First of all, I’d like to thank the entire team for creating this wonderful project. Great job.
I’d like both suggestions and information from you. I have a client who wants to detect a liquid that glows under UV light. I prepared a small dataset and converted it to a .hef file in DeGirum AI Hub. It works very well in Hailo8. My first question is, can we add a special property to the PRE_process section of the .json files when converting to a .hef file, or is this entirely a pySDK feature, or is it necessary? My second question is, my dataset is small, and I plan to create a human-verified dataset. Do I have to recreate the .hef file every time? It doesn’t exist. Is there such a process for fine-tuning a .pt file?
Hi @onryldrm016
Welcome to the DeGirum community. Thank you for your kind words.
Can you explain a little bit more on what special property you want to add to PRE_PROCESS section? Currently, it is used to specify various options in preprocessing the input. It is possible that what you want to do is already supported. If you share more details, we can help. Also, the preprocess section is added after compile and is not related to generating the hef.
yes, every time you have a newly trained checkpoint you need to compile it to hef again.
Hi, what I’d like to do is filter out some tones without actually feeding the data to the model.
If I set this in your pre_process and feed the frame to the model, does your SDK do this, or should I use cv2 to filter the frame and then feed it to the model?
I’d like to try training a model by suppressing the blue tones of UV light, and I’d hope to get better results if I apply the same filter to the incoming frame.
Hi @onryldrm016
You can do the pre-processing outside and send to our predict function. We can show you how to do it in an optimal way from a performance point of view, if you share a code snippet of the preprocessing.