is NVCaffe compatible with Caffe?

I got a pretrained model with DIGITS:19.03-caffe. I could run it with NVCaffe:19.03-py2. But it was not run with Caffe, which has been installed with conda. (caffe-gpu 1.0 py27h999e8d7_5). I matched the google protobuf version to 3.6.0.

Here are snippets from logs when executing ‘net = caffe.Classifier(MODEL, PRETRAINED, channel_swap=(2,1,0), raw_scale=255)’


I0325 22:36:48.378525 802 net.cpp:255] Network initialization done.
I0325 22:36:48.469805 802 net.cpp:744] Ignoring source layer train-data
F0325 22:36:48.469831 802 blob.cpp:496] Check failed: count_ == proto.data_size() (34848 vs. 0)
*** Check failure stack trace: ***
[I 22:36:51.386 NotebookApp] KernelRestarter: restarting kernel (1/5), keep random ports
WARNING:root:kernel e258c4df-120c-4b8d-ba7c-964a5d101d8b restarted

Is NVCaffe compatible with Caffe?

To run NVCaffe models on Caffe you’d need to use

store_blobs_in_old_format: true

when you train and snapshot.
See https://github.com/NVIDIA/caffe/blob/caffe-0.17/src/caffe/proto/caffe.proto#L514-L519 for details