search

IBM / federated-learning-lib

A library for federated learning (a distributed machine learning process) in an enterprise environment.
Other
498 stars 135 forks source link

YOLO #82

Open fernando080 opened 3 years ago

fernando080 commented 3 years ago

Hi from Spain! I am building a representation of a federated architecture using docker with the objective of training YOLO with COCO dataset in a federated manner for my master's thesis. Can you give me some hints or recommendations to do it with you tool?

Thank you very much! Best Regards, Fernando...

Yi-Zoey commented 3 years ago

Hi @fernando080, what's the underlying ML library for YOLO? Currently, we support TensorFlow/Keras, PyTorch, RLlib, and Scikit-learn.

fernando080 commented 3 years ago

My YOLO version is built in tebsorflow/keras. Thank you very much for your answer ^^

Yi-Zoey commented 3 years ago

If you try to use one of our fusion algorithms to train, for example, FedAvg, you can modify the corresponding file to define your own model specification. If you are using TensorFlow 1.5 or just Keras library, you can use KerasFLModel model class, so you will need to modify this file according to your use case. If you are working with Tensorflow 2.2 or 2.4, you will use TensorFlowFLModel model class, and modify this file accordingly.

Moreover, regarding the COCO dataset, you will also need your own data handler and modify the data section in the configs to let IBM FL load this custom dataset properly. You can find more details about custom data handler in our tutorials.