trying to fix mem alloc of torch

Author: C-Achard

docs/res/welcome.rst

@@ -68,6 +68,7 @@ Then go into Plugins > napari-cellseg3d, and choose which tool to use.
 - **Train**:  This module allows you to train segmentation algorithms from labeled volumes.
 - **Utilities**: This module allows you to use several utilities, e.g. to crop your volumes and labels, compute prediction scores or convert labels
 
+See above for links to detailed guides regarding the usage of the modules.
 
 Credits & acknowledgments
 ---------------------------------------------

src/napari_cellseg3d/model_workers.py

@@ -184,6 +184,9 @@ def inference(self):
 
         images_dict = self.create_inference_dict(self.images_filepaths)
 
+        if self.device =="cuda": # TODO : fix mem alloc
+            torch.backends.cudnn.benchmark = False
+
         # TODO : better solution than loading first image always ?
         data_check = LoadImaged(keys=["image"])(images_dict[0])
         # print(data)
@@ -284,6 +287,8 @@ def inference(self):
                 )
 
                 out = outputs.detach().cpu()
+                # del outputs # TODO fix memory ?
+                # outputs = None
 
                 if self.transforms["zoom"][0]:
                     zoom = self.transforms["zoom"][1]

src/napari_cellseg3d/plugin_model_inference.py

@@ -107,7 +107,7 @@ def __init__(self, viewer: "napari.viewer.Viewer"):
         )
 
         self.aniso_box_widgets = ui.make_n_spinboxes(
-            n=3, min=1.0, default=1.5, step=0.5, double=True
+            n=3, min=1.0, max=1000, default=1.5, step=0.5, double=True
         )
         self.aniso_box_lbl = [
             QLabel("Resolution in " + axis + " (microns) :") for axis in "xyz"

src/napari_cellseg3d/plugin_model_training.py

@@ -706,7 +706,7 @@ def start(self):
 
         TODO:
 
-        * Fix memory leak
+        * Fix memory allocation from torch
 
 
         Returns: Returns empty immediately if the file paths are not set correctly.

src/napari_cellseg3d/utils.py

@@ -146,7 +146,7 @@ def get_padding_dim(image_shape, anisotropy_factor=None):
         pad = -1
         size = image_shape[i]
         if anisotropy_factor is not None:
-            # TODO : GOING TO CAUSE ISSUES WITH CERTAIN ANISOTROPY FACTORS
+            # problems with zero divs avoided via params for spinboxes
             size = int(size / anisotropy_factor[i])
         while pad < size:
             pad = 2**n