usage: run.py [-h]
              [--model {unet,hrnet,setr,unet3p,segnet,hardnet,deeplab,pspnet,danet,eanet,ocrnet,resunet,ocnet,attunet,dense,dlink,ternaus,scseunet,r2,r2att,csnet,unetpp,unetppbig,multires,u2net,u2netp,onenet,lightnet,cenet,setr,hardalter,lrfea,simple}]
              [--pretrain] [--checkpoint CHECKPOINT] --dataset
              {xh,xh_hard,zs_big,zs_small,pancreas} --mode
              {train,test,predict,debug} [--load LOAD] [--aug] [--cuda]
              [--stn] [-o {Adam,SGD}] [-e EPOCHS] [-b BATCH_SIZE] [-l LR]
              [-c CLASS_NUM] [--loss LOSS] [-w BCE_WEIGHT] [-r RESULT_DIR]
              [--poly]

optional arguments:
  -h, --help            show this help message and exit
  --model {unet,hrnet,setr,unet3p,segnet,hardnet,deeplab,pspnet,danet,eanet,ocrnet,resunet,ocnet,attunet,dense,dlink,ternaus,scseunet,r2,r2att,csnet,unetpp,unetppbig,multires,u2net,u2netp,onenet,lightnet,cenet,setr,hardalter,lrfea,simple}
                        choose the model
  --pretrain            whether to use pretrained weights
  --checkpoint CHECKPOINT
                        the location of the pretrained weights
  --dataset {xh,xh_hard,zs_big,zs_small,pancreas}
                        choose a dataset
  --mode {train,test,predict,debug}
                        select a mode
  --load LOAD           the location of the model weights for testing
  --aug                 whether to use color augmentation
  --cuda                whether to use CUDA acceleration
  --stn                 whether to use spatial transformer network
  -o {Adam,SGD}, --optimizer {Adam,SGD}
                        select an optimizer
  -e EPOCHS, --epochs EPOCHS
                        num of training epochs
  -b BATCH_SIZE, --batch-size BATCH_SIZE
                        batch size for training
  -l LR, --learning-rate LR
                        learning rate
  -c CLASS_NUM, --class-num CLASS_NUM
                        pixel-wise classes
  --loss LOSS           Choose from 'ce', 'iou', 'dice', 'focal', if CE loss
                        is selected, you should use a `weight` parameter
  -w BCE_WEIGHT         use this weight if BCE loss is selected; if w is
                        given, then the weights for positive and negative
                        classes will be w and 2.0 - w respectively
  -r RESULT_DIR, --resultdir RESULT_DIR
                        test result output directory
  --poly                whether to use polynomial learning rate scheduler

usage: run_ssl.py [-h]
                  [--model {unet,hrnet,setr,unet3p,segnet,hardnet,deeplab,pspnet,danet,eanet,ocrnet,resunet,ocnet,attunet,dense,dlink,ternaus,scseunet,r2,r2att,csnet,unetpp,unetppbig,multires,u2net,u2netp,onenet,lightnet,cenet,setr,hardalter,lrfea,simple}]
                  --dataset {xh,xh_hard,zs_big,zs_small,pancreas}
                  [--save SAVE] [-e EPOCHS] [-c CLASS_NUM] [-b BATCH_SIZE]
                  [--channel EMBEDDING_CHANNEL] [--layer LAYER] [--lr LR]
                  [--pretrain]

optional arguments:
  -h, --help            show this help message and exit
  --model {unet,hrnet,setr,unet3p,segnet,hardnet,deeplab,pspnet,danet,eanet,ocrnet,resunet,ocnet,attunet,dense,dlink,ternaus,scseunet,r2,r2att,csnet,unetpp,unetppbig,multires,u2net,u2netp,onenet,lightnet,cenet,setr,hardalter,lrfea,simple}
                        choose a model network
  --dataset {xh,xh_hard,zs_big,zs_small,pancreas}
                        select a dataset
  --save SAVE           model weights save path
  -e EPOCHS, --epochs EPOCHS
                        number of training epochs
  -c CLASS_NUM, --class-num CLASS_NUM
                        class number
  -b BATCH_SIZE, --batch-size BATCH_SIZE
                        training batch size
  --channel EMBEDDING_CHANNEL
                        number of channels of embedded feature maps
  --layer LAYER         layer to extract features from
  --lr LR               learning rate
  --pretrain