I have a Google coral micro dev board and want to use it to identify food dishes like pasta, omelet, etc. I used the existing object detection model (tf2_ssd_mobilenet_v2_coco17) used in its “detect_object” example and retrained it for a custom dataset. My dataset has 6 classes and ~100 images for each class. The compiled edgetpu model size is ~5.6 Mb. I am able to now run the “detect_objects” example using my model and it gives me good results. However, I increased the number of classes to 8 and added about ~100 images for each new class. Now, my compiled edgetpu model size is ~5.8 Mb. Now, when I try to run the “detect_objects” example using my new model (with 8 classes), I get the following error on my serial console:
console error text: Bad Bulk InTransfer Internal
To check if this was a connection issue or my model/”detect_objects” example issue, I trained other object detection models with a larger number of classes (greater than 6 classes) and the resulting edgetpu compiled models are comparatively larger in size (greater than 5.6 Mb). I noticed that the error pops up on any object detection edgetpu compiled model over the size of 5.6 Mb and goes away when the size is lesser than or equal to 5.6 Mb.
So, is a model size limit for the google coral micro dev board? Any tips for reducing the size of my model?