Fix RISE algorithm for explain function

CHANGELOG.md

@@ -35,8 +35,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
   (<https://github.com/openvinotoolkit/datumaro/pull/1245>)
 - Enable image backend and color channel format to be selectable
   (<https://github.com/openvinotoolkit/datumaro/pull/1246>)
-- Enhance Datumaro data format detect() to be memory-bounded and performant
-  (<https://github.com/openvinotoolkit/datumaro/pull/1229>)
+- Enhance RISE algortihm for explainable AI
+  (<https://github.com/openvinotoolkit/datumaro/pull/1263>)
 
 ### Bug fixes
 - Fix wrong example of Datumaro dataset creation in document

src/datumaro/cli/commands/explain.py

@@ -98,7 +98,7 @@ def build_parser(parser_ctor=argparse.ArgumentParser):
     rise_parser.add_argument(
         "-s",
         "--max-samples",
-        default=None,
+        default=100,
         type=int,
         help="Number of algorithm iterations (default: mask size ^ 2)",
     )
@@ -203,13 +203,9 @@ def explain_command(args):
 
     rise = RISE(
         model,
-        max_samples=args.max_samples,
-        mask_width=args.mask_width,
-        mask_height=args.mask_height,
+        num_masks=args.max_samples,
+        mask_size=args.mask_width,
         prob=args.prob,
-        iou_thresh=args.iou_thresh,
-        nms_thresh=args.nms_iou_thresh,
-        det_conf_thresh=args.det_conf_thresh,
         batch_size=args.batch_size,
     )
 

src/datumaro/components/algorithms/rise.py

@@ -1,29 +1,20 @@
-# Copyright (C) 2019-2020 Intel Corporation
+# Copyright (C) 2019-2024 Intel Corporation
 #
 # SPDX-License-Identifier: MIT
 
 # pylint: disable=unused-variable
 
-from math import ceil
-
+import cv2
 import numpy as np
 
-from datumaro.components.annotation import AnnotationType
-from datumaro.util.annotation_util import nms
+from datumaro.components.dataset import Dataset
+from datumaro.components.dataset_base import DatasetItem
+from datumaro.components.media import Image
+from datumaro.util import take_by
 
 __all__ = ["RISE"]
 
 
-def _flatmatvec(mat):
-    return np.reshape(mat, (len(mat), -1))
-
-
-def _expand(array, axis=None):
-    if axis is None:
-        axis = len(array.shape)
-    return np.expand_dims(array, axis=axis)
-
-
 class RISE:
     """
     Implements RISE: Randomized Input Sampling for
@@ -34,186 +25,104 @@
     def __init__(
         self,
         model,
-        max_samples=None,
-        mask_width=7,
-        mask_height=7,
-        prob=0.5,
-        iou_thresh=0.9,
-        nms_thresh=0.0,
-        det_conf_thresh=0.0,
-        batch_size=1,
+        num_masks: int = 100,
+        mask_size: int = 7,
+        prob: float = 0.5,
+        batch_size: int = 1,
     ):
+        assert prob >= 0 and prob <= 1
         self.model = model
-        self.max_samples = max_samples
-        self.mask_height = mask_height
-        self.mask_width = mask_width
+        self.num_masks = num_masks
+        self.mask_size = mask_size
         self.prob = prob
-        self.iou_thresh = iou_thresh
-        self.nms_thresh = nms_thresh
-        self.det_conf_thresh = det_conf_thresh
         self.batch_size = batch_size
 
-    @staticmethod
-    def split_outputs(annotations):
-        labels = []
-        bboxes = []
-        for r in annotations:
-            if r.type is AnnotationType.label:
-                labels.append(r)
-            elif r.type is AnnotationType.bbox:
-                bboxes.append(r)
-        return labels, bboxes
-
-    def normalize_hmaps(self, heatmaps, counts):
-        eps = np.finfo(heatmaps.dtype).eps
-        mhmaps = _flatmatvec(heatmaps)
-        mhmaps /= _expand(counts * self.prob + eps)
-        mhmaps -= _expand(np.min(mhmaps, axis=1))
-        mhmaps /= _expand(np.max(mhmaps, axis=1) + eps)
-        return np.reshape(mhmaps, heatmaps.shape)
+    def normalize_saliency(self, saliency):
+        normalized_saliency = np.empty_like(saliency)
+        for idx, sal in enumerate(saliency):
+            normalized_saliency[idx, ...] = (sal - np.min(sal)) / (np.max(sal) - np.min(sal))
+        return normalized_saliency
 
-    def apply(self, image, progressive=False):
-        import cv2
+    def generate_masks(self, image_size):
+        cell_size = np.ceil(np.array(image_size) / self.mask_size).astype(np.int8)
+        up_size = tuple([(self.mask_size + 1) * cs for cs in cell_size])
+
+        grid = np.random.rand(self.num_masks, self.mask_size, self.mask_size) < self.prob
+        grid = grid.astype("float32")
+
+        masks = np.empty((self.num_masks, *image_size))
+        for i in range(self.num_masks):
+            # Random shifts
+            x = np.random.randint(0, cell_size[0])
+            y = np.random.randint(0, cell_size[1])
 
+            # Linear upsampling and cropping
+            masks[i, ...] = cv2.resize(grid[i], up_size, interpolation=cv2.INTER_LINEAR)[
+                x : x + image_size[0], y : y + image_size[1]
+            ]
+
+        return masks
+
+    def generate_masked_dataset(self, image, image_size, masks):
+        input_image = cv2.resize(image, image_size, interpolation=cv2.INTER_LINEAR)
+
+        items = []
+        for id, mask in enumerate(masks):
+            masked_image = np.expand_dims(mask, axis=-1) * input_image
+            items.append(
+                DatasetItem(
+                    id=id,
+                    media=Image.from_numpy(masked_image),
+                )
+            )
+        return Dataset.from_iterable(items)
+
+    def apply(self, image, progressive=False):
         assert len(image.shape) in [2, 3], "Expected an input image in (H, W, C) format"
         if len(image.shape) == 3:
             assert image.shape[2] in [3, 4], "Expected BGR or BGRA input"
         image = image[:, :, :3].astype(np.float32)
 
         model = self.model
-        iou_thresh = self.iou_thresh
-
-        image_size = np.array((image.shape[:2]))
-        mask_size = np.array((self.mask_height, self.mask_width))
-        cell_size = np.ceil(image_size / mask_size)
-        upsampled_size = np.ceil((mask_size + 1) * cell_size)
-
-        rng = lambda shape=None: np.random.rand(*shape)
-        samples = np.prod(image_size)
-        if self.max_samples is not None:
-            samples = min(self.max_samples, samples)
-        batch_size = self.batch_size
-
-        # model is expected to get NxCxHxW shaped input tensor
-        pred = next(iter(model.infer(_expand(np.transpose(image, (2, 0, 1)), 0))))
-        result = model.postprocess(pred, None)
-        result_labels, result_bboxes = self.split_outputs(result)
-        if 0 < self.det_conf_thresh:
-            result_bboxes = [
-                b for b in result_bboxes if self.det_conf_thresh <= b.attributes["score"]
-            ]
-        if 0 < self.nms_thresh:
-            result_bboxes = nms(result_bboxes, self.nms_thresh)
-
-        predicted_labels = set()
-        if len(result_labels) != 0:
-            predicted_label = max(result_labels, key=lambda r: r.attributes["score"]).label
-            predicted_labels.add(predicted_label)
-        if len(result_bboxes) != 0:
-            for bbox in result_bboxes:
-                predicted_labels.add(bbox.label)
-        predicted_labels = {label: idx for idx, label in enumerate(predicted_labels)}
-
-        predicted_bboxes = result_bboxes
-
-        heatmaps_count = len(predicted_labels) + len(predicted_bboxes)
-        heatmaps = np.zeros((heatmaps_count, *image_size), dtype=np.float32)
-        total_counts = np.zeros(heatmaps_count, dtype=np.int32)
-        confs = np.zeros(heatmaps_count, dtype=np.float32)
-
-        heatmap_id = 0
-
-        # label_heatmaps = None
-        label_total_counts = None
-        label_confs = None
-        if len(predicted_labels) != 0:
-            step = len(predicted_labels)
-            # label_heatmaps = heatmaps[heatmap_id : heatmap_id + step]
-            label_total_counts = total_counts[heatmap_id : heatmap_id + step]
-            label_confs = confs[heatmap_id : heatmap_id + step]
-            heatmap_id += step
-
-        # bbox_heatmaps = None
-        bbox_total_counts = None
-        bbox_confs = None
-        if len(predicted_bboxes) != 0:
-            step = len(predicted_bboxes)
-            # bbox_heatmaps = heatmaps[heatmap_id : heatmap_id + step]
-            bbox_total_counts = total_counts[heatmap_id : heatmap_id + step]
-            bbox_confs = confs[heatmap_id : heatmap_id + step]
-            heatmap_id += step
-
-        ups_mask = np.empty(upsampled_size.astype(int), dtype=np.float32)
-        masks = np.empty((batch_size, *image_size), dtype=np.float32)
-
-        full_batch_inputs = np.empty((batch_size, *image.shape), dtype=np.float32)
-        current_heatmaps = np.empty_like(heatmaps)
-        for b in range(ceil(samples / batch_size)):
-            batch_pos = b * batch_size
-            current_batch_size = min(samples - batch_pos, batch_size)
-
-            batch_masks = masks[:current_batch_size]
-            for i in range(current_batch_size):
-                mask = (rng(mask_size) < self.prob).astype(np.float32)
-                cv2.resize(mask, (int(upsampled_size[1]), int(upsampled_size[0])), ups_mask)
-
-                offsets = np.round(rng((2,)) * cell_size)
-                mask = ups_mask[
-                    int(offsets[0]) : int(image_size[0] + offsets[0]),
-                    int(offsets[1]) : int(image_size[1] + offsets[1]),
-                ]
-                batch_masks[i] = mask
-
-            batch_inputs = full_batch_inputs[:current_batch_size]
-            np.multiply(_expand(batch_masks), _expand(image, 0), out=batch_inputs)
-
-            preds = model.infer(np.transpose(batch_inputs, (0, 3, 1, 2)))
-            results = [model.postprocess(pred, None) for pred in preds]
-            for mask, result in zip(batch_masks, results):
-                result_labels, result_bboxes = self.split_outputs(result)
-
-                confs.fill(0)
-                if len(predicted_labels) != 0:
-                    for r in result_labels:
-                        idx = predicted_labels.get(r.label, None)
-                        if idx is not None:
-                            label_total_counts[idx] += 1
-                            label_confs[idx] += r.attributes["score"]
-                    for r in result_bboxes:
-                        idx = predicted_labels.get(r.label, None)
-                        if idx is not None:
-                            label_total_counts[idx] += 1
-                            label_confs[idx] += r.attributes["score"]
-
-                if len(predicted_bboxes) != 0 and len(result_bboxes) != 0:
-                    if 0 < self.det_conf_thresh:
-                        result_bboxes = [
-                            b
-                            for b in result_bboxes
-                            if self.det_conf_thresh <= b.attributes["score"]
-                        ]
-                    if 0 < self.nms_thresh:
-                        result_bboxes = nms(result_bboxes, self.nms_thresh)
-
-                    for detection in result_bboxes:
-                        for pred_idx, pred in enumerate(predicted_bboxes):
-                            if pred.label != detection.label:
-                                continue
-
-                            iou = pred.iou(detection)
-                            assert iou == -1 or 0 <= iou and iou <= 1
-                            if iou < iou_thresh:
-                                continue
-
-                            bbox_total_counts[pred_idx] += 1
-
-                            conf = detection.attributes["score"]
-                            bbox_confs[pred_idx] += conf
-
-                np.multiply.outer(confs, mask, out=current_heatmaps)
-                heatmaps += current_heatmaps
+
+        image_size = model.inputs[0].shape
+        logit_size = model.outputs[0].shape
+
+        batch_size = image_size[0]
+        if image_size[1] in [1, 3]:  # for CxHxW
+            image_size = (image_size[2], image_size[3])
+        elif image_size[3] in [1, 3]:  # for HxWxC
+            image_size = (image_size[1], image_size[2])
+
+        masks = self.generate_masks(image_size=image_size)
+        masked_dataset = self.generate_masked_dataset(image, image_size, masks)
+
+        saliency = np.zeros((logit_size[1], *image_size), dtype=np.float32)
+        for batch_id, batch in enumerate(take_by(masked_dataset, batch_size)):
+            outputs = model.launch(batch)
+
+            for sample_id in range(len(batch)):
+                mask = masks[batch_size * batch_id + sample_id]
+                for class_idx in range(logit_size[1]):
+                    score = outputs[sample_id][class_idx].attributes["score"]
+                    saliency[class_idx, ...] += score * mask
+
+                # [TODO] wonjuleee: support DRISE for detection model explainability
+                # if isinstance(self.target, Label):
+                #     logits = outputs[sample_id][0].vector
+                #     max_score = logits[self.target.label]
+                # elif isinstance(self.target, Bbox):
+                #     preds = outputs[sample_id][0]
+                #     max_score = 0
+                #     for box in preds:
+                #         if box[0] == self.target.label:
+                #             confidence, box = box[1], box[2]
+                #             score = iou(self.target.get_bbox, box) * confidence
+                #             if score > max_score:
+                #                 max_score = score
+                # saliency += max_score * mask
 
                 if progressive:
-                    yield self.normalize_hmaps(heatmaps.copy(), total_counts)
+                    yield self.normalize_saliency(saliency)
 
-        yield self.normalize_hmaps(heatmaps, total_counts)
+        yield self.normalize_saliency(saliency)

src/datumaro/components/shift_analyzer.py

@@ -4,7 +4,6 @@
 
 # ruff: noqa: E501
 
-import itertools
 from collections import defaultdict
 from typing import TYPE_CHECKING, Dict, List, Optional
 
@@ -79,8 +78,9 @@ def get_activation_stats(self, dataset: IDataset) -> RunningStats1D:
         running_stats = RunningStats1D()
 
         for batch in take_by(dataset, self._batch_size):
-            features = self.model.launch(batch)
-            running_stats.add(list(itertools.chain(*features)))
+            outputs = self.model.launch(batch)[0]
+            features = [outputs[-1]]  # extracted feature vector of googlenet-v4
+            running_stats.add(features)
 
         return running_stats
 
@@ -99,10 +99,11 @@ def get_activation_stats(self, dataset: IDataset) -> Dict[int, RunningStats1D]:
                     inputs.append(np.atleast_3d(item.media.data))
                     targets.append(ann.label)
 
-            features = self.model.launch(batch)
+            outputs = self.model.launch(batch)[0]
+            features = [outputs[-1]]  # extracted feature vector of googlenet-v4
 
-            for feat, target in zip(features, targets):
-                running_stats[target].add(feat)
+            for target in targets:
+                running_stats[target].add(features)
 
         return running_stats
 

src/datumaro/plugins/openvino_plugin/launcher.py

@@ -208,6 +208,14 @@ def __init__(
         self._check_model_support(self._network, self._device)
         self._load_executable_net()
 
+    @property
+    def inputs(self):
+        return self._network.inputs
+
+    @property
+    def outputs(self):
+        return self._network.outputs
+
     def _check_model_support(self, net, device):
         not_supported_layers = set(
             name for name, dev in self._core.query_model(net, device).items() if not dev