from peft import PeftModel, PeftConfig
from transformers import AutoModelForTokenClassification
config = PeftConfig.from_pretrained("bite-the-byte/byt5-small-deASCIIfy-TR")
model = AutoModelForTokenClassification.from_pretrained("google/byt5-small")
model = PeftModel.from_pretrained(model, "bite-the-byte/byt5-small-deASCIIfy-TR")
def test_mask(device, sample):
"""
Masks the padded tokens in the input.
Args:
data (list): List of strings.
Returns:
dataset (list): List of dictionaries.
"""
tokens = dict()
input_tokens = [i + 3 for i in sample.encode('utf-8')]
input_tokens.append(0) # eos token
tokens['input_ids'] = torch.tensor([input_tokens], dtype=torch.int64, device=device)
# Create attention mask
tokens['attention_mask'] = torch.ones_like(tokens['input_ids'], dtype=torch.int64, device=device)
return tokens
def rewrite(model, data):
"""
Rewrites the input text with the model.
Args:
model (torch.nn.Module): Model.
data (dict): Dictionary containing 'input_ids' and 'attention_mask'.
Returns:
output (str): Rewritten text.
"""
with torch.no_grad():
pred = torch.argmax(model(**data).logits, dim=2).squeeze(0)
output = list() # save the indices of the characters as list of integers
# Conversion table for Turkish characters {100: [300, 350], ...}
en2tr = {en: tr for tr, en in zip(list(map(list, map(str.encode, list('ÜİĞŞÇÖüığşçö')))), list(map(ord, list('UIGSCOuigsco'))))}
for inp, lab in zip((data['input_ids'].squeeze(0) - 3).tolist(), pred.tolist()):
if lab and inp in en2tr:
# if the model predicts a diacritic, replace it with the corresponding Turkish character
output.extend(en2tr[inp])
elif inp >= 0: output.append(inp)
return bytes(output).decode()
def try_it(text, model):
sample = test_mask(model.device, text)
return rewrite(model, sample)
try_it('Cekoslovakyalilastiramadiklarimizdan misiniz?', model)
- Downloads last month
- 24
Inference API (serverless) does not yet support peft models for this pipeline type.