The project pipeline worked as following:
First we have to train the models on some good GPU since the smallest model (2-4 filter) already takes few hours on an NVIDIA Titan X .
Models are constructed by calling models.py
After training, they are saved as .h5
All inference files are saved in their respective folders "./WeedAI_Inference/Inferences/..."
Use script from_h5_to_frozen_graph.py to load the .h5 model and further convert it into flex_model
All converted frozen graphs are going to be saved within "./Models/Frozen_graphs/*"
Important Remark: The vanilla resnet is converted to flex model here. Thus, the exported frozen graph is already a flex model on full-precision mode (FP32)
Call predict_on_full_images.py to generate annotations, inference times and confusion matrices. You can give the command line argument -p FP16 to use half-precision or -p FP32 for full-precision. All results generated at this stage are saved within "./Results/seeds/*"