Bubbliiiing

icon indicating a website link https://blog.csdn.net/weixin_44791964

icon location Bubbliiiing哔哩哔哩空间： https://space.bilibili.com/472467171

Results 2197 comments of Bubbliiiing

博主大佬，yolov4单小目标检测怎么找指标最好的权重，是根据F1指标还是？

如果没有difficult参数，这两个应该差距不会很大

训练过程中报错：ValueError: Dimensions must be equal, but are 512 and 38 for 'conv4_3_norm/mul' (op: 'Mul') with input shapes: [?,38,38,512], [38].

版本

训练过程中报错：ValueError: Dimensions must be equal, but are 512 and 38 for 'conv4_3_norm/mul' (op: 'Mul') with input shapes: [?,38,38,512], [38].

我咋感觉你好像在B站问过我

训练过程中报错：ValueError: Dimensions must be equal, but are 512 and 38 for 'conv4_3_norm/mul' (op: 'Mul') with input shapes: [?,38,38,512], [38].

一般而言yolov4效果好，但是yolov4网络比ssd大很多，如果输入图片特征比较不明显（不需要高的语义信息），可能无法使得yolov4拟合，此时SSD就会好很多

训练过程中报错：ValueError: Dimensions must be equal, but are 512 and 38 for 'conv4_3_norm/mul' (op: 'Mul') with input shapes: [?,38,38,512], [38].

keras版本

It is better to write “*/datasets/%s/%sA/*” as “/your/cyclegan/datasets/%s/%sA/*”, otherwise path_A and path_B is empty

Windows or Ubuntu？It runs greatly in my computer……It might be the problem of system

80.36的mAP有用预训练权重吗？

有

80.36的mAP有用预训练权重吗？

……很正常

就没人问问这是python几吗？？？

你训练了多久？多少代？多少数据集？

就没人问问这是python几吗？？？

300多张是非常少的，你可以试试单阶段的先。