fast.ai ULMFiT¶

from fastai.text import *

twitter_data_path = "."

data_lm = (TextList
           .from_csv("./twitter-data/", 'train-processed.csv', cols=5)
           .split_by_rand_pct()
           .label_for_lm()
           .databunch(bs=32))

batchsize = 32

learn = language_model_learner(data_lm, AWD_LSTM, drop_mult=0.3)

learn.lr_find()

learn.recorder.plot()

learn.fit_one_cycle(10, 1e-2)

learn.save_encoder('fine_tuned_enc')

data_class = (TextList
           .from_csv(twitter_data_path, 'train-processed.csv', cols=5, vocab=data_lm.vocab)
           .split_by_rand_pct()
           .label_from_df(cols=0)
           .databunch())

twitter_classifer_learner = text_classifier_learner(data_class, AWD_LSTM, drop_mult=0.5)
twitter_classifer_learner.load_encoder('fine_tuned_enc')

twitter_classifer_learner.lr_find()

twitter_classifer_learner.recorder.plot()

twitter_classifer_learner.fit_one_cycle(5, 1e-3)

twitter_classifer_learner.freeze_to(-2)

twitter_classifer_learner.fit_one_cycle(1, 1e-3)