gjm441

I met the same problem earlier. It means that your weight value of net is type of float , but the tensor of net is fp16. You can print the...

@hfutxrg Try to use the function "HalfToFloat", converting the net to float instead.

You can fork latest caffe from BVLC, and then add layers such as 'interp layer' related files to that caffe. Recompile and work.

Thanks for reply.When i change to op11,it occurs: [05/27/2021-04:47:17] [V] [TRT] ModelImporter.cpp:119: Searching for input: add_8:0 [05/27/2021-04:47:17] [V] [TRT] ModelImporter.cpp:119: Searching for input: f3/kernel/read/_20__cf__20:0 [05/27/2021-04:47:17] [V] [TRT] ModelImporter.cpp:125: f3/MatMul [MatMul]...

**I tried opset with 11，occurs error： [Recurrence]'s output can not be used as shape tensor** &&&& RUNNING TensorRT.trtexec # trtexec --onnx=onnx/dien_op11.onnx --minShapes=Inputs/uid_batch_ph:0:1,Inputs/mid_batch_ph:0:1,Inputs/cat_batch_ph:0:1,Inputs/mid_his_batch_ph:0:19,Inputs/cat_his_batch_ph:0:19,Inputs/mask:0:19,Inputs/seq_len_ph:0:1 --optShapes=Inputs/uid_batch_ph:0:1,Inputs/mid_batch_ph:0:1,Inputs/cat_batch_ph:0:1,Inputs/mid_his_batch_ph:0:19,Inputs/cat_his_batch_ph:0:19,Inputs/mask:0:19,Inputs/seq_len_ph:0:1 --maxShapes=Inputs/uid_batch_ph:0:1,Inputs/mid_batch_ph:0:1,Inputs/cat_batch_ph:0:1,Inputs/mid_his_batch_ph:0:19,Inputs/cat_his_batch_ph:0:19,Inputs/mask:0:19,Inputs/seq_len_ph:0:1 --saveEngine=engine/dien_b1_fp32_static.engine [06/16/2021-08:17:03] [I] === Model Options...

I trained shufflenet-ssd on widerface, easy and medium AP only got 0.6 .It got bad performance, the trainning loss stayed at 3.7 at last and hard to descend .May be...

As I know, shufflenet-ssd is faster than mobilenet-ssd on mobile devices,especially for ncnn.