-
Notifications
You must be signed in to change notification settings - Fork 579
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
nvdla_small running on FPGA, wrong inference compared to golden results #254
Comments
@nookfoo hi, did you use the flatbuff under sw? Did you boot a linux4.3.3 on the ARM and use the KMD and UMD? Sorry, I think it's a stupid question. I know for now if you do not hack the sofeware routine, it's a must. And I also found your steps in another post. |
@shgangchen Yeah i used flatbuff under sw. I think I used petalinux 2018.2, to boot on xilinx zcu102. I had to cross compile kmd kernel before I could insert it on Petalinux. Don't worry about the questions, I also had no clue in the start. I found this thesis project about nvdla on zynq ultrascale where most of the setup steps are described in detail. Maybe it will help: https://webthesis.biblio.polito.it/9033/ |
@nookfoo thanks for your reply. So there is another question. As far as I know, the software of NVDLA under the sw reposity is currently only supporting the nvdla_full version which is the nvdla1. So how can you run the kmd and umd for the nvdla_small configuration (I think "nvdla_small" you meant is that under "master" branch)? Any modifications? I think this information is quite exciting. |
Hi @shgangchen, the current KMD and UMD actually does have support(*) for nv_small - what is missing for public is the compiler that would generate those flatbufs (like NN_L0_1_small_flatbuf). The available flatbuffs come from here: https://github.com/nvdla/sw/tree/master/regression/flatbufs/kmd From other threads I learned that compiler for nv_small should be released by end of Q1. (*) as for KMD - compile with |
@mmaciag thanks. I think your explaination is valuable for people like me. Thanks a lot, and I will try it firstly on the virtual platform and then FPGA. Lots of work to do. |
@nookfoo Thank you very much. |
@nookfoo @shgangchen @mmaciag Dead here, can't keep running. What could be the cause of this problem? Has anyone ever had the same problem? How to fix it. |
Same issue,is there a solution? |
I am running nvdla_small spec on Xiliinx Ultrascale Board ZCU102 and am able to successfully pass NN_L0_1_small_flatbuf test. Unfortunately the resulting output.dimg has a different md5 hash compared to golden result /o_000000.dimg
As I understand, this means my inference results are wrong or deviating. Is this to be expected since inference results vary depending on hardware architecture?
Would appreciate if anyone has ideas on what causes this or how to attempt to troubleshoot this case.
Thanks and happy holidays!
The text was updated successfully, but these errors were encountered: