Developing and Deploying Your Custom Action Recognition Application Without Any AI Expertise Using NVIDIA TAO and NVIDIA DeepStream

Originally published at:

Build an action recognition app with pretrained models, the TAO Toolkit, and DeepStream without large training data sets or deep AI expertise.

1 Like

I was trying to recreate the pre-trained model. I trained the net with the same 5 classes (ride bike, walk, fall floor, run and push) and the default hyperparameters. Then I use this model as a pre-trained model and train again with different classes. However, the accuracy and F1 I get with my recreated model don’t even come close to the ones I get with the downloaded pre-trained model. I thought this could be cause the hyperparameters used to train the model were different from the ones I used, but I couldn’t find any information on how the pre-trained model was trained.
I would highly appreciate if anyone could provide me with this information. The hyperparameters I am referring to are the ones in the train_rgb_3d_finetune.yaml file, such as the learning rate, epochs, momentum, weight_decay, dropout ratio, steps, batch size etc…

Thank you!