I am trying to reproduce your method on VVEnc, I just use your training code and the same network, except that I generate residual with IME for each 8x8 in the 64x64 and use slice qp instead of sequence qp.
Then I try to predict with your provided model, both LDP-CNN and LDP-LSTM. For 32x32 block, setting threshold to 0.5, LDP-CNN model can get about 75% accuracy, while LSTM get a very low accuracy. I also try to train LDP-CNN and LDP-LSTM with my data, similarly, LDP-CNN can get about 75% accuracy, but LSTM get about 60%.
I use the same training and validation dataset for LDP-CNN and LDP-LSTM, I just don't know what is the problem with LSTM? Is LSTM a good method for prediction from your point of view?
Besides, how do you calculate accuracy? The average accuracy for all the samples, or the average of negative sample accuracy and positive sample accuracy?
Hi Tianyi,
I am trying to reproduce your method on VVEnc, I just use your training code and the same network, except that I generate residual with IME for each 8x8 in the 64x64 and use slice qp instead of sequence qp.
Then I try to predict with your provided model, both LDP-CNN and LDP-LSTM. For 32x32 block, setting threshold to 0.5, LDP-CNN model can get about 75% accuracy, while LSTM get a very low accuracy. I also try to train LDP-CNN and LDP-LSTM with my data, similarly, LDP-CNN can get about 75% accuracy, but LSTM get about 60%.
I use the same training and validation dataset for LDP-CNN and LDP-LSTM, I just don't know what is the problem with LSTM? Is LSTM a good method for prediction from your point of view?
Besides, how do you calculate accuracy? The average accuracy for all the samples, or the average of negative sample accuracy and positive sample accuracy?
Looking for your reply, best regards !!!