yyccr / yolov5_in_tf2_keras

YOLOv5 in TF2 > TFLite > ONNX > TensorRT

YOLOv5 in tesnorflow2.x-keras

• yolov5数据增强jupyter示例
• Bilibili视频讲解地址: 《yolov5 解读,训练,复现》
• Bilibili视频讲解PPT文件: yolov5_bilibili_talk_ppt.pdf
• Bilibili视频讲解PPT文件: yolov5_bilibili_talk_ppt.pdf (gitee链接)
• yolov5 tflite android 部署
• yolov5 ncnn android 部署

模型测试

• 训练 猫狗检测 3.7k

• 检测效果

()

• TFLite Android 测试效果：

• [email protected]/[email protected]:0.95/精度/召回率

class	[email protected]	[email protected]:0.95	precision	recall
cat	0.962680	0.672483	0.721003	0.958333
dog	0.934285	0.546893	0.770701	0.923664
total	0.948482	0.609688	0.745852	0.940999

Requirements

pip3 install -r requirements.txt

Get start

0. 下载数据集

https://www.kaggle.com/datasets/andrewmvd/dog-and-cat-detection/download
或者从releasev1.0下载:
https://github.com/yyccR/yolov5_in_tf2_keras/releases/download/v1.0/JPEGImages.zip

解压数据将images目录修改为JPEGImages, 放到 ./data/cat_dog_face_data下

1. 训练

python3 train.py

2. tensorboard

tensorboard --host 0.0.0.0 --logdir ./logs/ --port 8053 --samples_per_plugin=images=40

3. 查看

http://127.0.0.1:8053

4. 测试, 修改detect.py里面input_image和model_path

python3 detect.py

5. 评估验证

python3 val.py

6. 导出TFLite格式

cd ./data
python3 ./h5_to_tfite.py

7. 导出ONNX格式

cd ./data
python3 ./h5_to_onnx.py

训练自己的数据

1. labelme打标自己的数据
2. 打开data/labelme2coco.py脚本, 修改如下地方

input_dir = '这里写labelme打标时保存json标记文件的目录'
output_dir = '这里写要转CoCo格式的目录，建议建一个空目录'
labels = "这里是你打标时所有的类别名, txt文本即可, 每行一个类, 类名无需加引号"

3. 执行data/labelme2coco.py脚本会在output_dir生成对应的json文件和图片
4. 修改train.py文件中train_coco_json, val_coco_json, num_class, classes
5. 开始训练, python3 train.py

yolov5_in_tf2_keras's Issues

关于类别损失计算的问题

您好，loss这段不是很理解，

yolov5_in_tf2_keras/loss.py

Lines 142 to 152 in 3e6645c

	if w_ratio_bool and h_ratio_bool:
	grid_xy = np.array(box_xy // (1 / grid_size), np.int32)
	# 0,1,2 % 3 = 0,1,2 3,4,5 % 3 = 0,1,2 6,7,8 % 3 = 0,1,2
	best_anchor_id = anchor_idx % len(self.anchor_masks[0])
	if self.only_best_anchor:
	y_true_out[i, grid_xy[0], grid_xy[1], best_anchor_id, :] = \
	[box[0], box[1], box[2], box[3], 1, gt_box_class_anchor[i, j, 4]]
	else:
	y_true_out[i, grid_xy[0], grid_xy[1], :, :] = \
	np.array([[box[0], box[1], box[2], box[3], 1, gt_box_class_anchor[i, j, 4]]])

请问targets最后两位应该是置信度1和最佳的anchor索引吗？

yolov5_in_tf2_keras/loss.py

Lines 288 to 293 in 3e6645c

	# ----------------- 这里处理target数据 --------------------------
	# [batch, grid, grid, anchors, 6(x1, y1, x2, y2, obj, class)]
	target = targets[i]
	true_box, true_obj, true_cls = tf.split(target, (4, 1, 1), axis=-1)
	true_xy = (true_box[..., 0:2] + true_box[..., 2:4]) / 2
	true_wh = true_box[..., 2:4] - true_box[..., 0:2]

那这边split出来的true_obj, true_cls应该就是对应的置信度1和最佳的anchor索引吧。
那这个类别损失

yolov5_in_tf2_keras/loss.py

Line 356 in 3e6645c

class_loss = obj_mask * tf.keras.losses.sparse_categorical_crossentropy(true_cls, pred_cls)

计算的不是最佳anchor索引吗，是跟obj_mask 有关系吗

能使用tfserving部署吗？

如题

大佬ppt失效了

sparse_categorical_crossentropy训练时有nan结果

有的数据会在这行出现nan

yolov5_in_tf2_keras/loss.py

Line 357 in 033a115

class_loss = obj_mask * tf.keras.losses.sparse_categorical_crossentropy(true_cls, pred_cls)

查看了input的is_nan，都正常。而且把sparse_categorical_crossentropy换成binary_crossentropy就好了。
请问这两者在这里计算有差别吗，是否可以进行替换

What is the mAP on COCO17 val ?

Hi @yyccR, thanks for your repo. I want to know if you can reach the same mAP as in original YOLOV5 (Train on COCO17 train and test on COCO17 val)? And do you have plan to release some pretrained checkpoint ?

layers.py

根據ultralytics/yolov5:

https://github.com/ultralytics/yolov5/blob/63ddb6f0d06f6309aa42bababd08c859197a27af/models/common.py#L70-L73

這一段程式:

yolov5_in_tf2_keras/layers.py

Lines 127 to 132 in 46298d7

	def call(self, inputs, *args, **kwargs):
	x = self.multiheadAttention(self.q(inputs), self.v(inputs), self.k(inputs)) + inputs
	x = self.fc1(x)
	x = self.fc2(x)
	x = x + inputs
	return x

是不是應該改寫成:

    def call(self, inputs, *args, **kwargs):
        y = self.multiheadAttention(self.q(inputs), self.v(inputs), self.k(inputs)) + inputs
        x = self.fc1(x)
        x = self.fc2(x)
        x = x +  y
        return x

generate_anchors.py 还存在吗？？

Source of loss function

From where you refered the loss function ?

怎么添加关键点分支可以讲解一下吗?

lebelme2coco处理逻辑有误

我在实际使用您的代码训练自己的数据集时发现，labelme2coco.py 好像缺少对shape_type == "rectangle"时的处理，导致我最后生成的json文件annotations项为空。
以下是labelme2coco.py文件100行到124行代码：
` if shape_type == "polygon":
mask = labelme.utils.shape_to_mask(
img.shape[:2], points, shape_type
)
# cv2.imshow("",np.array(mask, dtype=np.uint8)*255)
# cv2.waitKey(0)

            if group_id is None:
                group_id = uuid.uuid1()

            instance = (label, group_id)
            # print(instance)

            if instance in masks:
                masks[instance] = masks[instance] | mask
            else:
                masks[instance] = mask
            # print(masks[instance].shape)

            if shape_type == "rectangle":
                (x1, y1), (x2, y2) = points
                x1, x2 = sorted([x1, x2])
                y1, y2 = sorted([y1, y2])
                points = [x1, y1, x2, y1, x2, y2, x1, y2]
            if shape_type == "circle": 
            ....

`
代码永远不会执行到shape_type == "rectangle"或shape_type == "circle"。