sample_embedding_folder2/0702651.txt

Ticket Name: TDA2: Some question about implement deep learning algorithm by ourselves

Query Text:
Part Number: TDA2 Hi, I found that if we want to implement deep learning algorithm by ourselves, we must need to use TIDL. But if i implement it directly without TIDL, and just run it by CPU, it seems that i can't do this? I know that TIDL can help us accelerate the algorithm, but i just want to know that if i don't use this library and run on CPU can i implement my neural network on this board ?

Responses:
Hi Chen-Wei, you can create new deep learning algorithm and integrate it in your usecase. You can check documentation at vision_sdk\docs to get idea how to add new links/chains and new usecases. But please note that if you implement your deep learning algorithm on CPU you might not get very good performance (for example TIDL runs on DSP or EVE which are better suited for such calculations). Regards, Yordan

Hi Yordan, i have this question because i modify the original usecase, replace the original function(canny) with tiny-YOLO and get the following result So, i think maybe i can't implement it on CPU. I think the best way to solve this problem is creating a new usecase rather than modifying the original usecase right? Thanks for your reply, Chen-Wei Lai

Hi Chen-Wei , I have pinged VisionSDK experts for further help. Regards, Yordan

Hi Yordan, There is no further reply for so many days, should i keep waiting or just open another thread for this issue? Thanks, Chen-Wei Lai

darket net is not suported by TIDL. You may create nw uses case with out using TIDl