Closed jbericat closed 2 years ago
Tota la recerca necessària es troba al llibre de DL de Mohamed Elgendy (Deep Learning for Vision Systems), capítols 4 i 5 quasi sencers.
S'han generat els issues #80, #82 i #83, dels quals ja hi ha un esborrany complet enfocat a la PoC (en paper) i queda pendent acabar d'estudiar el capítol 5 per poder fer la #81
S'actualitza el document de la memòria parcial FITA#05_ESBORRANY.PDF i de moment es pot donar la recerca per acabada (si cal fer-ne més ja es reobrirà aquest issue)
Es reobre issue ja que cal actualitzar la secció de recerca del document de memòria parcial de la FITA#5 amb info de les noves TASK creades per a la fita.
Fer recerca dels acceleradors de deep learning TDA4VM amb CortexA9 i els models DL que pot manegar (Zoo, etc). Podria ser interessant considerar que els drons porten aquests acceleradors incorporats al companion-computer (dispositiu IoT / Edge Computing corrent TI-RTOS o un equivalent open source) que els permeten realitzar la detecció de característiques en temps real durant la explotació del model.
Info:
Hello everyone,
Hope you are all doing fine. In case, you have missed the information on our E2E forums, I wanted to pass the information that our next Edge AI monthly webinar is next week on Dec 9th 2021. Based on the questions and interest we are seeing from developers, this month we will focus on Yolov5, a popular object detection deep learning model, deployment on TDA4x platform.
We are excited to have Debapriya Maji, who will walk through the steps of optimizing deep learning models to take advantage of TI deep learning hardware accelerator in TDA4VM processor using industry standard APIs. With Yolov5 as an example, we will show step-by-step approach of doing minor modification to the model architecture such that it runs efficiently and at best performance on TDA4x processor without compromising on accuracy.
Webinar outline: •TI Model Zoo •Yolov5 model overview •Optimizations for TI’s deep learning accelerator •Compiling the model using open-source run time •Model benchmarking on Free Cloud tool
We have two time zones planned for this webinar and the registration links are in the E2E post below.
Looking forward to seeing you!
Best regards,
Srik Gurrapu
Jacinto SW Applications
per a la construcció del model es seguirà el següent tutorial:
https://docs.microsoft.com/en-us/windows/ai/windows-ml/tutorials/pytorch-train-mod
Per a l'adaptació del model a les condicions del dataset, així com per a la seva optimització, el següent enllaç proporciona informació valuosa sobre casos d'ús concrets:
https://www.pyimagesearch.com/2021/05/14/convolutional-neural-networks-cnns-and-layer-types/
Tasques de recerca finalitzades, resta pendent documentar
Receca finalitzada i documentació actualitzada
Tasques de recerca i document de memòria parcial actualitzat
cal documentar quines tasques de recerca s'efectuen i incloure links, keywords de cerca a Google, etc