1_coefficient', type=float, default=0.0,
help='Weight multiplying L1 loss added to SSIM loss')
self.add_argument('--q_loss', type=str, default=None,
help='Select a q weighted loss')
self.add_argument('--loss_denominator_power', default=1.0, type=float,
help="q weight for the q_loss, changing relative weighting of low vs high contrast regions")
self.add_argument('--loss_kernel', type=float, default=1.5,
help='Gaussian kernel size in pixel stds used for loss calculations')
self.add_argument('--gradient_loss', type=bool, default=False,
help="Penalize differences in the average magnitude of the gradient")
# DicomWriter
self.add_argument('--model_checkpoint', default=None, type=Path,
help='Path to the model checkpoint to use for reconstruction'
'(Used by DicomWriter)')
self.add_argument('--dicom_save_dir', default=None, type=Path,
help='Path to the directory where dicoms should be saved')
self.add_argument('--save_ground_truth', default=False, type=bool,
help='Whether to save DICOMs of ground truth images (default=False)')
self.add_argument('--reconstruct_all', default=False, type=bool,
help='Save DICOMs of reconstructions for every test case.'
'Default is to only save the six common test cases'
'used in the image review sessions.')
self.add_argument('--noise_levels', nargs='*', default=[], type=float,
help='List of noise levels used when producing DICOMS')
self.add_argument('--series_number', default=0, type=int,
help='Initial series number (useful when presenting '
'multiple methods side-by-side)')
self.add_argument('--series_description', default='', type=str,
help='Descriptive name for this series')
### Cluster
self.add_argument('--checkpoint_type', default="none", choices=["resume", "restart", "none"],
help='Resume (keeping all runinfo) or restart (keeping only model weights) from a previous '
'model checkpoint. "--checkpoint" should be set with this')
self.add_argument('--checkpoint', type=str,
help='Path to an existing checkpoint. Used along with "--resume"')
self.add_argument('--auto_requeue', default=False, type=bool,
help='If job is killed by slurm, reschedule it')
##############################
### Direct zeronet parameters
self.add_argument('--first_layer_planes', default=128, type=int)
self.add_argument('--cascades', default=3, type=int)
self.add_argument('--blocks_middle', default=4, type=int)
self.add_argument('--blocks_inner_middle', default=2, type=int)
self.add_argument('--blocks_inner', default=2, type=int)
self.add_argument('--blocks_outer', default=1, type=int)
self.add_argument('--use_fixed_conv', default=False, type=bool)
self.add_argument('--bottle_neck_factor', default=1, type=int, help="resnet block expansion factor")
self.add_argument('--groupnorm', default=True, type=bool)
self.add_argument('--use_fixed_conv_block', default=False, type=bool)
self.add_argument('--block_type', default='bottleneck', type=str)
self.add_argument('--tall_convs', default=False, type=bool)
self.add_argument('--autoregressive_sample_rate', default=0.01, type=float)
self.add_argument('--ar_channels', default=256, type=int)
self.add_argument('--box_sample', default=True, type=bool,
help="Do AR sampling simultanously for each box, with the boxes tiling the image")
self.add_argument('--boxes_per_dim', default=10, type=int,
help="Tile image with the square of this many boxes. Must evenly divide height/width")
#### Autocalibration
self.add_argument('--autocal_subsample', default=False, type=bool,
help="Train autocalibration model on subsampled k-space or not")
self.add_argument('--autocal_soft_normalization', default=False, type=bool)
## Direct reconstruction
self.add_argument('--whiten_coils', default=True, type=bool,
help="Pre-whiten all coils for the direct transforms")
self.add_argument('--sensitivity_target', default=True, type=bool,
help="use SENSE-REDUCE based target instead of RSS for the direct transforms")
self.add_argument('--imagespace_projection', default=False, type=bool,
help="Use imagespace projection instead of fourier based projection")
self.add_argument('--rss_output', default=False, type=bool,
help="Use RSS instead of SENSE based output on the last layer of the cascade")
self.add_argument('--gradient_checkpointing', default=False, type=bool,
help="Greatly reduce memory using checkpointing. Requires support by the model used.")
### Architecture search
self.add_argument('--arch_seed', default=None, type=int,
help="Optionanlly fix architecture every iteration with the archsearch model, for debugging")
self.add_argument('--arch_eval_size', default=10, type=int,
help="Size of subset of the training data used to choose best architecture (per gpu)")
self.add_argument('--arch_eval_nseeds', default=1000, type=int,
help="Number of seeds to consider during post-run search for best architecture")
#### Orientation adversary
self.add_argument('--orientation_save_images', default=None, type=str,
help="For one-off use, save out pngs for debugging")
self.add_argument('--orientation_save_images_from', default=None, type=int,
help="skip this number of initial batches")
self.add_argument('--orientation_adversary', default=False, type=bool,
help="Use the orientation detection adversary")
self.add_argument('--number_of_adversaries', default=1, type=int,
help="Potentially use an ensemble of multiple adversaries")
self.add_argument('--adversary_lr_scale', default=1.0, type=float,
help="The learning rate for the adversary is this multiple of the main learning rate")
self.add_argument('--adversary_epoch_from', default=0, type=int,
help="Start training adversary at a certain epoch")
self.add_argument('--warmup_adversary_from', default=0, type=int,
help="Start training adversary from this epoch, potentially from an earlier epoch than "
"when it's used to regulate the predictor via adversary_epoch_from")
self.add_argument('--adversary_strength', default=1.0, type=float,
help='strength of adversary_mixin adv training in loss')
self.add_argument('--reg_param', default=0.0, type=float,
help="Regularize the adversary's gradient norm")
self.add_argument('--adversary_weight_decay', default=0.0, type=float,
help="Apply this weight decay to the adversary (decay param effects predictor only)")
self.add_argument('--adv_target_uncertain', default=False, type=bool,
help="Train predictor to encourage 0.5 prop output from adv instead of 0/1")
self.add_argument('--adversary_model', default="unpooled_resnet50", type=str)
self.add_argument('--dont_learn_predictor', default=False, type=bool,
help="Only learn adversary for debugging purposes")
# multi-slice paarams
self.add_argument('--before_slices', default=0, type=int,
help="number of slices to grab for prediction before the slice being predicted")
self.add_argument('--after_slices', default=0, type=int,
help="number of slices to grab for prediction after the slice being predicted")
# Override defaults with passed overrides
self.set_defaults(**overrides)
## Some run specific context that we want globally accessible and saved out
## at the end of the run.
self.set_defaults(
main_pid = os.getpid(),
cwd = os.getcwd()
)