Skip to content

Evaluator

The Evaluator class is used to evaluate the performance of a model. This class is the parent class for all other evaluators, and acts as an evaluator for object detection models.

Code Documentation

Evaluates the output of a model on a dataset.

Source code in evaluations/evaluator.py
 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
class Evaluator:
    """
    Evaluates the output of a model on a dataset.
    """

    def __init__(
        self,
        predictions: dict = {},
        ground_truth: dict = {},
        class_names: list = [],
        confidence_threshold: int = 0.2,
        mode: str = "interactive",
        name: str = "",
        model_type: str = "",
    ) -> None:
        if mode not in ACCEPTED_MODES:
            raise ValueError(f"mode must be one of {ACCEPTED_MODES}")

        if model_type not in ["multiclass", "classification", "object-detection", "segmentation"]:
            raise ValueError(
                f"model_type must be one of ['multiclass', 'classification', 'object-detection', 'segmentation']"
            )

        self.confidence_threshold = confidence_threshold
        self.mode = mode
        self.class_names = class_names
        self.name = name
        self.model_type = model_type

        if not os.path.exists("./output/images"):
            os.makedirs("./output/images")

        if not os.path.exists("./output/matrices"):
            os.makedirs("./output/matrices")

        merged_data = {}

        for key in predictions.keys():
            gt = ground_truth[key.split("/")[-1]]

            gts = [list(item) for item in gt.xyxy]

            for idx, item in enumerate(gt.class_id):
                gts[idx].append(item)

            merged_data[key] = {
                "ground_truth": gts,
                "predictions": predictions[key]["predictions"],
                "filename": key,
            }

        self.data = merged_data

    def eval_model_predictions(self) -> dict:
        """
        Compute confusion matrices for a Roboflow, Grounding DINO, and CLIP model.

        Returns:
            combined_cf (dict): A dictionary with a confusion matrix composed of the result of all inferences.
        """

        confusion_matrices = {}

        combined_cf = {"fn": 0, "fp": 0}
        combined_cf = {}

        for i, _ in enumerate(self.class_names):
            for j, _ in enumerate(self.class_names):
                combined_cf[(i, j)] = 0

        for key, value in self.data.items():
            print(
                "evaluating image predictions against ground truth",
                key,
                "...",
                self.class_names,
            )

            gt = value["ground_truth"]
            pred = value["predictions"]

            cf = self.compute_confusion_matrix([gt, pred], self.class_names)

            for k, v in cf.items():
                combined_cf[k] += v

            confusion_matrices[key] = cf

            if self.model_type == "object-detection":
                visualize_image_experiment(value, self.class_names)

            plot_confusion_matrix(cf, self.class_names, False, key, self.mode)

        plot_confusion_matrix(
            combined_cf, self.class_names, True, "aggregate", self.mode
        )

        self.combined_cf = combined_cf

        return combined_cf

    def compute_confusion_matrix(
        self,
        image_eval_data,
        class_names,
    ) -> dict:
        # image eval data looks like:
        # {filename: "path/to/image.jpg", ground_truth: [{}, {}, {}, ...], predictions: [{},{},{},...]}
        # each ground truth is a tuple/list of (x_min, y_min, x_max, y_max, label)
        # each prediction is a tuple/list of (x_min, y_min, x_max, y_max, label, confidence)

        predictions = copy.deepcopy(image_eval_data[1])

        all_predictions = []

        for i in range(len(predictions)):
            if (
                predictions.confidence[i] > self.confidence_threshold
                and predictions.class_id[i] is not None
            ):
                merged_prediction = predictions.xyxy[i].tolist() + [
                    predictions.class_id[i]
                ]

                all_predictions.append(merged_prediction)

        ground_truths = copy.deepcopy(image_eval_data[0])

        confusion_data = {}
        for i, _ in enumerate(class_names):
            for j, _ in enumerate(class_names):
                confusion_data[(i, j)] = 0

        for gt_box in ground_truths:
            match_idx, match = find_best_prediction(gt_box, all_predictions)

            if match is None:
                confusion_data[gt_box[4], len(class_names) - 1] += 1
            else:
                all_predictions.pop(match_idx)
                confusion_data[gt_box[4], match[4]] += 1

        for p in all_predictions:
            confusion_data[len(class_names) - 1, p[4]] += 1

        return confusion_data

    def calculate_statistics(self) -> tuple:
        """
        Calculate precision, recall, and f1 score for the evaluation.

        Returns:
            precision: The precision of the model
            recall: The recall of the the model
            f1: The f1 score of the model
        """
        cf = self.combined_cf

        # compute precision, recall, and f1 score
        tp = 0
        fp = 0
        fn = 0

        for x in range(len(self.class_names)):  # ground truth
            for y in range(len(self.class_names)):  # predictions
                if (
                    x == len(self.class_names) - 1
                ):  # last column / prediction with no ground truth
                    fp += cf[(x, y)]
                elif (
                    y == len(self.class_names) - 1
                ):  # bottom row / ground truth with no prediction
                    fn += cf[(x, y)]
                elif x == y:  # true positives across the diagonal
                    tp += cf[(x, y)]
                else:  # misclassification
                    fp += cf[(x, y)]

        if tp + fp == 0:
            precision = 1
        else:
            precision = tp / (tp + fp)

        if tp + fn == 0:
            recall = 1
        else:
            recall = tp / (tp + fn)

        if precision + recall == 0:
            f1 = 0
        else:
            f1 = 2 * (precision * recall) / (precision + recall)

        return EvaluatorResponse(
            true_positives=tp,
            false_positives=fp,
            false_negatives=fn,
            precision=precision,
            recall=recall,
            f1=f1,
        )

calculate_statistics()

Calculate precision, recall, and f1 score for the evaluation.

Returns:

Name Type Description
precision tuple

The precision of the model

recall tuple

The recall of the the model

f1 tuple

The f1 score of the model

Source code in evaluations/evaluator.py
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
def calculate_statistics(self) -> tuple:
    """
    Calculate precision, recall, and f1 score for the evaluation.

    Returns:
        precision: The precision of the model
        recall: The recall of the the model
        f1: The f1 score of the model
    """
    cf = self.combined_cf

    # compute precision, recall, and f1 score
    tp = 0
    fp = 0
    fn = 0

    for x in range(len(self.class_names)):  # ground truth
        for y in range(len(self.class_names)):  # predictions
            if (
                x == len(self.class_names) - 1
            ):  # last column / prediction with no ground truth
                fp += cf[(x, y)]
            elif (
                y == len(self.class_names) - 1
            ):  # bottom row / ground truth with no prediction
                fn += cf[(x, y)]
            elif x == y:  # true positives across the diagonal
                tp += cf[(x, y)]
            else:  # misclassification
                fp += cf[(x, y)]

    if tp + fp == 0:
        precision = 1
    else:
        precision = tp / (tp + fp)

    if tp + fn == 0:
        recall = 1
    else:
        recall = tp / (tp + fn)

    if precision + recall == 0:
        f1 = 0
    else:
        f1 = 2 * (precision * recall) / (precision + recall)

    return EvaluatorResponse(
        true_positives=tp,
        false_positives=fp,
        false_negatives=fn,
        precision=precision,
        recall=recall,
        f1=f1,
    )

eval_model_predictions()

Compute confusion matrices for a Roboflow, Grounding DINO, and CLIP model.

Returns:

Name Type Description
combined_cf dict

A dictionary with a confusion matrix composed of the result of all inferences.

Source code in evaluations/evaluator.py
 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
def eval_model_predictions(self) -> dict:
    """
    Compute confusion matrices for a Roboflow, Grounding DINO, and CLIP model.

    Returns:
        combined_cf (dict): A dictionary with a confusion matrix composed of the result of all inferences.
    """

    confusion_matrices = {}

    combined_cf = {"fn": 0, "fp": 0}
    combined_cf = {}

    for i, _ in enumerate(self.class_names):
        for j, _ in enumerate(self.class_names):
            combined_cf[(i, j)] = 0

    for key, value in self.data.items():
        print(
            "evaluating image predictions against ground truth",
            key,
            "...",
            self.class_names,
        )

        gt = value["ground_truth"]
        pred = value["predictions"]

        cf = self.compute_confusion_matrix([gt, pred], self.class_names)

        for k, v in cf.items():
            combined_cf[k] += v

        confusion_matrices[key] = cf

        if self.model_type == "object-detection":
            visualize_image_experiment(value, self.class_names)

        plot_confusion_matrix(cf, self.class_names, False, key, self.mode)

    plot_confusion_matrix(
        combined_cf, self.class_names, True, "aggregate", self.mode
    )

    self.combined_cf = combined_cf

    return combined_cf

Evaluator Response

Source code in evaluations/evaluator.py
11
12
13
14
15
16
17
18
@dataclass
class EvaluatorResponse:
    true_positives: int
    false_positives: int
    false_negatives: int
    precision: float
    recall: float
    f1: float