A PyTorch wrapper for unbalanced training/classification on image rois
usage: pytorchunicas.py [-h] <command> ...
optional arguments:
-h, --help show this help message and exit
subcommands:
<command> description
train Train or finetune a model
test Score a model
crossvalid Cross-validation (includes multiple iterations of training and
testing)
Example of usage for the crossvalid
modality (includes both training and testing phases using k-fold cross-validation):
usage: pytorchunicas.py crossvalid [-h] --model
{CAEP16,CAEP23L2,CAEP23L2F2,CAEP23L3,CAEP48,CAE_MNIST,VAE_MNIST,VGGnetP12L2,VGGnetP16L2,VGGnetP16L2BN,VGGnetP23L2,VGGnetP23L3}
--roi_files ROI_FILES --img_folder
IMG_FOLDER --workspace WORKSPACE
[-b BATCH_SIZE] [--gpu GPU]
[--img_prefix IMG_PREFIX]
[--img_suffix IMG_SUFFIX]
[--img_list IMG_LIST]
[--img_channel IMG_CHANNEL]
[--max_epochs MAX_EPOCHS]
[--class_weights CLASS_WEIGHTS]
[--class_max_counts CLASS_MAX_COUNTS]
[--class_min_counts CLASS_MIN_COUNTS]
[--preprocessing {ImageStandardization,MinMaxNormalization,NoPreprocessing,PixelStandardization}]
[--augmentation {FlipRotate,Replicate}]
[--optimizer {ASGD,Adadelta,Adagrad,Adam,Adamax,LBFGS,RMSprop,Rprop,SGD,SparseAdam}]
[--lr_base LR_BASE]
[--lr_stepsize LR_STEPSIZE]
[--lr_gamma LR_GAMMA] [--momentum MOMENTUM]
[--weight_decay WEIGHT_DECAY]
[--display DISPLAY]
[--loss_weights LOSS_WEIGHTS] [--resume]
[--start_epoch N] [-j NUM_WORKERS]
[--no_reshuffle] [--data_stats]
[--cross_valid_N CROSS_VALID_N]
[--fold_start FOLD_START]
[--fold_end FOLD_END] [--skip_training]
[--skip_testing]
optional arguments:
-h, --help show this help message and exit
-b BATCH_SIZE, --batch_size BATCH_SIZE
mini-batch size (default: 32)
--gpu GPU GPU id to use.
--img_prefix IMG_PREFIX
Prefix to be added to all the images. (default: )
--img_suffix IMG_SUFFIX
Suffix to be added to all the images. (default: )
--img_list IMG_LIST Image inclusion list (only load samples if they belong
to the images contained in this list).
--img_channel IMG_CHANNEL
Image channel selection according to BGR color space
indexing (B = 0, G = 1, R = 2).
--max_epochs MAX_EPOCHS
Maximum number of epochs to train or test. In the
latter case, this determines the model to load.
(default: 30)
--class_weights CLASS_WEIGHTS
Class weights determining the class sampling
probability at training time.
--class_max_counts CLASS_MAX_COUNTS
Max number of samples that can be loaded, for each
class.
--class_min_counts CLASS_MIN_COUNTS
Min number of samples that can be loaded, for each
class.
--preprocessing {ImageStandardization,MinMaxNormalization,NoPreprocessing,PixelStandardization}
Data preprocessing. (default: PixelStandardization)
--augmentation {FlipRotate,Replicate}
Data augmentation. (default: FlipRotate)
--optimizer {ASGD,Adadelta,Adagrad,Adam,Adamax,LBFGS,RMSprop,Rprop,SGD,SparseAdam}
Optimizer. (default: SGD)
--lr_base LR_BASE Base learning rate. (default: 0.001)
--lr_stepsize LR_STEPSIZE
Period of learning rate decay (in epochs). (default:
6)
--lr_gamma LR_GAMMA Multiplicative factor of learning rate decay.
(default: 0.1)
--momentum MOMENTUM Momentum factor. (default: 0.9)
--weight_decay WEIGHT_DECAY
Weight decay (L2 penalty). (default: 0)
--display DISPLAY The number of epochs between displaying info.
(default: 0.1)
--loss_weights LOSS_WEIGHTS
Class weights assigned when computing the loss
function (weights < 0 will be assigned automatically)
--resume Resume training from last epoch model. (default:
False)
--start_epoch N Manual epoch number (useful on restarts) (default: 0)
-j NUM_WORKERS, --num_workers NUM_WORKERS
Number of workers for mini batch loading (default: 2)
--no_reshuffle Disables rehuffling of training samples at the
beginning of each epoch. (default: False)
--data_stats Print data statistics (min, max, mean, std) during
training (default: False)
--cross_valid_N CROSS_VALID_N
Number of cross-validation folds. (default: 2)
--fold_start FOLD_START
Fold start (default: 1)
--fold_end FOLD_END Fold end (default: -1)
--skip_training Skip training phase. (default: False)
--skip_testing Skip testing phase. (default: False)
required named arguments:
--model {CAEP16,CAEP23L2,CAEP23L2F2,CAEP23L3,CAEP48,CAE_MNIST,VAE_MNIST,VGGnetP12L2,VGGnetP16L2,VGGnetP16L2BN,VGGnetP23L2,VGGnetP23L3}
Model architecture. (default: None)
--roi_files ROI_FILES
Roi files separated by ',' one for each class.
--img_folder IMG_FOLDER
Directory with all the images.
--workspace WORKSPACE
Folder containing input and output files.