File size: 15,106 Bytes
18dd6ad
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
# Copyright (c) Facebook, Inc. and its affiliates.
import copy
import itertools
import json
import logging
import os
import pickle
from collections import OrderedDict
import torch

import annotator.oneformer.detectron2.utils.comm as comm
from annotator.oneformer.detectron2.config import CfgNode
from annotator.oneformer.detectron2.data import MetadataCatalog
from annotator.oneformer.detectron2.structures import Boxes, BoxMode, pairwise_iou
from annotator.oneformer.detectron2.utils.file_io import PathManager
from annotator.oneformer.detectron2.utils.logger import create_small_table

from .coco_evaluation import instances_to_coco_json
from .evaluator import DatasetEvaluator


class LVISEvaluator(DatasetEvaluator):
    """
    Evaluate object proposal and instance detection/segmentation outputs using
    LVIS's metrics and evaluation API.
    """

    def __init__(
        self,
        dataset_name,
        tasks=None,
        distributed=True,
        output_dir=None,
        *,
        max_dets_per_image=None,
    ):
        """
        Args:
            dataset_name (str): name of the dataset to be evaluated.
                It must have the following corresponding metadata:
                "json_file": the path to the LVIS format annotation
            tasks (tuple[str]): tasks that can be evaluated under the given
                configuration. A task is one of "bbox", "segm".
                By default, will infer this automatically from predictions.
            distributed (True): if True, will collect results from all ranks for evaluation.
                Otherwise, will evaluate the results in the current process.
            output_dir (str): optional, an output directory to dump results.
            max_dets_per_image (None or int): limit on maximum detections per image in evaluating AP
                This limit, by default of the LVIS dataset, is 300.
        """
        from lvis import LVIS

        self._logger = logging.getLogger(__name__)

        if tasks is not None and isinstance(tasks, CfgNode):
            self._logger.warn(
                "COCO Evaluator instantiated using config, this is deprecated behavior."
                " Please pass in explicit arguments instead."
            )
            self._tasks = None  # Infering it from predictions should be better
        else:
            self._tasks = tasks

        self._distributed = distributed
        self._output_dir = output_dir
        self._max_dets_per_image = max_dets_per_image

        self._cpu_device = torch.device("cpu")

        self._metadata = MetadataCatalog.get(dataset_name)
        json_file = PathManager.get_local_path(self._metadata.json_file)
        self._lvis_api = LVIS(json_file)
        # Test set json files do not contain annotations (evaluation must be
        # performed using the LVIS evaluation server).
        self._do_evaluation = len(self._lvis_api.get_ann_ids()) > 0

    def reset(self):
        self._predictions = []

    def process(self, inputs, outputs):
        """
        Args:
            inputs: the inputs to a LVIS model (e.g., GeneralizedRCNN).
                It is a list of dict. Each dict corresponds to an image and
                contains keys like "height", "width", "file_name", "image_id".
            outputs: the outputs of a LVIS model. It is a list of dicts with key
                "instances" that contains :class:`Instances`.
        """
        for input, output in zip(inputs, outputs):
            prediction = {"image_id": input["image_id"]}

            if "instances" in output:
                instances = output["instances"].to(self._cpu_device)
                prediction["instances"] = instances_to_coco_json(instances, input["image_id"])
            if "proposals" in output:
                prediction["proposals"] = output["proposals"].to(self._cpu_device)
            self._predictions.append(prediction)

    def evaluate(self):
        if self._distributed:
            comm.synchronize()
            predictions = comm.gather(self._predictions, dst=0)
            predictions = list(itertools.chain(*predictions))

            if not comm.is_main_process():
                return
        else:
            predictions = self._predictions

        if len(predictions) == 0:
            self._logger.warning("[LVISEvaluator] Did not receive valid predictions.")
            return {}

        if self._output_dir:
            PathManager.mkdirs(self._output_dir)
            file_path = os.path.join(self._output_dir, "instances_predictions.pth")
            with PathManager.open(file_path, "wb") as f:
                torch.save(predictions, f)

        self._results = OrderedDict()
        if "proposals" in predictions[0]:
            self._eval_box_proposals(predictions)
        if "instances" in predictions[0]:
            self._eval_predictions(predictions)
        # Copy so the caller can do whatever with results
        return copy.deepcopy(self._results)

    def _tasks_from_predictions(self, predictions):
        for pred in predictions:
            if "segmentation" in pred:
                return ("bbox", "segm")
        return ("bbox",)

    def _eval_predictions(self, predictions):
        """
        Evaluate predictions. Fill self._results with the metrics of the tasks.

        Args:
            predictions (list[dict]): list of outputs from the model
        """
        self._logger.info("Preparing results in the LVIS format ...")
        lvis_results = list(itertools.chain(*[x["instances"] for x in predictions]))
        tasks = self._tasks or self._tasks_from_predictions(lvis_results)

        # LVIS evaluator can be used to evaluate results for COCO dataset categories.
        # In this case `_metadata` variable will have a field with COCO-specific category mapping.
        if hasattr(self._metadata, "thing_dataset_id_to_contiguous_id"):
            reverse_id_mapping = {
                v: k for k, v in self._metadata.thing_dataset_id_to_contiguous_id.items()
            }
            for result in lvis_results:
                result["category_id"] = reverse_id_mapping[result["category_id"]]
        else:
            # unmap the category ids for LVIS (from 0-indexed to 1-indexed)
            for result in lvis_results:
                result["category_id"] += 1

        if self._output_dir:
            file_path = os.path.join(self._output_dir, "lvis_instances_results.json")
            self._logger.info("Saving results to {}".format(file_path))
            with PathManager.open(file_path, "w") as f:
                f.write(json.dumps(lvis_results))
                f.flush()

        if not self._do_evaluation:
            self._logger.info("Annotations are not available for evaluation.")
            return

        self._logger.info("Evaluating predictions ...")
        for task in sorted(tasks):
            res = _evaluate_predictions_on_lvis(
                self._lvis_api,
                lvis_results,
                task,
                max_dets_per_image=self._max_dets_per_image,
                class_names=self._metadata.get("thing_classes"),
            )
            self._results[task] = res

    def _eval_box_proposals(self, predictions):
        """
        Evaluate the box proposals in predictions.
        Fill self._results with the metrics for "box_proposals" task.
        """
        if self._output_dir:
            # Saving generated box proposals to file.
            # Predicted box_proposals are in XYXY_ABS mode.
            bbox_mode = BoxMode.XYXY_ABS.value
            ids, boxes, objectness_logits = [], [], []
            for prediction in predictions:
                ids.append(prediction["image_id"])
                boxes.append(prediction["proposals"].proposal_boxes.tensor.numpy())
                objectness_logits.append(prediction["proposals"].objectness_logits.numpy())

            proposal_data = {
                "boxes": boxes,
                "objectness_logits": objectness_logits,
                "ids": ids,
                "bbox_mode": bbox_mode,
            }
            with PathManager.open(os.path.join(self._output_dir, "box_proposals.pkl"), "wb") as f:
                pickle.dump(proposal_data, f)

        if not self._do_evaluation:
            self._logger.info("Annotations are not available for evaluation.")
            return

        self._logger.info("Evaluating bbox proposals ...")
        res = {}
        areas = {"all": "", "small": "s", "medium": "m", "large": "l"}
        for limit in [100, 1000]:
            for area, suffix in areas.items():
                stats = _evaluate_box_proposals(predictions, self._lvis_api, area=area, limit=limit)
                key = "AR{}@{:d}".format(suffix, limit)
                res[key] = float(stats["ar"].item() * 100)
        self._logger.info("Proposal metrics: \n" + create_small_table(res))
        self._results["box_proposals"] = res


# inspired from Detectron:
# https://github.com/facebookresearch/Detectron/blob/a6a835f5b8208c45d0dce217ce9bbda915f44df7/detectron/datasets/json_dataset_evaluator.py#L255 # noqa
def _evaluate_box_proposals(dataset_predictions, lvis_api, thresholds=None, area="all", limit=None):
    """
    Evaluate detection proposal recall metrics. This function is a much
    faster alternative to the official LVIS API recall evaluation code. However,
    it produces slightly different results.
    """
    # Record max overlap value for each gt box
    # Return vector of overlap values
    areas = {
        "all": 0,
        "small": 1,
        "medium": 2,
        "large": 3,
        "96-128": 4,
        "128-256": 5,
        "256-512": 6,
        "512-inf": 7,
    }
    area_ranges = [
        [0**2, 1e5**2],  # all
        [0**2, 32**2],  # small
        [32**2, 96**2],  # medium
        [96**2, 1e5**2],  # large
        [96**2, 128**2],  # 96-128
        [128**2, 256**2],  # 128-256
        [256**2, 512**2],  # 256-512
        [512**2, 1e5**2],
    ]  # 512-inf
    assert area in areas, "Unknown area range: {}".format(area)
    area_range = area_ranges[areas[area]]
    gt_overlaps = []
    num_pos = 0

    for prediction_dict in dataset_predictions:
        predictions = prediction_dict["proposals"]

        # sort predictions in descending order
        # TODO maybe remove this and make it explicit in the documentation
        inds = predictions.objectness_logits.sort(descending=True)[1]
        predictions = predictions[inds]

        ann_ids = lvis_api.get_ann_ids(img_ids=[prediction_dict["image_id"]])
        anno = lvis_api.load_anns(ann_ids)
        gt_boxes = [
            BoxMode.convert(obj["bbox"], BoxMode.XYWH_ABS, BoxMode.XYXY_ABS) for obj in anno
        ]
        gt_boxes = torch.as_tensor(gt_boxes).reshape(-1, 4)  # guard against no boxes
        gt_boxes = Boxes(gt_boxes)
        gt_areas = torch.as_tensor([obj["area"] for obj in anno])

        if len(gt_boxes) == 0 or len(predictions) == 0:
            continue

        valid_gt_inds = (gt_areas >= area_range[0]) & (gt_areas <= area_range[1])
        gt_boxes = gt_boxes[valid_gt_inds]

        num_pos += len(gt_boxes)

        if len(gt_boxes) == 0:
            continue

        if limit is not None and len(predictions) > limit:
            predictions = predictions[:limit]

        overlaps = pairwise_iou(predictions.proposal_boxes, gt_boxes)

        _gt_overlaps = torch.zeros(len(gt_boxes))
        for j in range(min(len(predictions), len(gt_boxes))):
            # find which proposal box maximally covers each gt box
            # and get the iou amount of coverage for each gt box
            max_overlaps, argmax_overlaps = overlaps.max(dim=0)

            # find which gt box is 'best' covered (i.e. 'best' = most iou)
            gt_ovr, gt_ind = max_overlaps.max(dim=0)
            assert gt_ovr >= 0
            # find the proposal box that covers the best covered gt box
            box_ind = argmax_overlaps[gt_ind]
            # record the iou coverage of this gt box
            _gt_overlaps[j] = overlaps[box_ind, gt_ind]
            assert _gt_overlaps[j] == gt_ovr
            # mark the proposal box and the gt box as used
            overlaps[box_ind, :] = -1
            overlaps[:, gt_ind] = -1

        # append recorded iou coverage level
        gt_overlaps.append(_gt_overlaps)
    gt_overlaps = (
        torch.cat(gt_overlaps, dim=0) if len(gt_overlaps) else torch.zeros(0, dtype=torch.float32)
    )
    gt_overlaps, _ = torch.sort(gt_overlaps)

    if thresholds is None:
        step = 0.05
        thresholds = torch.arange(0.5, 0.95 + 1e-5, step, dtype=torch.float32)
    recalls = torch.zeros_like(thresholds)
    # compute recall for each iou threshold
    for i, t in enumerate(thresholds):
        recalls[i] = (gt_overlaps >= t).float().sum() / float(num_pos)
    # ar = 2 * np.trapz(recalls, thresholds)
    ar = recalls.mean()
    return {
        "ar": ar,
        "recalls": recalls,
        "thresholds": thresholds,
        "gt_overlaps": gt_overlaps,
        "num_pos": num_pos,
    }


def _evaluate_predictions_on_lvis(
    lvis_gt, lvis_results, iou_type, max_dets_per_image=None, class_names=None
):
    """
    Args:
        iou_type (str):
        max_dets_per_image (None or int): limit on maximum detections per image in evaluating AP
            This limit, by default of the LVIS dataset, is 300.
        class_names (None or list[str]): if provided, will use it to predict
            per-category AP.

    Returns:
        a dict of {metric name: score}
    """
    metrics = {
        "bbox": ["AP", "AP50", "AP75", "APs", "APm", "APl", "APr", "APc", "APf"],
        "segm": ["AP", "AP50", "AP75", "APs", "APm", "APl", "APr", "APc", "APf"],
    }[iou_type]

    logger = logging.getLogger(__name__)

    if len(lvis_results) == 0:  # TODO: check if needed
        logger.warn("No predictions from the model!")
        return {metric: float("nan") for metric in metrics}

    if iou_type == "segm":
        lvis_results = copy.deepcopy(lvis_results)
        # When evaluating mask AP, if the results contain bbox, LVIS API will
        # use the box area as the area of the instance, instead of the mask area.
        # This leads to a different definition of small/medium/large.
        # We remove the bbox field to let mask AP use mask area.
        for c in lvis_results:
            c.pop("bbox", None)

    if max_dets_per_image is None:
        max_dets_per_image = 300  # Default for LVIS dataset

    from lvis import LVISEval, LVISResults

    logger.info(f"Evaluating with max detections per image = {max_dets_per_image}")
    lvis_results = LVISResults(lvis_gt, lvis_results, max_dets=max_dets_per_image)
    lvis_eval = LVISEval(lvis_gt, lvis_results, iou_type)
    lvis_eval.run()
    lvis_eval.print_results()

    # Pull the standard metrics from the LVIS results
    results = lvis_eval.get_results()
    results = {metric: float(results[metric] * 100) for metric in metrics}
    logger.info("Evaluation results for {}: \n".format(iou_type) + create_small_table(results))
    return results