I am trying use you code, however even after 5 epochs the accuracy is not imporving. The acc_e is 20% and acc_g is 0% from the begining of the training.
I have done minor changes related to dataset and num_classes per episode. I am using only voxceleb1 for training due to limited resources. I kept num_claases to 5 which I observe in many of meta leraning papers related to face recognition.
Moreover, I want to understand the training stategy for meta learning. Even using only voxceleb1 created approximately 10000 episodes per epoch which consumes 2 hours of time to train on GPU. If I include the voxceleb2, num_episodeds can reach to more than 100000. How much time it has taken for you to train one epoch (definitely it depends on the capacity of resource, so if you could give high level specification of resources used that would be really beneficial)? Any suggestion on speeding up the training?
Thanks for open-sourcing implementation.
I am trying use you code, however even after 5 epochs the accuracy is not imporving. The acc_e is 20% and acc_g is 0% from the begining of the training.
I have done minor changes related to dataset and num_classes per episode. I am using only voxceleb1 for training due to limited resources. I kept num_claases to 5 which I observe in many of meta leraning papers related to face recognition.
Moreover, I want to understand the training stategy for meta learning. Even using only voxceleb1 created approximately 10000 episodes per epoch which consumes 2 hours of time to train on GPU. If I include the voxceleb2, num_episodeds can reach to more than 100000. How much time it has taken for you to train one epoch (definitely it depends on the capacity of resource, so if you could give high level specification of resources used that would be really beneficial)? Any suggestion on speeding up the training?