Set default val_split to 0.2 for bundle trainer (#1118)

* set default val_split to 0.2 for bundle trainer

Signed-off-by: tangy5 <[email protected]>

* set default val_split to 0.2 for bundle trainer

Signed-off-by: tangy5 <[email protected]>

Signed-off-by: tangy5 <[email protected]>
Co-authored-by: SACHIDANAND ALLE <[email protected]>

Author: tangy5

monailabel/tasks/train/bundle.py

@@ -112,7 +112,7 @@ def _partition_datalist(self, datalist, request, shuffle=False):
         datalist = [{"image": d["image"], "label": d["label"]} for d in datalist if d]
         logger.info(f"Total Records in Dataset: {len(datalist)}")
 
-        val_split = request.get("val_split", 0.0)
+        val_split = request.get("val_split", 0.2)
         if val_split > 0.0:
             train_datalist, val_datalist = partition_dataset(
                 datalist, ratios=[(1 - val_split), val_split], shuffle=shuffle

sample-apps/endoscopy/README.md

@@ -118,12 +118,17 @@ export MONAI_LABEL_DATASTORE_PASSWORD=mypass
 monailabel start_server \
   --app workspace/endoscopy \
   --studies workspace/images \
+  --conf models tooltracking \
   --conf epistemic_enabled true \
-  --conf epistemic_top_k 3 \
+  --conf epistemic_top_k 10 \
   --conf auto_finetune_models tooltracking \
   --conf auto_finetune_check_interval 30
 ```
 
+Note: the argument ```epistemic_top_k``` specifies the number of images to label in each iteration, provide a number above 5 if there is no externally provided validation data
+MONAI Label server will automatically split annotated data into train and validation by percentage. No external validation data and ```epistemic_top_k``` < 5 will lead to traning with no
+saved model.
+
 #### Fetch and Publish latest model to CVAT/Nuclio
 After re-train the fine-tuned model meets all the conditions to be considered as a good model.  You can push the model to cvat/nuclio function container.
 ```bash