wassname/multifit
1
0
Fork 0
mirror of https://github.com/wassname/multifit.git synced 2026-09-09 11:27:26 +08:00
Code Issues Packages Projects Releases Wiki Activity

Working version of BILM - probably won't train well yet

Browse Source
This commit is contained in:
Piotr Czapla
2018-11-14 14:55:56 +01:00
parent a85800610b
commit 5ba83b1d4e
3 changed files with 74 additions and 27 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/test_text_train.py
+12 -13
View File
@@ -43,23 +43,22 @@ def learn():
def test_val_loss(learn):
assert learn.validate()[1] > 0.5
def test_bwdlm_lstm_can_be_trained():
manual_seed()
path, df_trn, df_val = prep_human_numbers()
data = TextLMDataBunch.from_df(path, df_trn, df_val, tokenizer=Tokenizer(BaseTokenizer),
lm_type = contrib_data.LanguageModelType.BiLM,
ld_cls = contrib_data.LanguageModelLoader)
learn = bilm_learner(data, emb_sz=100, nl=1, drop_mult=0.1, qrnn=False)
learn.fit_one_cycle(4, 5e-3)
assert learn.validate()[1] > 0.5
def test_bilm_lstm_can_be_trained():
manual_seed()
path, df_trn, df_val = prep_human_numbers()
data = TextLMDataBunch.from_df(path, df_trn, df_val, tokenizer=Tokenizer(BaseTokenizer),
lm_type = contrib_data.LanguageModelType.BwdLM,
ld_cls = contrib_data.LanguageModelLoader)
lm_type = contrib_data.LanguageModelType.BiLM)
learn = bilm_learner(data, emb_sz=100, nl=1, drop_mult=0.1, qrnn=False)
learn.metrics = []
learn.fit_one_cycle(4, 5e-3)
assert learn.validate()[0] < 2 #TODO Change to accuracy once it is fixed
def test_bwdlm_lstm_can_be_trained():
manual_seed()
path, df_trn, df_val = prep_human_numbers()
data = TextLMDataBunch.from_df(path, df_trn, df_val, tokenizer=Tokenizer(BaseTokenizer),
lm_type = contrib_data.LanguageModelType.BwdLM)
learn = language_model_learner(data, emb_sz=100, nl=1, drop_mult=0.1, qrnn=False)
learn.fit_one_cycle(4, 5e-3)