search

airockchip / rknn-toolkit2

Other
996 stars 104 forks source link

Converting ViT_T results in a too large file size. #214

Open cao-nv opened 1 week ago

cao-nv commented 1 week ago

Hi Rockchip devs, I exported some models including ViT-T and R100 for face recognition to RKNN for deploying on RK3568. Everything worked well with R100, but the exported ViT-T file size is about 2.6GB while the input ONNX file size is only 76MB. The exporting process takes a much longer time than usual. Could you give me some suggestions? Thanks.

yuyun2000 commented 1 week ago

Update the version and try again

cao-nv commented 1 week ago

Update the version and try again

I'm using version 2.2.0

yuyun2000 commented 1 week ago

It's very likely that you encountered an internal bug.

cao-nv commented 1 week ago

It's very likely that you encountered an internal bug.

E RKNN: [15:06:13.670] REGTASK: The bit width of field value exceeds the limit, target: lite, offset: 0x1014, shift = 0, limit: 0x7, value: 0x9
E RKNN: [15:06:13.670] REGTASK: The bit width of field value exceeds the limit, target: lite, offset: 0x1014, shift = 3, limit: 0x7, value: 0x9

These are errors message that I during the export

yuyun2000 commented 1 week ago

It has nothing to do with this. It should be that some operators are stuck in a loop due to internal operator optimization, so the exported model becomes very large.