I tried to work with ULMFiT with a highly imbalance dataset. After fine tuning the language model and in training the classifier; I tried to change the sampler of the training dataloader from SortishSampler to a weighted sampler and keep the validation dataloader with the default SortSampler . My question is simple. When I tried to predict the validation dataset without a sampler it gives a difference results (accuracy) from the one computed after each epochs.
I dont have an answer to your question but could you please tell me what is the particular fine tuning you are referring to while working on a highly imbalanced dataset for ULMFit? Thanks in advance!