increase the learning rate of prostate

fixing a bug with infinite sampler

semi_seg/__init__.py

@@ -1,5 +1,5 @@
 acdc_ratios = [0.01, 0.015, 0.025, 1.0]
-prostate_ratio = [0.08, 0.1, 0.13, 0.18, 1.0]  # 3 5 7
+prostate_ratio = [0.08, 0.13, 0.18, 1.0]  # 3 5 7
 
 ratio_zoom = {"acdc": acdc_ratios,
               "prostate": prostate_ratio}

semi_seg/dsutils.py

@@ -38,7 +38,7 @@ def get_dataloaders(config, group_val_patient=True):
                      f"given {labeled_data_ratio} and {unlabeled_data_ratio}")
     data_manager = datainterface(root_dir=DATA_PATH, labeled_data_ratio=labeled_data_ratio,
                                  unlabeled_data_ratio=unlabeled_data_ratio, verbose=False,
-                                 seed=0 if dataset_name == "acdc" else 1)# avoid bad random seed for prostate
+                                 seed=0 if dataset_name == "acdc" else 10)# avoid bad random seed for prostate
 
     label_set, unlabel_set, val_set = data_manager._create_semi_supervised_datasets(  # noqa
         labeled_transform=augmentinferface.pretrain,

semi_seg/scripts/helper.py

@@ -12,7 +12,7 @@
     "mmwhs": 5,
 }
 ft_lr_zooms = {"acdc": 0.0000001,
-               "prostate": 0.00000025,
+               "prostate": 0.0000005,
                "spleen": 0.000001,
                "mmwhs": 0.000001}
 

semi_seg/trainers/_helper.py

@@ -2,6 +2,7 @@
 from pathlib import Path
 from typing import List, Dict, Any, Callable
 
+from deepclustering2.dataloader.sampler import InfiniteRandomSampler
 from deepclustering2.dataset import PatientSampler
 from deepclustering2.meters2 import StorageIncomeDict, Storage, EpochResultDict
 from deepclustering2.tqdm import item2str
@@ -29,18 +30,21 @@ def _get_contrastive_dataloader(partial_loader, config):
     dataset_name = config["Data"]["name"]
     batch_sampler = None
     batch_size = contrastive_config["group_sample_num"] * {"acdc": 3, "prostate": 7}[dataset_name]
+    sampler = InfiniteRandomSampler(dataset, shuffle=True)
+
     if dataset_name == "acdc":
         # only group the acdc dataset
         batch_sampler = ContrastBatchSampler(
             dataset=dataset,
             **contrastive_config
-        )
+        )  # this batch sampler is without end
         batch_size = 1
+        sampler = None
 
     contrastive_loader = DataLoader(
-        dataset, batch_sampler=batch_sampler,
+        dataset, batch_sampler=batch_sampler, sampler=sampler,
         num_workers=num_workers, batch_size=batch_size,
-        pin_memory=True, shuffle=False if batch_sampler else True,
+        pin_memory=True, shuffle=False,
     )
 
     from contrastyou.augment import ACDCStrongTransforms

semi_seg/transfer.sh

@@ -10,8 +10,8 @@ local_folder=./runs
 #  "${local_folder}/0402_semi_acdc/"
 
 rsync -azP --exclude "*/*.png"  --exclude "*/tra/*/*.pth" \
-  --exclude "*/patient*"  --exclude "*/*calculquebec.ca" \
-  beluga:/lustre04/scratch/jizong/Contrast-You/semi_seg/runs/0415_prostate \
+  --exclude "*/patient*"  \
+  [email protected]:/root/main/runs/0416_prostate \
   "${local_folder}"
 
 #rsync -azP --exclude "*/*.png"  --exclude "*/*.pth" \