Denoise 3D
usage: denoise3d [-h] [-o OUTPUT] [--suffix SUFFIX] [-m MODEL]
[-a EVEN_TRAIN_PATH] [-b ODD_TRAIN_PATH] [--N-train N_TRAIN]
[--N-test N_TEST] [-c CROP]
[--base-kernel-width BASE_KERNEL_WIDTH]
[--optim {adam,adagrad,sgd}] [--lr LR] [--criteria {L1,L2}]
[--momentum MOMENTUM] [--batch-size BATCH_SIZE]
[--num-epochs NUM_EPOCHS] [-w WEIGHT_DECAY]
[--save-interval SAVE_INTERVAL] [--save-prefix SAVE_PREFIX]
[--num-workers NUM_WORKERS] [-j NUM_THREADS] [-g GAUSSIAN]
[-s PATCH_SIZE] [-p PATCH_PADDING] [-d DEVICE]
[volumes ...]
Positional Arguments
- volumes
volumes to denoise
Named Arguments
- -o, --output
directory to save denoised volumes
- --suffix
optional suffix to append to file paths. if not output is specfied, denoised volumes are written to the same location as the input with the suffix appended to the name (default .denoised)
- -m, --model
use pretrained denoising model. accepts path to a previously saved model or one of the provided pretrained models. pretrained model options are: unet-3d, unet-3d-10a, unet-3d-20a (default: unet-3d)
Default: “unet-3d”
- -a, --even-train-path
path to even training data
- -b, --odd-train-path
path to odd training data
- --N-train
Number of train points per volume (default: 1000)
Default: 1000
- --N-test
Number of test points per volume (default: 200)
Default: 200
- -c, --crop
training tile size (default: 96)
Default: 96
- --base-kernel-width
width of the base convolutional filter kernel in the U-net model (default: 11)
Default: 11
- --optim
Possible choices: adam, adagrad, sgd
optimizer (default: adagrad)
Default: “adagrad”
- --lr
learning rate for the optimizer (default: 0.001)
Default: 0.001
- --criteria
Possible choices: L1, L2
training criteria (default: L2)
Default: “L2”
- --momentum
momentum parameter for SGD optimizer (default: 0.8)
Default: 0.8
- --batch-size
minibatch size (default: 10)
Default: 10
- --num-epochs
number of training epochs (default: 500)
Default: 500
- -w, --weight_decay
L2 regularizer on the generative network (default: 0)
Default: 0
- --save-interval
save frequency in epochs (default: 10)
Default: 10
- --save-prefix
path prefix to save denoising model
- --num-workers
number of workers for dataloader (default: 1)
Default: 1
- -j, --num-threads
number of threads for pytorch, 0 uses pytorch defaults, <0 uses all cores (default: 0)
Default: 0
- -g, --gaussian
standard deviation of Gaussian filter postprocessing, 0 means no postprocessing (default: 0)
Default: 0
- -s, --patch-size
denoises volumes in patches of this size. not used if <1 (default: 96)
Default: 96
- -p, --patch-padding
padding around each patch to remove edge artifacts (default: 48)
Default: 48
- -d, --device
compute device/s to use (default: -2, multi gpu), set to >= 0 for single gpu, set to -1 for cpu
Default: -2