12-14-2018 06:04 AM
I've tried quite a few things with the DNNDK kit, but have trouble getting some other model to run on it; The decent or dnnc tools fail at some point due to the tools not handling the layers I provide. So I wonder if you could provide some better guidance on steps to take to get around these issues. I'll provide one example. I followed this guide, and intended to run this model on DNNDK, but get this issue :
*** Deephi DNNDK could not support LRN Layer, (pool1/norm1), please delete it or replace it with BatchNorm + Scale Layer and retrain.
Here is the guide I tried to follow :
The question is - how would I modify the train_val.prototxt to remove LRN and replace it with BatchNorm + Scale? And why doesn't DNNDK support LRN?
I assume that if succesful, I would modify train_val.prototxt so uses BatchNorm and Scale instead of LRN, then run training and finally I might be able to run decent and dnnc on my Caffe model.
Hope you can help - I'm sure others have the same questions.
12-17-2018 11:19 PM
@dsskjelbMay I know the model you want to run through DNNDK?
BTW, the reason DNNDK does not support LRN is that it is seldom used in the modern networks. Most of the networks use Batchnorm which removes the need for LRN.
04-24-2019 03:02 AM
04-24-2019 03:13 AM
I realized I didn't answer the question regarding what model I tried to run. I tried various already trained caffee models - e.g. dogs vs cats pretrained model that I found in model zoo I believe it was (https://github.com/BVLC/caffe/wiki/Model-Zoo) But I had issues getting any of the model zoo models to run on DNNDK because it didn't support the now outdated layers. So I had to modify the model and retrain on my laptop just to get it deployed on DNNDK.
So testing out a pre-trained model on DNNDK was much more difficult than I anticipated.