In one of the lectures, Jeremy mentioned that we can improve the model performance by fixing the high loss data points and training the model with a lower learning rate. I am wondering how to do this.
learn = text_classifier_learner(data_clas, AWD_LSTM, drop_mult=0.4, callback_fns=[csv_logger]) learn.load_encoder(os.path.join(data_enc_path, data_enc_name)) learn.fit_one_cycle(cycle, 1e-2, moms=(0.8,0.7))
I have exported the above using
learn.export(file = ‘labeled_only_model.pkl’)
I would like to load a new data clas with just corrected high loss file and the previously exported learn (labeled_only_model.pkl) and train the again. How do I correctly load the weights and the most recent vocab correctly?