File size: 8,223 Bytes
098c98c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from typing import *
from pathlib import Path
import importlib

import albumentations as A
import cv2
from omegaconf import DictConfig, OmegaConf
import numpy as np

from .utils import read_image, standardize, normalize_cdf, normalize_min_max, get_boxes

if importlib.util.find_spec("openvino") is not None:
    import openvino as ov
else:
    raise ImportError("OpenVINO is not installed.")

class OpenVINOInferencer():
    """For perform inference with OpenVINO.

    Args:
        path (str | Path): Folder path to exported OpenVINO model. Must contains
            model.xml, model.bin, and metadata.json.
        device (str): Device to run inference on. Defaults to "AUTO".
        cache_dir (str | Path): Cache directory for OpenVINO 
    """
    def __init__(self,
                 path: Union[str, Path],
                 device: str="AUTO",
                 cache_dir: Union[str, Path, None]=None
                 ) -> None:
        if isinstance(path, str):
            path = Path(path)

        self.model = self._load_model(path, device, cache_dir)
        self.metadata = self._load_metadata(path)

        # Note: Transformation require Albumentations package
        self.transform = A.from_dict(self.metadata["transform"])
        self.metadata["expand_offset"] = self._get_expand_offset(self.transform)

        # Record input & output blob (key)
        self.metadata["input_blob"] = self.model.input(0).get_names().pop()
        self.metadata["output_blob"] = self.model.output(0).get_names().pop()

    def _load_model(self, path: Path, device: str, cache_dir: Union[str, Path, None]) -> ov.CompiledModel:
        xml_path = path / "model.xml"
        bin_path = path / "model.bin"

        ov_core = ov.Core()
        model = ov_core.read_model(xml_path, bin_path)

        # Create cache directory
        if cache_dir is None:
            cache_dir = "cache"
        if isinstance(cache_dir, str):
            cache_dir = Path(cache_dir)
        cache_dir.mkdir(parents=True, exist_ok=True)
        ov_core.set_property({"CACHE_DIR": cache_dir})

        model = ov_core.compile_model(model=model, device_name=device.upper())
        return model

    def _load_metadata(self, path: Path) -> DictConfig:
        metadata = path / "metadata.json"
        metadata = OmegaConf.load(metadata)
        metadata = cast(DictConfig, metadata)
        return metadata

    def _get_expand_offset(self, transform):
        is_center_cropped = False
        for t in reversed(transform.transforms):
            if isinstance(t, A.CenterCrop):
                is_center_cropped = True
                cropped_h = t.height
                cropped_w = t.width
            elif isinstance(t, A.Resize) and is_center_cropped:
                return (t.height - cropped_h) // 2, (t.width - cropped_w) // 2

    def predict(self, image: Union[str, Path, np.ndarray]) -> Dict[str, np.ndarray]:
        if isinstance(image, (str, Path)):
            image = read_image(image)

        # Record input image size
        self.metadata["image_shape"] = image.shape[:2]

        inputs = self._pre_process(image, self.transform)
        predictions = self.model(inputs)
        outputs = self._post_processs(predictions, self.metadata)
        outputs.update({"image": image})
        return outputs

    def __call__(self, image: Union[str, Path, np.ndarray]) -> Dict[str, np.ndarray]:
        return self.predict(image)

    def _pre_process(self, image: np.ndarray, transform=None) -> np.ndarray:
        if transform is not None:
            image = transform(image=image)["image"]

        if len(image.shape) == 3:
            # Add batch_size axis
            image = np.expand_dims(image, axis=0)

        if image.shape[3] == 3:
            # Transpose the color_channel axis
            # Expected shape: [b, c, h, w]
            image = image.transpose(0, 3, 1, 2)

        return image

    def _post_processs(self, predictions: np.ndarray, metadata: DictConfig) -> Dict[str, np.ndarray]:
        predictions = predictions[metadata["output_blob"]]

        anomaly_map: np.ndarray = None
        pred_label: float = None
        pred_mask: float = None

        if metadata["task"] == "classification":
            pred_score = predictions
        else:
            anomaly_map = predictions.squeeze()
            pred_score = anomaly_map.reshape(-1).max()

        if "image_threshold" in metadata:
            # Assign anomalous label to predictions with score >= threshold
            pred_label = pred_score >= metadata["image_threshold"]

        if metadata["task"] == "classification":
            _, pred_score = self._normalize(pred_scores=pred_score, metadata=metadata)
        else:
            if "pixel_threshold" in metadata:
                pred_mask = (anomaly_map >= metadata["pixel_threshold"]).astype(np.uint8)

            anomaly_map, pred_score = self._normalize(
                pred_scores=pred_score,
                metadata=metadata,
                anomaly_map=anomaly_map
            )

            if "image_shape" in metadata and anomaly_map.shape != metadata["image_shape"]:
                if "expand_offset" in metadata and metadata["expand_offset"] is not None:
                    anomaly_map = self._expand(anomaly_map, metadata["expand_offset"][0], metadata["expand_offset"][1])
                    pred_mask = self._expand(pred_mask, metadata["expand_offset"][0], metadata["expand_offset"][1])
                h, w = metadata["image_shape"] # Fix: cv2.resize take (w, h) as argument
                anomaly_map = cv2.resize(anomaly_map, (w, h))
                if pred_mask is not None:
                    pred_mask = cv2.resize(pred_mask, (w, h))

        if metadata["task"] == "detection":
            pred_boxes = get_boxes(pred_mask)
            box_labels = np.ones(pred_boxes.shape[0])
        else:
            pred_boxes: np.ndarray | None = None
            box_labels: np.ndarray | None = None

        return {
            "anomaly_map": anomaly_map,
            "pred_label": pred_label,
            "pred_score": pred_score,
            "pred_mask": pred_mask,
            "pred_boxes": pred_boxes,
            "box_labels": box_labels
        }

    @staticmethod
    def _expand(map, offset_h, offset_w):
        h, w = map.shape
        if map is not None:
            expanded_map = np.zeros((h + offset_h * 2, w + offset_w * 2), dtype=map.dtype)
            expanded_map[offset_h:offset_h+h, offset_w:offset_w+w] = map
            return expanded_map

    @staticmethod
    def _normalize(
        pred_scores: np.float32,
        metadata: DictConfig,
        anomaly_map: np.ndarray | None = None
        ) -> Tuple[Union[np.ndarray, None], float]:
        # Min-max normalization
        if "min" in metadata and "max" in metadata:
            if anomaly_map is not None:
                anomaly_map = normalize_min_max(
                    anomaly_map,
                    metadata["pixel_threshold"],
                    metadata["min"],
                    metadata["max"]
                )
            pred_scores = normalize_min_max(
                pred_scores,
                metadata["image_threshold"],
                metadata["min"],
                metadata["max"]
            )

        # Standardize pixel scores
        if "pixel_mean" in metadata and "pixel_std" in metadata:
            if anomaly_map is not None:
                anomaly_map = standardize(
                    anomaly_map,
                    metadata["pixel_mean"],
                    metadata["pixel_std"],
                    center_at=metadata["image_mean"]
                )
                anomaly_map = normalize_cdf(
                    anomaly_map,
                    metadata["pixel_threshold"]
                )

        # Standardize image scores
        if "image_mean" in metadata and "image_std" in metadata:
            pred_scores = standardize(
                pred_scores,
                metadata["image_mean"],
                metadata["image_std"]
            )
            pred_scores = normalize_cdf(
                pred_scores,
                metadata["image_threshold"]
            )

        return anomaly_map, float(pred_scores)