fix wrapper constructor in ErrorAnalysisManager and RAIVisionInsights…

… load call for RAI Vision Dashboard (#2560)
microsoft · May 1, 2024 · 93eacce · 93eacce
1 parent 6026326
commit 93eacce
Show file tree

Hide file tree

Showing 4 changed files with 52 additions and 7 deletions.
diff --git a/responsibleai_vision/responsibleai_vision/managers/error_analysis_manager.py b/responsibleai_vision/responsibleai_vision/managers/error_analysis_manager.py
@@ -316,13 +316,21 @@ def _load(path, rai_insights):
         feature_names = list(dataset.columns)
         inst.__dict__['_feature_names'] = feature_names
         task_type = rai_insights.task_type
-        wrapped_model = wrap_model(rai_insights.model, dataset,
-                                   task_type,
-                                   classes=rai_insights._classes,
-                                   device=rai_insights.device)
+        classes = rai_insights._classes
+        device = rai_insights.device
+
+        test = rai_insights.test
+        image_mode = rai_insights.image_mode
+        transformations = rai_insights._transformations
+        sample = test.iloc[0:2]
+        sample = get_images(sample, image_mode, transformations)
+        wrapped_model = wrap_model(
+            rai_insights.model, sample, task_type, classes=classes,
+            device=device)
+
         inst.__dict__['_task_type'] = task_type
-        index_classes = rai_insights._classes
-        index_dataset = rai_insights.test
+        index_classes = classes
+        index_dataset = test
         if isinstance(target_column, list):
             # create copy of dataset as we will make modifications to it
             index_dataset = index_dataset.copy()

diff --git a/responsibleai_vision/responsibleai_vision/rai_vision_insights/rai_vision_insights.py b/responsibleai_vision/responsibleai_vision/rai_vision_insights/rai_vision_insights.py
@@ -1095,7 +1095,9 @@ def load(path):
         # load current state
         RAIBaseInsights._load(
             path, inst, manager_map, RAIVisionInsights._load_metadata)
-        inst._wrapped_model = wrap_model(inst.model, inst.test, inst.task_type,
+        sample = inst.test.iloc[0:2]
+        sample = get_images(sample, inst.image_mode, inst._transformations)
+        inst._wrapped_model = wrap_model(inst.model, sample, inst.task_type,
                                          classes=inst._classes,
                                          device=inst.device)
         inst.automl_image_model = is_automl_image_model(inst._wrapped_model)

diff --git a/responsibleai_vision/tests/common_vision_utils.py b/responsibleai_vision/tests/common_vision_utils.py
@@ -164,6 +164,15 @@ def create_dummy_model(df):
     return DummyFlowersClassifier()
 
 
+def create_raw_torchvision_classification_model():
+    """Creates a dummy torchvision model for testing purposes.
+
+    :return: dummy torchvision model
+    :rtype: torchvision.models.resnet.ResNet
+    """
+    return torchvision_models.vgg16(pretrained=False, num_classes=2)
+
+
 def retrieve_unzip_file(download_url, data_file):
     fetch_dataset(download_url, data_file)
     # extract files
@@ -486,6 +495,14 @@ def _get_model_path(self, path):
         return os.path.join(path, 'image-classification-model')
 
 
+class TorchvisionDummyPipelineSerializer(object):
+    def save(self, model, path):
+        pass
+
+    def load(self, path):
+        return create_raw_torchvision_classification_model()
+
+
 class ObjectDetectionPipelineSerializer(object):
     def save(self, model, path):
         pass

diff --git a/responsibleai_vision/tests/test_rai_vision_insights_save_and_load_scenarios.py b/responsibleai_vision/tests/test_rai_vision_insights_save_and_load_scenarios.py
@@ -12,8 +12,10 @@
 from common_vision_utils import (DummyFlowersPipelineSerializer,
                                  ImageClassificationPipelineSerializer,
                                  ObjectDetectionPipelineSerializer,
+                                 TorchvisionDummyPipelineSerializer,
                                  create_dummy_model,
                                  create_image_classification_pipeline,
+                                 create_raw_torchvision_classification_model,
                                  load_flowers_dataset,
                                  load_fridge_object_detection_dataset,
                                  load_imagenet_dataset, load_imagenet_labels,
@@ -49,6 +51,22 @@ def test_rai_insights_empty_save_load_save(self):
         run_and_validate_serialization(
             pred, test, task_type, class_names, label, serializer)
 
+    def test_rai_insights_pytorch_empty_save_load_save(self):
+        data = load_flowers_dataset(upscale=False)
+        data = data[0:1]
+        # stack two of the same image since we need same
+        # image sizes for pytorch model
+        data = data.append(data).reset_index(drop=True)
+        pred = create_raw_torchvision_classification_model()
+        test = data
+        class_names = data[ImageColumns.LABEL.value].unique()
+        task_type = ModelTask.IMAGE_CLASSIFICATION
+        label = ImageColumns.LABEL
+        serializer = TorchvisionDummyPipelineSerializer()
+
+        run_and_validate_serialization(
+            pred, test, task_type, class_names, label, serializer)
+
     @pytest.mark.skip("Insufficient memory on test machines to load images")
     def test_rai_insights_large_images_save_load_save(self):
         PIL.Image.MAX_IMAGE_PIXELS = None