Commit ·
c8a375c
1
Parent(s): 5412e5c
add bloomz summarization
Browse files- app/config/config.py +3 -0
- app/inference/bloomz.py +55 -0
- app/main.py +3 -0
- app/models/models.py +4 -0
app/config/config.py
CHANGED
|
@@ -13,10 +13,13 @@ class Settings:
|
|
| 13 |
mBART_LORA_SUMMARIZATION_MODEL_PATH: str = os.getenv('mBART_LORA_SUMMARIZATION_MODEL_PATH')
|
| 14 |
mT5_BASE_MODEL_PATH: str = os.getenv('mT5_BASE_MODEL_PATH')
|
| 15 |
mT5_LORA_SUMMARIZATION_MODEL_PATH: str = os.getenv('mT5_LORA_SUMMARIZATION_MODEL_PATH')
|
|
|
|
|
|
|
| 16 |
TRANSFORMER_SUMMARIZATION_MODEL_PATH: str = os.path.join(ROOT_LEVEL_PATH, 'outputs/transformer/TransformerSummarizationModel.keras')
|
| 17 |
|
| 18 |
mBART_TOKENIZER_PATH: str = os.getenv('mBART_BASE_MODEL_PATH')
|
| 19 |
mT5_TOKENIZER_PATH: str = os.getenv('mT5_BASE_MODEL_PATH')
|
|
|
|
| 20 |
NEPALI_SUMMARIZATION_TOKENIZER: str = os.path.join(ROOT_LEVEL_PATH, 'outputs/transformer/tokenizer/summarization_52000.model')
|
| 21 |
|
| 22 |
settings = Settings()
|
|
|
|
| 13 |
mBART_LORA_SUMMARIZATION_MODEL_PATH: str = os.getenv('mBART_LORA_SUMMARIZATION_MODEL_PATH')
|
| 14 |
mT5_BASE_MODEL_PATH: str = os.getenv('mT5_BASE_MODEL_PATH')
|
| 15 |
mT5_LORA_SUMMARIZATION_MODEL_PATH: str = os.getenv('mT5_LORA_SUMMARIZATION_MODEL_PATH')
|
| 16 |
+
BLOOMZ_LORA_SUMMARIZATION_MODEL_PATH: str = os.getenv('BLOOMZ_LORA_SUMMARIZATION_MODEL_PATH')
|
| 17 |
+
BLOOMZ_BASE_MODEL_PATH: str = os.getenv('BLOOMZ_BASE_MODEL_PATH')
|
| 18 |
TRANSFORMER_SUMMARIZATION_MODEL_PATH: str = os.path.join(ROOT_LEVEL_PATH, 'outputs/transformer/TransformerSummarizationModel.keras')
|
| 19 |
|
| 20 |
mBART_TOKENIZER_PATH: str = os.getenv('mBART_BASE_MODEL_PATH')
|
| 21 |
mT5_TOKENIZER_PATH: str = os.getenv('mT5_BASE_MODEL_PATH')
|
| 22 |
+
BLOOMZ_TOKENIZER_PATH: str = os.getenv('BLOOMZ_BASE_MODEL_PATH')
|
| 23 |
NEPALI_SUMMARIZATION_TOKENIZER: str = os.path.join(ROOT_LEVEL_PATH, 'outputs/transformer/tokenizer/summarization_52000.model')
|
| 24 |
|
| 25 |
settings = Settings()
|
app/inference/bloomz.py
ADDED
|
@@ -0,0 +1,55 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 2 |
+
from peft import PeftModel
|
| 3 |
+
from app.config.config import settings
|
| 4 |
+
from app.preprocess.preprocess import nepali_perprocessing
|
| 5 |
+
import torch
|
| 6 |
+
import gc
|
| 7 |
+
|
| 8 |
+
torch.classes.__path__ = []
|
| 9 |
+
|
| 10 |
+
def tokenize(news: str, tokenizer, news_sequence_length: int=523, title_prompt_sequence_length: int=3):
|
| 11 |
+
title_prompt_ids = tokenizer(
|
| 12 |
+
[" Headline:"],
|
| 13 |
+
max_length=title_prompt_sequence_length,
|
| 14 |
+
padding=False,
|
| 15 |
+
truncation=True,
|
| 16 |
+
add_special_tokens=False
|
| 17 |
+
)
|
| 18 |
+
|
| 19 |
+
inputs = tokenizer(
|
| 20 |
+
news,
|
| 21 |
+
max_length=news_sequence_length,
|
| 22 |
+
add_special_tokens=True,
|
| 23 |
+
truncation=True,
|
| 24 |
+
padding=False
|
| 25 |
+
)
|
| 26 |
+
|
| 27 |
+
input_ids = inputs["input_ids"][0] + title_prompt_ids['input_ids'][0]
|
| 28 |
+
attn_mask = [1] * len(input_ids)
|
| 29 |
+
|
| 30 |
+
return {
|
| 31 |
+
"input_ids": torch.tensor([input_ids]),
|
| 32 |
+
"attention_mask": torch.tensor([attn_mask]),
|
| 33 |
+
}
|
| 34 |
+
|
| 35 |
+
def generate(text: str, input_sequence_length: int=512, output_sequence_length: int=26, title_prompt_sequence_length: int=3, num_beams: int=3, temperature: float=1, top_k: int=50):
|
| 36 |
+
model = AutoModelForCausalLM.from_pretrained(settings.BLOOMZ_BASE_MODEL_PATH)
|
| 37 |
+
tokenizer = AutoTokenizer.from_pretrained(settings.BLOOMZ_BASE_MODEL_PATH)
|
| 38 |
+
model = PeftModel.from_pretrained(model, settings.BLOOMZ_LORA_SUMMARIZATION_MODEL_PATH, token=settings.HF_ACCESS_TOKEN)
|
| 39 |
+
text = nepali_perprocessing(text)
|
| 40 |
+
inputs = tokenize(text, tokenizer, input_sequence_length, title_prompt_sequence_length)
|
| 41 |
+
generated_tokens = model.generate(
|
| 42 |
+
input_ids=inputs["input_ids"],
|
| 43 |
+
attention_mask=inputs["attention_mask"],
|
| 44 |
+
max_new_tokens=output_sequence_length,
|
| 45 |
+
num_beams=num_beams,
|
| 46 |
+
do_sample=True,
|
| 47 |
+
temperature=temperature,
|
| 48 |
+
top_k=top_k,
|
| 49 |
+
)
|
| 50 |
+
input_token_length = inputs["input_ids"].shape[1]
|
| 51 |
+
summary_tokens = generated_tokens[0][input_token_length:]
|
| 52 |
+
summary = tokenizer.decode(summary_tokens, skip_special_tokens=True).strip()
|
| 53 |
+
del model, tokenizer
|
| 54 |
+
gc.collect()
|
| 55 |
+
return summary
|
app/main.py
CHANGED
|
@@ -3,6 +3,7 @@ import os, uvicorn
|
|
| 3 |
from fastapi.middleware.trustedhost import TrustedHostMiddleware
|
| 4 |
from app.inference import mbart as mbart
|
| 5 |
from app.inference import mt5 as mt5
|
|
|
|
| 6 |
from app.middleware.auth import AuthMiddleware
|
| 7 |
from fastapi.middleware.cors import CORSMiddleware
|
| 8 |
from app.models.models import SummarizationRequest, SummarizationResponse
|
|
@@ -25,6 +26,8 @@ def summarize_nepali_news(body: SummarizationRequest) -> SummarizationResponse:
|
|
| 25 |
res = mbart.generate(body.text)
|
| 26 |
elif body.model == 'mt5':
|
| 27 |
res = mt5.generate(body.text)
|
|
|
|
|
|
|
| 28 |
else:
|
| 29 |
raise HTTPException(400, "Bad Request")
|
| 30 |
|
|
|
|
| 3 |
from fastapi.middleware.trustedhost import TrustedHostMiddleware
|
| 4 |
from app.inference import mbart as mbart
|
| 5 |
from app.inference import mt5 as mt5
|
| 6 |
+
from app.inference import bloomz as bloomz
|
| 7 |
from app.middleware.auth import AuthMiddleware
|
| 8 |
from fastapi.middleware.cors import CORSMiddleware
|
| 9 |
from app.models.models import SummarizationRequest, SummarizationResponse
|
|
|
|
| 26 |
res = mbart.generate(body.text)
|
| 27 |
elif body.model == 'mt5':
|
| 28 |
res = mt5.generate(body.text)
|
| 29 |
+
elif body.model == 'bloomz':
|
| 30 |
+
res = bloomz.generate(body.text, num_beams=body.num_beams, temperature=body.temperature, top_k=body.top_k)
|
| 31 |
else:
|
| 32 |
raise HTTPException(400, "Bad Request")
|
| 33 |
|
app/models/models.py
CHANGED
|
@@ -6,10 +6,14 @@ class SummarizationModels(str, Enum):
|
|
| 6 |
transformer = "transformer"
|
| 7 |
mbart = 'mbart'
|
| 8 |
mt5 = 'mt5'
|
|
|
|
| 9 |
|
| 10 |
class SummarizationRequest(BaseModel):
|
| 11 |
text: str
|
| 12 |
model: Optional[SummarizationModels] = 'mbart'
|
|
|
|
|
|
|
|
|
|
| 13 |
|
| 14 |
class SummarizationResponse(BaseModel):
|
| 15 |
prediction: str
|
|
|
|
| 6 |
transformer = "transformer"
|
| 7 |
mbart = 'mbart'
|
| 8 |
mt5 = 'mt5'
|
| 9 |
+
blooomz = 'bloomz'
|
| 10 |
|
| 11 |
class SummarizationRequest(BaseModel):
|
| 12 |
text: str
|
| 13 |
model: Optional[SummarizationModels] = 'mbart'
|
| 14 |
+
num_beams: Optional[int] = 3
|
| 15 |
+
temperature: Optional[float] = 1.0
|
| 16 |
+
top_k: Optional[int] = 50
|
| 17 |
|
| 18 |
class SummarizationResponse(BaseModel):
|
| 19 |
prediction: str
|