Create compare.py
Browse files- compare.py +105 -0
compare.py
ADDED
|
@@ -0,0 +1,105 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import time
|
| 2 |
+
import torch
|
| 3 |
+
from transformers import MT5ForConditionalGeneration, T5ForConditionalGeneration, AutoTokenizer, MT5Tokenizer
|
| 4 |
+
from datasets import load_dataset
|
| 5 |
+
|
| 6 |
+
# Пути к моделям
|
| 7 |
+
mt5_path = "./mt5" # Локальная MT5 модель
|
| 8 |
+
byt5_path = "./unzipped_model" # Локальная или скачанная ByT5 модель
|
| 9 |
+
|
| 10 |
+
# Путь к данным
|
| 11 |
+
validation_file = "mt5_validation_data-1.jsonl"
|
| 12 |
+
|
| 13 |
+
# Загрузка моделей и токенизаторов
|
| 14 |
+
mt5_tokenizer = MT5Tokenizer.from_pretrained(mt5_path)
|
| 15 |
+
mt5_model = MT5ForConditionalGeneration.from_pretrained(mt5_path).eval()
|
| 16 |
+
|
| 17 |
+
byt5_tokenizer = AutoTokenizer.from_pretrained(byt5_path)
|
| 18 |
+
byt5_model = T5ForConditionalGeneration.from_pretrained(byt5_path).eval()
|
| 19 |
+
|
| 20 |
+
# Выбор устройства
|
| 21 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
| 22 |
+
mt5_model.to(device)
|
| 23 |
+
byt5_model.to(device)
|
| 24 |
+
|
| 25 |
+
# Загрузка валидационной выборки
|
| 26 |
+
dataset = load_dataset("json", data_files={"validation": validation_file})
|
| 27 |
+
val_data = dataset["validation"]
|
| 28 |
+
|
| 29 |
+
# Функция предсказания
|
| 30 |
+
def predict(model, tokenizer, text):
|
| 31 |
+
inputs = tokenizer(
|
| 32 |
+
text,
|
| 33 |
+
return_tensors="pt",
|
| 34 |
+
truncation=True,
|
| 35 |
+
padding=True,
|
| 36 |
+
max_length=256
|
| 37 |
+
).to(device)
|
| 38 |
+
|
| 39 |
+
outputs = model.generate(
|
| 40 |
+
**inputs,
|
| 41 |
+
max_length=64,
|
| 42 |
+
num_beams=5,
|
| 43 |
+
early_stopping=True
|
| 44 |
+
)
|
| 45 |
+
return tokenizer.decode(outputs[0], skip_special_tokens=True)
|
| 46 |
+
|
| 47 |
+
# Статистика
|
| 48 |
+
country_match = 0
|
| 49 |
+
city_match = 0
|
| 50 |
+
full_match = 0
|
| 51 |
+
mismatch_samples = []
|
| 52 |
+
|
| 53 |
+
start_time = time.time()
|
| 54 |
+
|
| 55 |
+
for example in val_data:
|
| 56 |
+
text = example["text"]
|
| 57 |
+
|
| 58 |
+
# Предсказания от MT5 и ByT5
|
| 59 |
+
mt5_pred = predict(mt5_model, mt5_tokenizer, text)
|
| 60 |
+
byt5_pred = predict(byt5_model, byt5_tokenizer, text)
|
| 61 |
+
|
| 62 |
+
def split_prediction(pred):
|
| 63 |
+
parts = pred.split(":", 1)
|
| 64 |
+
if len(parts) == 2:
|
| 65 |
+
return parts[0].strip(), parts[1].strip()
|
| 66 |
+
else:
|
| 67 |
+
return parts[0].strip(), "unknown"
|
| 68 |
+
|
| 69 |
+
mt5_country, mt5_city = split_prediction(mt5_pred)
|
| 70 |
+
byt5_country, byt5_city = split_prediction(byt5_pred)
|
| 71 |
+
|
| 72 |
+
if mt5_country == byt5_country:
|
| 73 |
+
country_match += 1
|
| 74 |
+
if mt5_city == byt5_city:
|
| 75 |
+
city_match += 1
|
| 76 |
+
if mt5_country == byt5_country and mt5_city == byt5_city:
|
| 77 |
+
full_match += 1
|
| 78 |
+
else:
|
| 79 |
+
mismatch_samples.append({
|
| 80 |
+
"text": text,
|
| 81 |
+
"mt5_prediction": f"{mt5_country}:{mt5_city}",
|
| 82 |
+
"byt5_prediction": f"{byt5_country}:{byt5_city}"
|
| 83 |
+
})
|
| 84 |
+
|
| 85 |
+
end_time = time.time()
|
| 86 |
+
total_time = end_time - start_time
|
| 87 |
+
num_examples = len(val_data)
|
| 88 |
+
time_per_example = total_time / num_examples if num_examples > 0 else 0
|
| 89 |
+
|
| 90 |
+
# Вывод различий
|
| 91 |
+
print("Примеры, где хотя бы что-то не совпало между MT5 и ByT5 (макс. 80):")
|
| 92 |
+
for i, item in enumerate(mismatch_samples[:80]):
|
| 93 |
+
print(f"\nПример {i+1}:")
|
| 94 |
+
print(f"Текст: {item['text']}")
|
| 95 |
+
print(f"MT5 предсказал: {item['mt5_prediction']}")
|
| 96 |
+
print(f"ByT5 предсказал: {item['byt5_prediction']}")
|
| 97 |
+
|
| 98 |
+
# Итоги
|
| 99 |
+
print("\nРезультаты сравнения MT5 vs ByT5:")
|
| 100 |
+
print(f"Всего примеров: {num_examples}")
|
| 101 |
+
print(f"Совпало стран: {country_match}")
|
| 102 |
+
print(f"Совпало городов: {city_match}")
|
| 103 |
+
print(f"Полных совпадений: {full_match}")
|
| 104 |
+
print(f"Общее время выполнения: {total_time:.4f} сек.")
|
| 105 |
+
print(f"Время на одно сравнение: {time_per_example:.6f} сек.")
|