DNN initialization Error : Simple NVCaffe pretrained model does not run on Driveworks


I am trying to run custom Caffe simple model on Drive PX2 using Driveworks DNN API, but failed.
I trained the simplified cifar10 model in NVCaffe and used the model and parameter in driveworks.

And I got the following Error Log

Network Initialization done.
Ignoring source layer cifar
Copying source layer ip1
Cannot initialize Caffe Network: DW_FAILURE

Looks like network model is loaded well,
but when trying to load weights, something error happened.

Hi, updating issue.

I installed pycaffe and used pycaffe interface to load my custom model as follows.

net = caffe.Net(‘cifar10.prototxt’, ‘cifar10_weights.caffemodel’, caffe.TEST);

My sample cifar10 like custom model loads find in nv-caffe 0.16 version showing following logs.

I0711 16:07:17.792974 3381 net.cpp:404] Network initialization done.
I0711 16:07:17.793540 3381 net.cpp:1067] Ignoring source layer cifar
I0711 16:07:17.793553 3381 net.cpp:1075] Copying source layer conv1 Type:Convolution #blobs=2
… …
I0711 16:07:17.793663 3381 net.cpp:1075] Copying source layer ip1 Type:InnerProduct #blobs=2
I0711 16:07:17.793696 3381 net.cpp:1075] Copying source layer ip2 Type:InnerProduct #blobs=2
I0711 16:07:17.793709 3381 net.cpp:1067] Ignoring source layer loss

And pycaffe also load driveworks dnn sample drivenet model predict.prototxt and weidhts.caffemodel successfully.

But when i try to use the custom model and weights in driveworks in drivepx2,
above issue happens.

Any ideas?

Hello kakpple,

Please let me know your using PDK version.
If you are using Beta 1.0, could you please update Beta 2.0?

If you are using Beta 2.0, could you please check below? Thanks.

  1. If you use TensorRT 2.0 RC on DrivePX2, could you please update TensorRT 2.0 GA via https://developer.nvidia.com/tensorrt?
  2. Could you please check the cifar10.prototxt has type : “Python” in any layer?
    If yes, please remove the layer and re-try it.