rknn-toolkit
rknn-toolkit copied to clipboard
rockchip-linux
yolov5的onnx转rknn后，优化有错误
toolkit 1.7.0版本 下面是错误摘要。这个装下转换出来的V6.0的yolov5n yolov5s的性能在rk1109下只有6fps和4fps --> Loading model W Call onnx.optimizer.optimize fail, skip optimize
W Warning: Axis may need to be adjusted according to original model shape. W Warning: Axis may need to be adjusted according to original model shape. W Warning: Axis may need to be adjusted according to original model shape. W Warning: Axis may need to be adjusted according to original model shape.
--> Init runtime environment librknn_runtime version 1.7.0 (9ad5c07 build: 2021-08-06 15:20:31 base: 1131) E [op_optimize:428]CONCAT, uid 7 must have same quantize parameter! E [op_optimize:428]CONCAT, uid 6 must have same quantize parameter! E [op_optimize:428]CONCAT, uid 5 must have same quantize parameter! E [op_optimize:428]CONCAT, uid 5 must have same quantize parameter! E [op_optimize:428]CONCAT, uid 6 must have same quantize parameter! E [op_optimize:428]CONCAT, uid 7 must have same quantize parameter!
老哥，你装rknn-1.7.0遇到问题了吗？我是TB3399Pro的板子，安装rknn1.7.0的时候，一直有问题。
我这边是1126和1109，目前设备连接评估什么都没问题
我这边TB-3399pro，yolov5s转rknn后，帧率也就是6FPS。 单就inference来说，每一帧耗时100ms左右（640*640）
我这边TB-3399pro，yolov5s转rknn后，帧率也就是6FPS。 单就inference来说，每一帧耗时100ms左右（640*640）
我这里用的yolov5n，用rknntoolkit的工具测试下来，320x320是48.x 的fps
@chohou 那直接export后onnx转换rknn会出现问题吧
我用i16的精度的话，没有问题，用u8和i8的话，没有结果。 据说是yolo官网V5的节点的问题，不改的话用u8和i8会有精度问题。
@chohou 那直接export后onnx转换rknn会出现问题吧
我用i16的精度的话，没有问题，用u8和i8的话，没有结果。 据说是yolo官网V5的节点的问题，不改的话用u8和i8会有精度问题。
这个情况请参考 #93 问题
toolkit 1.7.0版本 下面是错误摘要。这个装下转换出来的V6.0的yolov5n yolov5s的性能在rk1109下只有6fps和4fps --> Loading model W Call onnx.optimizer.optimize fail, skip optimize
W Warning: Axis may need to be adjusted according to original model shape. W Warning: Axis may need to be adjusted according to original model shape. W Warning: Axis may need to be adjusted according to original model shape. W Warning: Axis may need to be adjusted according to original model shape.
--> Init runtime environment librknn_runtime version 1.7.0 (9ad5c07 build: 2021-08-06 15:20:31 base: 1131) E [op_optimize:428]CONCAT, uid 7 must have same quantize parameter! E [op_optimize:428]CONCAT, uid 6 must have same quantize parameter! E [op_optimize:428]CONCAT, uid 5 must have same quantize parameter! E [op_optimize:428]CONCAT, uid 5 must have same quantize parameter! E [op_optimize:428]CONCAT, uid 6 must have same quantize parameter! E [op_optimize:428]CONCAT, uid 7 must have same quantize parameter!
@chohou Did you solve the problem? I had the same errors. I cant go forward..
toolkit 1.7.0版本 下面是错误摘要。这个装下转换出来的V6.0的yolov5n yolov5s的性能在rk1109下只有6fps和4fps --> Loading model W Call onnx.optimizer.optimize fail, skip optimize
W Warning: Axis may need to be adjusted according to original model shape. W Warning: Axis may need to be adjusted according to original model shape. W Warning: Axis may need to be adjusted according to original model shape. W Warning: Axis may need to be adjusted according to original model shape.
--> Init runtime environment librknn_runtime version 1.7.0 (9ad5c07 build: 2021-08-06 15:20:31 base: 1131) E [op_optimize:428]CONCAT, uid 7 must have same quantize parameter! E [op_optimize:428]CONCAT, uid 6 must have same quantize parameter! E [op_optimize:428]CONCAT, uid 5 must have same quantize parameter! E [op_optimize:428]CONCAT, uid 5 must have same quantize parameter! E [op_optimize:428]CONCAT, uid 6 must have same quantize parameter! E [op_optimize:428]CONCAT, uid 7 must have same quantize parameter!
我修改一部分之后会出错，加载程序会被killed，全部u8则不会
