keras-yolo3 エラーValueError: Dimension 0 in both shapes must be equal, but are 1 and 42. 〜の解決法を教えて下さい

前提・実現したいこと

keras-yolo3を使用してオリジナルデータをwebカメラを通し、物体の識別をリアルタイムで行いたいです。

発生している問題・エラーメッセージ

参考サイト１と参考サイト２を参考にしてオリジナルデータを作成しました。
参考サイト３の最後の方に書いてある「独自データによる学習済みモデルを読み込んで、物体検出を行う」の項目と同じようにプログラムコードを書き換えました。
webカメラを通してリアルタイム検出を行いたいので、参考サイト４の⑤に書いてあるwebカメラを使ってリアルタイム検知するプログラムで実行したところエラーが発生しました。

ValueError: Dimension 0 in both shapes must be equal, but are 1 and 42. Shapes are [1,1,1024,45] and [42,1024,1,1]. for 'Assign_360' (op: 'Assign') with input shapes: [1,1,1024,45], [42,1024,1,1].

該当のソースコード

python
1# -*- coding: utf-8 -*-
2"""
3Class definition of YOLO_v3 style detection model on image and video
4"""
5
6import colorsys
7import os
8from timeit import default_timer as timer
9
10import numpy as np
11from keras import backend as K
12from keras.models import load_model
13from keras.layers import Input
14from PIL import Image, ImageFont, ImageDraw
15
16from yolo3.model import yolo_eval, yolo_body, tiny_yolo_body
17from yolo3.utils import letterbox_image
18import os
19from keras.utils import multi_gpu_model
20
21class YOLO(object):
22    _defaults = {
23        "model_path": 'model_data/trained_weights_final.h5',
24        "anchors_path": 'model_data/yolo_anchors.txt',
25        "classes_path": 'model_data/my_classes.txt',
26        "score" : 0.3,
27        "iou" : 0.45,
28        "model_image_size" : (416, 416),
29        "gpu_num" : 1,
30    }
31
32    @classmethod
33    def get_defaults(cls, n):
34        if n in cls._defaults:
35            return cls._defaults[n]
36        else:
37            return "Unrecognized attribute name '" + n + "'"
38
39    def __init__(self, **kwargs):
40        self.__dict__.update(self._defaults) # set up default values
41        self.__dict__.update(kwargs) # and update with user overrides
42        self.class_names = self._get_class()
43        self.anchors = self._get_anchors()
44        self.sess = K.get_session()
45        self.boxes, self.scores, self.classes = self.generate()
46
47    def _get_class(self):
48        classes_path = os.path.expanduser(self.classes_path)
49        with open(classes_path) as f:
50            class_names = f.readlines()
51        class_names = [c.strip() for c in class_names]
52        return class_names
53
54    def _get_anchors(self):
55        anchors_path = os.path.expanduser(self.anchors_path)
56        with open(anchors_path) as f:
57            anchors = f.readline()
58        anchors = [float(x) for x in anchors.split(',')]
59        return np.array(anchors).reshape(-1, 2)
60
61    def generate(self):
62        model_path = os.path.expanduser(self.model_path)
63        assert model_path.endswith('.h5'), 'Keras model or weights must be a .h5 file.'
64
65        # Load model, or construct model and load weights.
66        num_anchors = len(self.anchors)
67        num_classes = len(self.class_names)
68        is_tiny_version = num_anchors==6 # default setting
69        try:
70            self.yolo_model = load_model(model_path, compile=False)
71        except:
72            self.yolo_model = tiny_yolo_body(Input(shape=(None,None,3)), num_anchors//2, num_classes) \
73                if is_tiny_version else yolo_body(Input(shape=(None,None,3)), num_anchors//3, num_classes)
74            self.yolo_model.load_weights(self.model_path) # make sure model, anchors and classes match
75        else:
76            assert self.yolo_model.layers[-1].output_shape[-1] == \
77                num_anchors/len(self.yolo_model.output) * (num_classes + 5), \
78                'Mismatch between model and given anchor and class sizes'
79
80        print('{} model, anchors, and classes loaded.'.format(model_path))
81
82        # Generate colors for drawing bounding boxes.
83        hsv_tuples = [(x / len(self.class_names), 1., 1.)
84                      for x in range(len(self.class_names))]
85        self.colors = list(map(lambda x: colorsys.hsv_to_rgb(*x), hsv_tuples))
86        self.colors = list(
87            map(lambda x: (int(x[0] * 255), int(x[1] * 255), int(x[2] * 255)),
88                self.colors))
89        np.random.seed(10101)  # Fixed seed for consistent colors across runs.
90        np.random.shuffle(self.colors)  # Shuffle colors to decorrelate adjacent classes.
91        np.random.seed(None)  # Reset seed to default.
92
93        # Generate output tensor targets for filtered bounding boxes.
94        self.input_image_shape = K.placeholder(shape=(2, ))
95        if self.gpu_num>=2:
96            self.yolo_model = multi_gpu_model(self.yolo_model, gpus=self.gpu_num)
97        boxes, scores, classes = yolo_eval(self.yolo_model.output, self.anchors,
98                len(self.class_names), self.input_image_shape,
99                score_threshold=self.score, iou_threshold=self.iou)
100        return boxes, scores, classes
101
102    def detect_image(self, image):
103        start = timer()
104
105        if self.model_image_size != (None, None):
106            assert self.model_image_size[0]%32 == 0, 'Multiples of 32 required'
107            assert self.model_image_size[1]%32 == 0, 'Multiples of 32 required'
108            boxed_image = letterbox_image(image, tuple(reversed(self.model_image_size)))
109        else:
110            new_image_size = (image.width - (image.width % 32),
111                              image.height - (image.height % 32))
112            boxed_image = letterbox_image(image, new_image_size)
113        image_data = np.array(boxed_image, dtype='float32')
114
115        print(image_data.shape)
116        image_data /= 255.
117        image_data = np.expand_dims(image_data, 0)  # Add batch dimension.
118
119        out_boxes, out_scores, out_classes = self.sess.run(
120            [self.boxes, self.scores, self.classes],
121            feed_dict={
122                self.yolo_model.input: image_data,
123                self.input_image_shape: [image.size[1], image.size[0]],
124                K.learning_phase(): 0
125            })
126
127        print('Found {} boxes for {}'.format(len(out_boxes), 'img'))
128
129        font = ImageFont.truetype(font='font/FiraMono-Medium.otf',
130                    size=np.floor(3e-2 * image.size[1] + 0.5).astype('int32'))
131        thickness = (image.size[0] + image.size[1]) // 300
132
133        for i, c in reversed(list(enumerate(out_classes))):
134            predicted_class = self.class_names[c]
135            box = out_boxes[i]
136            score = out_scores[i]
137
138            label = '{} {:.2f}'.format(predicted_class, score)
139            draw = ImageDraw.Draw(image)
140            label_size = draw.textsize(label, font)
141
142            top, left, bottom, right = box
143            top = max(0, np.floor(top + 0.5).astype('int32'))
144            left = max(0, np.floor(left + 0.5).astype('int32'))
145            bottom = min(image.size[1], np.floor(bottom + 0.5).astype('int32'))
146            right = min(image.size[0], np.floor(right + 0.5).astype('int32'))
147            print(label, (left, top), (right, bottom))
148
149            if top - label_size[1] >= 0:
150                text_origin = np.array([left, top - label_size[1]])
151            else:
152                text_origin = np.array([left, top + 1])
153
154            # My kingdom for a good redistributable image drawing library.
155            for i in range(thickness):
156                draw.rectangle(
157                    [left + i, top + i, right - i, bottom - i],
158                    outline=self.colors[c])
159            draw.rectangle(
160                [tuple(text_origin), tuple(text_origin + label_size)],
161                fill=self.colors[c])
162            draw.text(text_origin, label, fill=(0, 0, 0), font=font)
163            del draw
164
165        end = timer()
166        print(end - start)
167        return image
168
169    def close_session(self):
170        self.sess.close()
171
172def detect_video(yolo, video_path, output_path=""):
173    import cv2
174    vid = cv2.VideoCapture(video_path)
175    if not vid.isOpened():
176        raise IOError("Couldn't open webcam or video")
177    video_FourCC    = int(vid.get(cv2.CAP_PROP_FOURCC))
178    video_fps       = vid.get(cv2.CAP_PROP_FPS)
179    video_size      = (int(vid.get(cv2.CAP_PROP_FRAME_WIDTH)),
180                        int(vid.get(cv2.CAP_PROP_FRAME_HEIGHT)))
181    isOutput = True if output_path != "" else False
182    if isOutput:
183        print("!!! TYPE:", type(output_path), type(video_FourCC), type(video_fps), type(video_size))
184        out = cv2.VideoWriter(output_path, video_FourCC, video_fps, video_size)
185    accum_time = 0
186    curr_fps = 0
187    fps = "FPS: ??"
188    prev_time = timer()
189    while True:
190        return_value, frame = vid.read()
191        image = Image.fromarray(frame)
192        image = yolo.detect_image(image)
193        result = np.asarray(image)
194        curr_time = timer()
195        exec_time = curr_time - prev_time
196        prev_time = curr_time
197        accum_time = accum_time + exec_time
198        curr_fps = curr_fps + 1
199        if accum_time > 1:
200            accum_time = accum_time - 1
201            fps = "FPS: " + str(curr_fps)
202            curr_fps = 0
203        cv2.putText(result, text=fps, org=(3, 15), fontFace=cv2.FONT_HERSHEY_SIMPLEX,
204                    fontScale=0.50, color=(255, 0, 0), thickness=2)
205        cv2.namedWindow("result", cv2.WINDOW_NORMAL)
206        cv2.imshow("result", result)
207        if isOutput:
208            out.write(result)
209        if cv2.waitKey(1) & 0xFF == ord('q'):
210            break
211    yolo.close_session()
212

試したこと

オリジナルデータに書き換える前に元々のサンプルでwebカメラでのリアルタイム検知をやってみたところ普通に動きました。

補足情報

MacBookPro 2020

Anaconda

python3.7

Tensorflow 1.14.0

Keras 2.2.4

もっと良い方法がある場合は追加で教えて頂けると幸いです。

temps1101

2020/10/24 23:43

おそらくエラーが出るところにtry文を足せばうまくいくと思います。でもkatamuki7様のエラーメッセージが何行目に起こったかわからないのでどこを直せば良いのかわからないです。できればエラーが起こった行数を載せてほしいです。こんな感じ⇓ File "〇〇.py", line 〇〇〇〇〇 ValueError: Dimension 0 in both shapes must be equal, but are 1 and 42. Shapes are [1,1,1024,45] and [42,1024,1,1]. for 'Assign_360' (op: 'Assign') with input shapes: [1,1,1024,45], [42,1024,1,1].