can not reproduce
dkqkxx opened this issue · 3 comments
provide hyperparameters to achieve results in paper pls
Thank you for your interest in our work. The best hyperparameters on the MELD dataset are in config.py, It can produce a w-F1
of 67.52.
As we describe in Limitations, SPCL involves too much randomness, we do grad search for key hyper-parameters:
epochs we use epochs to control the difficulty.
psz size of representations queue
ssz size of support set
temp the temperature used in contrastive loss
This can be done on a single V100 GPU.
can you provide hyperparameter for IEMOCAP dataset? i can only reproduce results in paper for EmoryNLP and MELD, for IEMOCAP, i used hyperparameter in config.py and weighted-f1 result of IEMOCAP is around 67.17
can you provide hyperparameter for IEMOCAP dataset? i can only reproduce results in paper for EmoryNLP and MELD, for IEMOCAP, i used hyperparameter in config.py and weighted-f1 result of IEMOCAP is around 67.17
Sorry, I only kept the hyper-parameters for MELD, for IEMOCAP you may need run grad search again... :(, I have finished my internship, so I don't have enough computing resources to help you reproduce.