efficient ad tutorial fixes (#1187)

tutorial fixes
add anomaly det to readme

tutorials/mct_model_garden/evaluation_metrics/anomaly_eval.py

@@ -31,15 +31,15 @@ def benchmark(unified_model, name, q_st_start=None, q_st_end=None, q_ae_start=No
     auc = test(test_set=test_set, unified_model=unified_model, test_output_dir=test_output_dir, q_st_start=None, q_st_end=None, q_ae_start=None, q_ae_end=None, desc='Final inference')
     print('Final image auc: {:.4f}'.format(auc))
 
-def test(test_set, unified_model, test_output_dir=None, desc='Running inference'):
+def test(test_set, unified_model, test_output_dir=None, q_st_start=None, q_st_end=None, q_ae_start=None, q_ae_end=None, desc='Running inference'):
     """Test the model and calculate the AUC score."""
     y_true, y_score = [], []
     for image, target, path in tqdm(test_set, desc=desc):
         orig_width, orig_height = image.size
         image = DEFAULT_TRANSFORM(image)[None]  # Add batch dimension
         if torch.cuda.is_available():
             image = image.cuda()
-        map_combined = predict_combined(image, unified_model, q_st_start=None, q_st_end=None, q_ae_start=None, q_ae_end=None)
+        map_combined , _, _ = predict_combined(image, unified_model, q_st_start=None, q_st_end=None, q_ae_start=None, q_ae_end=None)
         map_combined = torch.nn.functional.interpolate(map_combined, (orig_height, orig_width), mode='bilinear')
         map_combined = map_combined[0, 0].detach().cpu().numpy()
         defect_class = os.path.basename(os.path.dirname(path))

tutorials/notebooks/imx500_notebooks/README.md

@@ -172,6 +172,16 @@ deployment performance.
         <td>50.4</td>
         <td>47.1</td>
     </tr>
+    <tr>
+        <td>Anomaly Detection</td>
+        <td>Efficient AD</td>
+        <td> <a href="pytorch/pytorch_efficient_anomaly_detection.ipynb">PyTorch</a></td>
+        <td><a href="https://arxiv.org/pdf/2303.08730v3">Ultralytics</a></td>
+        <td><a href="https://huggingface.co/SSI-DNN/Efficient_Anomaly_Detection">mct-model-garden</a></td>
+        <td>MvTech</td>
+        <td>98.56</td>
+        <td>97.85</td>
+    </tr>
 
 </table>
 

tutorials/notebooks/imx500_notebooks/pytorch/pytorch_efficient_anomaly_detection.ipynb

@@ -206,15 +206,15 @@
       "outputs": [],
       "source": [
         "from huggingface_hub import hf_hub_download\n",
-        "import tutorials.mct_model_garden.models_pytorch.Efficient_Anomaly_Det as efficient_ad\n",
+        "import tutorials.mct_model_garden.models_pytorch.Efficient_Anomaly_Det.efficient_ad import get_pdn_small, get_autoencoder, UnifiedAnomalyDetectionModel\n",
         "\n",
         "out_channels = 384\n",
         "\n",
         "model_path = hf_hub_download(repo_id=\"SSI-DNN/Efficient_Anomaly_Detection\", filename=\"efficientAD_bottle.pth\")\n",
         "\n",
-        "teacher = efficient_ad.get_pdn_small(out_channels)\n",
-        "student = efficient_ad.get_pdn_small(2 * out_channels)\n",
-        "autoencoder = efficient_ad.get_autoencoder(out_channels)\n",
+        "teacher = get_pdn_small(out_channels)\n",
+        "student = get_pdn_small(2 * out_channels)\n",
+        "autoencoder = get_autoencoder(out_channels)\n",
         "\n",
         "model = efficient_ad.UnifiedAnomalyDetectionModel.load_model(model_path, teacher, student, autoencoder)"
       ]
@@ -322,7 +322,9 @@
       "outputs": [],
       "source": [
         "#We first need to calculate the normalisation values\n",
+        "tutorials.resources.utils.efficient_ad_utils import map_normalization, teacher_normalization\n",
         "\n",
+        "teacher_mean, teacher_std = teacher_normalization(teacher, train_loader)\n",
         "q_st_start, q_st_end, q_ae_start, q_ae_end = map_normalization(\n",
         "    validation_loader=train_loader, teacher=teacher, student=student,\n",
         "    autoencoder=autoencoder, teacher_mean=teacher_mean,\n",
@@ -341,7 +343,7 @@
       },
       "outputs": [],
       "source": [
-        "from tutorials.mct_model_garden.anomaly_eval import benchmark\n",
+        "from tutorials.mct_model_garden.evaluation_metrics.anomaly_eval import benchmark\n",
         "benchmark(model, 'mvtec_ad', q_st_start, q_st_end, q_ae_start, q_ae_end)"
       ]
     },
@@ -392,8 +394,8 @@
       },
       "outputs": [],
       "source": [
-        "from tutorials.mct_model_garden.anomaly_eval import benchmark\n",
-        "benchmark(quant_model, 'mvtec_ad_quant')"
+        "from tutorials.mct_model_garden.evaluation_metrics.anomaly_eval import benchmark\n",
+        "benchmark(quant_model, 'mvtec_ad_quant', q_st_start, q_st_end, q_ae_start, q_ae_end)"
       ]
     },
     {
@@ -442,7 +444,7 @@
         "test_output_dir = os.path.join('output', 'anomaly_maps',\n",
         "                                name, 'bottle', 'test')\n",
         "model.eval()\n",
-        "visualize_anomalies(model, dataset_path, test_output_dir)"
+        "visualize_anomalies(model, dataset_path, test_output_dir, q_st_start, q_st_end, q_ae_start, q_ae_end)"
       ]
     },
     {

tutorials/resources/utils/efficient_ad_utils.py

@@ -69,7 +69,7 @@ def train_transform(image):
     """Apply transformations to the training images."""
     return DEFAULT_TRANSFORM(image), DEFAULT_TRANSFORM(TRANSFORM_AE(image))
 
-def visualize_anomalies(unified_model, dataset_path, test_output_dir=None, desc='Running inference'):
+def visualize_anomalies(unified_model, dataset_path, test_output_dir=None, q_st_start=None, q_st_end=None, q_ae_start=None, q_ae_end=None, desc='Running inference'):
     """Visualize anomalies by overlaying heatmaps on the original images."""
     test_set = ImageFolderWithPath(os.path.join(dataset_path, 'bottle', 'test'))
     images_to_display = random.sample(list(test_set), 10)  # Randomly select 10 images to display
@@ -78,7 +78,7 @@ def visualize_anomalies(unified_model, dataset_path, test_output_dir=None, desc=
     for image, target, path in tqdm(images_to_display, desc=desc):
         orig_width, orig_height = image.size
         image_tensor = DEFAULT_TRANSFORM(image)[None]  # Add batch dimension
-        map_combined = unified_model(image_tensor)
+        map_combined , _, _ = predict_combined(image_tensor, unified_model, q_st_start=None, q_st_end=None, q_ae_start=None, q_ae_end=None)
         map_combined = torch.nn.functional.interpolate(map_combined, (orig_height, orig_width), mode='bilinear')
         map_combined = map_combined[0, 0].detach().cpu().numpy()