daquexian / dnntools Goto Github PK

Hi, I downloaded squeezenet caffe model from https://github.com/DeepScale/SqueezeNet, and
Try to convert it to .daq model following the instructions you wrote.
But the transformation terminated with error message "core dumped"
Below is the last rows of printed message.

Is there any way I can check where the cause is ?
Thanks.

PS. I use Ubuntu 16.04LTS, python3.5, and Caffe has been successfully installed.
No GPU is in my notebook, I use CPU only.

I0518 15:20:05.292441 10474 net.cpp:242] This network produces output prob
I0518 15:20:05.292491 10474 net.cpp:255] Network initialization done.
I0518 15:20:05.296623 10474 net.cpp:744] Ignoring source layer label_data_1_split
I0518 15:20:05.297698 10474 net.cpp:744] Ignoring source layer pool10_pool10_0_split
I0518 15:20:05.297741 10474 net.cpp:744] Ignoring source layer loss
I0518 15:20:05.297746 10474 net.cpp:744] Ignoring source layer accuracy
I0518 15:20:05.297750 10474 net.cpp:744] Ignoring source layer accuracy_top5
程式記憶體區段錯誤 (core dumped)

Dilated conv can be implemented by shift-and-stitch, which relys on strided slice introduced in Android P SPACE_TO_BATCH

Currently some layers with unsupported parameters will not be refused to convert.

Hi
When I converted a pretrained model, an error message showed "Only depthwise convolution or
vanilla convolution are supported".
I found the message at line 118 in caffe2daq.py and check my deploy.prototxt of that pretrained model.
It should be "group: 2" parameter setting of the convolution layer causing the converting failure.

My question is : Is this also the limitation of NNAPI ? or User can additionally write code for this case and also other downward connections ?
Thanks in advance.