I just wanted to add a custom class for testing.
I've train the model on windows using Darkflow and is working relatively fine, I'm able to test locally for unseen images, video and usb camera.
Now from DarkFlow, I exported PB like this
flow --model cfg/train.cfg --load -xxx --savepb where xxx is desired checkpoint
I have installed NCSDK (v1) into VBOX and appears to working fine, except some examples that are not built because USB toggling.
I tried to convert PB to GRAPH like this:
mvNCCompile ./tiny-yolo-voc-test.pb -s12 -o ./tiny-yolo-voc-test.graph
Initially, errors, but after upgrade TENSORFLOW to the same version as in DARFLOW (1.13.1) it worked.
Now trying with that GRAPH on Raspberry, error
Error - could not load tiny yolo graph file: /opt/movidius/graphs/tiny_yolo/tiny-yolo-voc-test.graph
Traceback (most recent call last):
File "/opt/test/demo-tool/run.py", line 420, in
File "/opt/test/demo-tool/run.py", line 325, in main
File "/opt/test/demo-tool/tiny_yolo_processor.py", line 46, in init
self._fifo_in, self._fifo_out = self._ty_graph.allocate_with_fifos(ncs_device, blob)
File "/usr/local/lib/python3.5/dist-packages/mvnc/mvncapi.py", line 615, in allocate_with_fifos
Your graph file is likely corrupt since you're using a previous version of the NCSDK. Reinstall the latest version of the NCSDK here. And then try to recompile the graph file.
Let me know if this was helpful