精度下降严重

[YONGHUICAI]

作者您好，我根据您的步骤，发现剪枝后精度下降比较厉害，在剪枝率=0.1的时候，mAP=0.793179，当时推理时间反而变长了；在剪枝率=0.2的时候，mAP=0.693096 ；在剪枝率=0.5时，mAP=0，推理时已经完全没有输出检测框了。我运行的步骤是先运行第二步的稀疏化训练，然后执行第三步的规整剪枝，请问一下我是哪个环节出现问题了吗？

规整剪枝的结果： 1）percent=0.1时： +------------+----------+----------+ | Metric | Before | After | +------------+----------+----------+ | mAP | 0.812109 | 0.793179 | | Parameters | 61523734 | 60048534 | | Inference | 0.1162 | 0.1154 | +------------+----------+----------+ Config file has been saved: cfg/prune_0.1_yolov3-hand.cfg Compact model has been saved: weights/yolov3_hand_regular_pruning_0.1percent.weights

2）percent=0.2时： +------------+----------+----------+ | Metric | Before | After | +------------+----------+----------+ | mAP | 0.812109 | 0.693096 | | Parameters | 61523734 | 52141142 | | Inference | 0.1036 | 0.1764 | +------------+----------+----------+ Config file has been saved: cfg/prune_0.2_yolov3-hand.cfg Compact model has been saved: weights/yolov3_hand_regular_pruning_0.2percent.weights

3）percent=0.5时： +------------+----------+----------+ | Metric | Before | After | +------------+----------+----------+ | mAP | 0.812109 | 0.000000 | | Parameters | 61523734 | 14787023 | | Inference | 0.1010 | 0.0495 | +------------+----------+----------+ Config file has been saved: cfg/prune_0.5_yolov3-hand.cfg Compact model has been saved: weights/yolov3_hand_shortcut_pruning_0.5percent.weights

[CheungBH]

大概率没稀疏好，先判断以下稀疏效果再剪枝吧。实际上有些网络就算mAP变０，微调后依然能回到不错的水准

[forwardwfg]

请问解决了吗，我也遇到类似的情况

[Serissa]

大概率没稀疏好，先判断以下稀疏效果再剪枝吧。实际上有些网络就算mAP变０，微调后依然能回到不错的水准

请问稀疏化训练好-大概需要训练多少epoch？或者在loss或者mAP波动很小后还要训练多少epoch？