Update inference.json

481-Rotation of output images in prostate_mri_anatomy

- The transformations applied to the test data during preprocessing are not reversed during postprocessing. Thus, the resulting predictions have a different rotation compared to the original images and labels. I added an 'invertd()' function that reverses all the traceable transformations in the preprocessing.
- Running the inference on a PC without a GPU leads to an error because the handlers function 'CheckpointLoader' attempts to load the model onto the GPU. I set the map_location argument to torch.device(device) so that it automatically loads the model onto the CPU when a GPU is not available.
- Additionally, I changed some of the config directories as they were referring to the validation dataset rather than the test dataset.

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

models/prostate_mri_anatomy/configs/inference.json

@@ -5,8 +5,8 @@
     ],
     "bundle_root": "/workspace/data/prostate_mri_anatomy",
     "output_dir": "$@bundle_root + '/eval'",
-    "dataset_dir": "/workspace/data/prostate158/prostate158_train/",
-    "datalist": "$list(@dataset_dir + pd.read_csv(@dataset_dir + 'valid.csv').t2)",
+    "dataset_dir": "/workspace/data/prostate158/prostate158_test/",
+    "datalist": "$list(@dataset_dir + pd.read_csv(@dataset_dir + 'test.csv').t2)",
     "device": "$torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')",
     "network_def": {
         "_target_": "UNet",
@@ -114,6 +114,15 @@
                     2
                 ]
             },
+            {
+                "_target_": "Invertd",
+                "keys": "pred",
+                "transform": "@preprocessing",
+                "orig_keys": "image",
+                "meta_key_postfix": "meta_dict",
+                "nearest_interp": false,
+                "to_tensor": true
+            },
             {
                 "_target_": "SaveImaged",
                 "keys": "pred",
@@ -127,6 +136,10 @@
         {
             "_target_": "CheckpointLoader",
             "load_path": "$@bundle_root + '/models/model.pt'",
+            "map_location": {
+                "_target_": "torch.device",
+                "device": "@device"
+        },
             "load_dict": {
                 "model": "@network"
             }