-
Notifications
You must be signed in to change notification settings - Fork 5
neuralnet
Brett Graham edited this page Apr 6, 2020
·
1 revision
The neural network is a version of MobileNetv2 (224x224 input size) pre-trained on image net and then trained on a subset of images from the 2018 inaturalist dataset. The subset was selected by removing any:
- genus not observed in the US (GBIF query returns at least 1 occurance in US)
- genus not explicitly ignored
All images for species belonging to the same genus were combined resulting in 2988 categories. Some training parameters are:
- 32 image batch size
- 80/20 training/validation split
- 2D global avg pooling and dense layers added on top to adapt model
- Nadam optimizer with sparse_categorical_crossentry loss and sparse_categorical_accuracy
- 0.1 epsilon, 0.001 learning rate
- 20 training epochs
The trained network was then quantized to allow inference on a Google Coral EdgeTPU USB accelerator. Finally, the model is 'served' via a shared memory interface to allow multiple processes to access the USB accelerator.