Closed danschu closed 3 years ago
@danschu It looks like you model's outputs of size (512X512X151) is just too large to fit on the edgetpu's RAM. I suggest shrinking the model output size and try compiling again
I removed the last upscaling layer, but I still get the same error.
@Namburger: Can you please check the last attached file? I have removed the 512x512x151 layer, but I get the same error. Thank you :)
@danschu hey sorry, I forgot to update, this seems similar to an issue we've seen before and I've attached the model on the bug to have it looked at but haven't got a response yet. I'll ping it again today and will keep you updated
@Namburger Any news on the model? Is it a compiler bug? If so, when do we get a new version of the compiler?
@danschu hey, so sorry for the late response, unfortunately it is still the size limitation which we currently don't have a way to overcome via software. There are a couple of ops with inputs resolution of 65x65x960 and 257x257x96, these inputs are just a little too large, could you try shrinking those too? Apologies, I understand that could cause much precision drop
@Namburger I used the new compiler (v15.0) and splited the model into many parts. The problem is on the attached part of the model. I think that is not the input resolution. I still get an "internal compile error". Can you have a look on it?
Thank you mymodel_segment_21_of_24.zip
@Namburger Can you confirm this as a compiler bug?
Feel free to reopen if the issue still persists.
Edge TPU Compiler version 14.1.317412892
Model: blubuuu.zip
Based on https://tfhub.dev/sayakpaul/lite-model/deeplabv3-mobilenetv2-ade20k/1/default/1 with post training quantization.