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

Add loading weights and itos, so that we can extend training

Browse Source
This commit is contained in:
Piotr Czapla
2018-11-16 23:48:22 +01:00
parent 895a58c88d
commit 5cc4bfd444
1 changed files with 23 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ulmfit/pretrain_lm.py
+23 -15
View File
@@ -92,23 +92,26 @@ def pretrain_lm(dir_path, lang='en', cuda_id=0, qrnn=True, subword=False, max_vo
val_tok = val_tok[:max(20, int(len(val_tok) * ds_pct))]
print(f"Limiting data sets to {ds_pct*100}%, trn {len(trn_tok)}, val: {len(val_tok)}")
# create the vocabulary
cnt = Counter(word for sent in trn_tok for word in sent)
itos = [o for o,c in cnt.most_common(n=max_vocab)]
itos.insert(1, PAD) #  set pad id to 1 to conform to fast.ai standard
assert UNK in itos, f'Unknown words are expected to have been replaced with {UNK} in the data.'
stoi = {w: i for i, w in enumerate(itos)}
itos_fn=dir_path / model_dir / f'itos_{name}.pkl'
if not itos_fn.exists():
# create the vocabulary
cnt = Counter(word for sent in trn_tok for word in sent)
itos = [o for o,c in cnt.most_common(n=max_vocab)]
itos.insert(1, PAD) #  set pad id to 1 to conform to fast.ai standard
assert UNK in itos, f'Unknown words are expected to have been replaced with {UNK} in the data.'
# save vocabulary
print(f"Saving vocabulary as {dir_path / model_dir}")
results['itos_fname'] = itos_fn
with open(results['itos_fname'], 'wb') as f:
pickle.dump(itos, f)
else:
print("Loading itos:", itos_fn)
itos = np.load(itos_fn)
vocab = Vocab(itos)
stoi = vocab.stoi
# save vocabulary
print(f"Saving vocabulary as {dir_path / model_dir}")
results['itos_fname'] = dir_path / model_dir / f'itos_{name}.pkl'
with open(results['itos_fname'], 'wb') as f:
pickle.dump(itos, f)
trn_ids = np.array([([stoi.get(w, stoi[UNK]) for w in s]) for s in trn_tok])
val_ids = np.array([([stoi.get(w, stoi[UNK]) for w in s]) for s in val_tok])
@@ -152,10 +155,15 @@ def pretrain_lm(dir_path, lang='en', cuda_id=0, qrnn=True, subword=False, max_vo
else:
learn.metrics = [accuracy]
try:
learn.load(f'{model_name}_{name}')
print("Weights loaded")
except FileNotFoundError:
print("Starting from random weights")
pass
learn.fit_one_cycle(num_epochs, 5e-3, (0.8, 0.7), wd=1e-7)
if not subword and max_vocab is None:
# only if we use the unpreprocessed version and the full vocabulary
# are the perplexity results comparable to previous work