search

Gumpest / YOLOv5-Multibackbone-Compression

YOLOv5 Series Multi-backbone(TPH-YOLOv5, Ghostnet, ShuffleNetv2, Mobilenetv3Small, EfficientNetLite, PP-LCNet, SwinTransformer YOLO), Module(CBAM, DCN), Pruning (EagleEye, Network Slimming), Quantization (MQBench) and Deployment (TensorRT, ncnn) Compression Tool Box.
995 stars 202 forks source link

预训练权重 #42

Open Stepnen opened 2 years ago

Stepnen commented 2 years ago

您好，请问修改了网络结构之后（例如YOLOv5s-ghost），用原来的权重文件(例如yolov5s.pt)效果会不理想，这种情况该怎么办呀？非常感谢您能答复！

Gumpest commented 2 years ago

请问训练之后的精度下降多少呢

Stepnen commented 2 years ago

有点大，20%左右…

---原始邮件--- 发件人: @.> 发送时间: 2021年12月23日(周四) 晚上9:43 收件人: @.>; 抄送: @.**@.>; 主题: Re: [Gumpest/YOLOv5-Multibackbone-Compression] 预训练权重 (Issue #42)

请问训练之后的精度下降多少呢

— Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android. You are receiving this because you authored the thread.Message ID: @.***>

PeterSmith1 commented 2 years ago

请问训练之后的精度下降多少呢

作者你好，请问一下，你能分享下swin的权重吗？我拿官方权重精度有点低