Improving readability of charts for insertion and deletion

Author: rpytel1

docker-compose.yml

@@ -0,0 +1,23 @@
+version: "3.8"
+services:
+  trainer:
+    env_file: .env
+    image: "rfl/foxai-trainer:latest"
+    build: "./"
+    container_name: "foxai-trainer"
+    shm_size: 32gb
+    volumes:
+      - ".:/FoXAI"
+      - "${IMG_DIR}/:/home/user/Downloads"
+    ports:
+      - 8888:8888
+      - 6006:6006
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: 1
+              capabilities: [gpu]
+    privileged: true
+    command: ["bash", "-c", "source /etc/bash.bashrc && jupyter lab --notebook-dir=/FoXAI --ip 0.0.0.0 --no-browser --allow-root"]

example/notebooks/deletion_insertion_metric_example.ipynb

@@ -4,14 +4,14 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "# Example notebook for Image Classification (with imagenet-mini)"
+    "# Example notebook for usage of Deletion/Insertion"
    ]
   },
   {
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "In this tutorial, You will see how You can use explainable algorithms to study pre-trained model decision. We will take a pre-trained model, sample images and run several explainable methods."
+    "In this tutorial, You will see how You will understand how to setup to use insertion and deletion metrics for XAI. "
    ]
   },
   {
@@ -273,20 +273,13 @@
     "%matplotlib inline"
    ]
   },
-  {
-   "cell_type": "markdown",
-   "metadata": {},
-   "source": [
-    "Let's see how images from `ImageNet-Mini` looks like. We will display first few samples of dataset. In the following steps we will use them to explain model predictions using different explainable algorithms."
-   ]
-  },
   {
    "cell_type": "markdown",
    "metadata": {
     "tags": []
    },
    "source": [
-    "## Demo for general algorithms "
+    "## Demo for usage of Insertion and Deletion metrics for GradCAM explanations"
    ]
   },
   {
@@ -295,31 +288,8 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "# define list of explainers we want to use\n",
-    "# full list of supported explainers is present in `Explainers` enum class.\n",
-    "explainer_list = [\n",
-    "    ExplainerWithParams(explainer_name=CVClassificationExplainers.CV_GRADIENT_SHAP_EXPLAINER),\n",
-    "    ExplainerWithParams(explainer_name=CVClassificationExplainers.CV_INPUT_X_GRADIENT_EXPLAINER),\n",
-    "    ExplainerWithParams(explainer_name=CVClassificationExplainers.CV_INTEGRATED_GRADIENTS_EXPLAINER),\n",
-    "]"
-   ]
-  },
-  {
-   "cell_type": "markdown",
-   "metadata": {
-    "tags": []
-   },
-   "source": [
-    "## Demo for layer specific algorithms"
-   ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": [
-    "layer = [module for module in model.modules() if isinstance(module, torch.nn.Conv2d)][-1]"
+    "from foxai.metrics import insertion, deletion\n",
+    "from foxai.visualizer import visualize_metric"
    ]
   },
   {
@@ -328,8 +298,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "from foxai.metrics import insertion, deletion\n",
-    "from foxai.visualizer import visualize_metric"
+    "chosen_explainer = CVClassificationExplainers.CV_LAYER_GRADCAM_EXPLAINER"
    ]
   },
   {
@@ -338,7 +307,7 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "type(model)"
+    "layer = [module for module in model.modules() if isinstance(module, torch.nn.Conv2d)][-1]"
    ]
   },
   {
@@ -347,27 +316,28 @@
    "metadata": {},
    "outputs": [],
    "source": [
-    "\n",
     "# iterate over dataloader\n",
     "sample_batch = next(iter(val_dataloader))\n",
     "# iterate over all samples in batch\n",
     "sample, label = sample_batch[0][0], sample_batch[1][0]\n",
+    "sample = sample.to(device)\n",
     "# add batch size dimension to the data sample\n",
-    "input_data = sample.reshape(1, sample.shape[0], sample.shape[1], sample.shape[2]).to(device)\n",
+    "input_data = sample.reshape(1, sample.shape[0], sample.shape[1], sample.shape[2])\n",
     "category_name = categories[label.item()]\n",
     "with FoXaiExplainer(\n",
     "    model=model,\n",
-    "    explainers=[ExplainerWithParams(explainer_name=CVClassificationExplainers.CV_LAYER_GRADCAM_EXPLAINER, layer=layer)],\n",
+    "    explainers=[ExplainerWithParams(explainer_name=chosen_explainer, layer=layer)],\n",
     "    target=label,\n",
     ") as xai_model:\n",
+    "    chosen_explainer_key = f\"{chosen_explainer.name}_0\"\n",
     "    # calculate attributes for every explainer\n",
     "    first_output, attributes_dict = xai_model(input_data)\n",
-    "    value = attributes_dict[\"CV_LAYER_GRADCAM_EXPLAINER\"]\n",
+    "    value = attributes_dict[chosen_explainer_key]\n",
     "    figure = mean_channels_visualization(attributions=value[0], transformed_img=sample, title= f\"Mean of channels)\")\n",
     "    # save figure to artifact directory\n",
     "    show_figure(figure) \n",
     "    \n",
-    "    gradcam_maps = attributes_dict[\"CV_LAYER_GRADCAM_EXPLAINER\"]\n",
+    "    gradcam_maps = attributes_dict[chosen_explainer_key]\n",
     "    value = gradcam_maps[0]\n",
     "    chosen_class = first_output.argmax()\n",
     "    insertion_result, importance_lst = insertion(value, sample, model, chosen_class)\n",
@@ -376,13 +346,6 @@
     "    visualize_metric(importance_lst, deletion_result, metric_type=\"Deletion\")\n",
     "        \n"
    ]
-  },
-  {
-   "cell_type": "code",
-   "execution_count": null,
-   "metadata": {},
-   "outputs": [],
-   "source": []
   }
  ],
  "metadata": {
@@ -401,7 +364,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.7.13"
+   "version": "3.8.10"
   },
   "vscode": {
    "interpreter": {

foxai/metrics/insertion_deletion_metrics.py

@@ -39,21 +39,25 @@ def _metric_calculation(
     transformed_img: torch.Tensor,
     model: ModelType,
     chosen_class: int,
-    steps_num: int = 30,
-    metric_type: Metrics = Metrics.INSERTION,
+    steps_num=30,
+    metric_type=Metrics.INSERTION,
+    kernel=(101, 101),
 ) -> Tuple[np.ndarray, List]:
-    """Calculate metric (insertion or deletion) given importance map, image, model and chosen class.
+    """Calculate metric (insertion or deletion) given importance map, image, model and chosen class. Implementation of both metrics
+      (insertion and deletion) are inspired by the paper "RISE: Randomized Input Sampling for Explanation of Black-box Models"
+        : https://arxiv.org/abs/1806.07421
 
     Args:
         attributions: Torch Tensor corresponding to importance map.
         transformed_img: Torch Tensor corresponding to image.
         model: model which we are explaining.
         chosen_class: index of the class we are creating metric for.
         metric_type: type of metric presented using enum, supported ones are: Insertion and Deletion.
+        kernel: define a tuple regarding the used blurring kernel. Default value is 101 to produce very blurred value.
 
     Returns:
         metric: numerical value of chosen metric for given picture and explanation.
-        importance_lst: list of numpy elements corresponding to confidence value at each step.
+        importance_list: list of numpy elements corresponding to confidence value at each step.
 
     Raises:
         AttributeError: if metric type is not enum of Metrics.INSERTION or Metrics.DELETION
@@ -74,7 +78,7 @@ def _metric_calculation(
     sorted_attrs: np.ndarray = np.flip(np.sort(np.unique(preprocessed_attrs)))
     stepped_attrs: np.ndarray = _get_stepped_attrs(sorted_attrs, steps_num)
 
-    importance_lst: List[np.ndarray] = []
+    importance_list: List[Tuple[float, float]] = []
 
     cuda = next(model.parameters()).is_cuda
     device = torch.device("cuda" if cuda else "cpu")
@@ -83,7 +87,7 @@ def _metric_calculation(
     removed_img_part[:] = transformed_img.mean()
 
     if metric_type == Metrics.INSERTION:
-        removed_img_part = gaussian_blur(transformed_img, (101, 101))
+        removed_img_part = gaussian_blur(transformed_img, kernel)
 
     for val in stepped_attrs:
         attributes_map_np: np.ndarray = np.expand_dims(
@@ -115,11 +119,17 @@ def _metric_calculation(
 
         output = model(perturbed_img.unsqueeze(dim=0))
         softmax_output: torch.Tensor = torch.nn.functional.softmax(output)[0]
-        importance_lst.append(softmax_output[chosen_class].detach().numpy())
+        importance_val: float = float(
+            softmax_output[chosen_class].detach().cpu().numpy()
+        )
+        importance_list.append((val, importance_val))
 
-    metric: np.ndarray = np.round(np.trapz(importance_lst) / len(importance_lst), 4)
+    importance_values: List = [elem[0] for elem in importance_list]
+    metric: np.ndarray = np.round(
+        np.trapz(importance_values) / len(importance_values), 4
+    )
 
-    return metric, importance_lst
+    return metric, importance_list
 
 
 def deletion(
@@ -138,7 +148,7 @@ def deletion(
 
     Returns:
         metric: numerical value of chosen metric for given picture and explanation.
-        importance_lst: list of numpy elements corresponding to confidence value at each step.
+        importance_list: list of numpy elements corresponding to confidence value at each step.
 
     Raises:
         AttributeError: if metric type is not enum of Metrics.INSERTION or Metrics.DELETION
@@ -153,6 +163,7 @@ def insertion(
     transformed_img: torch.Tensor,
     model: ModelType,
     chosen_class: int,
+    kernel=(101, 101),
 ) -> Tuple[np.ndarray, List]:
     """Calculate insertion metric given importance map, image, model and chosen class.
 
@@ -164,7 +175,7 @@ def insertion(
 
     Returns:
         metric: numerical value of chosen metric for given picture and explanation.
-        importance_lst: list of numpy elements corresponding to confidence value at each step.
+        importance_list: list of numpy elements corresponding to confidence value at each step.
 
     Raises:
         AttributeError: if metric type is not enum of Metrics.INSERTION or Metrics.DELETION
@@ -175,4 +186,5 @@ def insertion(
         model,
         chosen_class,
         metric_type=Metrics.INSERTION,
+        kernel=kernel,
     )

foxai/visualizer.py

@@ -244,17 +244,29 @@ def single_channel_visualization(
 
 
 def visualize_metric(
-    importance_lst: List[np.ndarray],
+    importance_list: List[np.ndarray],
     metric_result: float,
     metric_type: str = "Deletion",
 ):
     """
-    Visualize graph for Insertion or deletion metric based on which area under the curve is caluclated.
+    Visualize chart for Insertion or deletion metric based on which area under the curve is caluclated.
+
+    Args:
+        importance_list: List of temporary results used for calculating metrics like deletion or insertion.
+        metric_result: Overall result of metric.
+        metric_type: String name of visualized metric (currently supported are Deletion and Insertion).
     """
+    x_vals = [elem[0] for elem in importance_list]
+    y_vals = [elem[1] for elem in importance_list]
     plt.ylim((0, 1))
-    plt.xlim((0, len(importance_lst)))
-    plt.plot(np.arange(len(importance_lst)), importance_lst)
+    plt.xlim((0, 1))
+    plt.plot(x_vals, y_vals)
     plt.title(f"{metric_type}: {metric_result}")
+    plt.ylabel("Probablity of predicting chosen class")
+    if metric_type == "Deletion":
+        plt.xlabel("Percentage of image removed")
+    else:
+        plt.xlabel("Percentage of image revealed")
     plt.show()