A-sentiment / app.py
Avisl's picture
Update app.py
72e769d verified
Raw
History Blame Contribute Delete
12.7 kB
import os
import traceback
from typing import Optional, Dict, Tuple, Any, List
import gradio as gr
from inference import (
clean_text,
load_sentiment_model,
infer_sentiment_single,
_SENTIMENT_NEGATIVE_THRESHOLD,
_SENTIMENT_POSITIVE_THRESHOLD,
)
SENTIMENT_REPO = "DanielNRU/aiforever-rubert-large-tbsa-avito-sentiment"
USE_CUDA_ENV = os.getenv("USE_CUDA", "1")
USE_CUDA = USE_CUDA_ENV == "1"
# ── Нормализация ──────────────────────────────────────────────────────────────
def normalize_tone_label(value: Any) -> str:
if value is None:
return "—"
text = str(value).strip().lower()
if not text or text in {"—", "-", "none", "null"}:
return "—"
if any(x in text for x in ["негатив", "negative", "neg"]):
return "негатив"
if any(x in text for x in ["нейтрал", "neutral", "neu"]):
return "нейтрально"
if any(x in text for x in ["позитив", "positive", "pos"]):
return "позитив"
if text in {"-1", "-1.0"}:
return "негатив"
if text in {"0", "0.0"}:
return "нейтрально"
if text in {"1", "1.0"}:
return "позитив"
return "—"
def normalize_prob_value(v: Any) -> Optional[float]:
try:
x = float(v)
except Exception:
return None
if x > 1.0:
x = x / 100.0
if 0.0 <= x <= 1.0:
return x
return None
def normalize_prob_dict(probs: Any) -> Dict[str, float]:
result: Dict[str, float] = {}
if not isinstance(probs, dict):
return result
for k, v in probs.items():
key = normalize_tone_label(k)
if key == "—":
continue
val = normalize_prob_value(v)
if val is None:
continue
result[key] = val
return result
def choose_tone_from_probs(tone_probs: Dict[str, float]) -> str:
if not tone_probs:
return "—"
best_label = None
best_prob = -1.0
for key in ["негатив", "нейтрально", "позитив"]:
val = tone_probs.get(key)
if val is not None and val > best_prob:
best_prob = val
best_label = key
return best_label if best_label is not None else "—"
def build_probs_block(tone_probs: Dict[str, float]) -> str:
if not tone_probs:
return "—"
lines = []
for key in ["негатив", "нейтрально", "позитив"]:
if key in tone_probs:
try:
lines.append(f"{key}: {tone_probs[key] * 100:.1f}%")
except Exception:
continue
return "\n".join(lines) if lines else "—"
# ── Логика порогов neg_thr / pos_thr ─────────────────────────────────────────
def _apply_thresholds(
tone_probs: Dict[str, float],
neg_thr: Optional[float] = None,
pos_thr: Optional[float] = None,
) -> str:
"""Определяет тональность по раздельным порогам neg_thr / pos_thr.
Логика (A-модель всегда работает в режиме neg/pos, без tone_thr):
- P(neg) >= neg_thr → 'негатив'
- P(pos) >= pos_thr → 'позитив'
- иначе → 'нейтрально'
Fallback при отсутствии порогов — argmax.
"""
if not tone_probs:
return "—"
p_neg = tone_probs.get("негатив", 0.0)
p_pos = tone_probs.get("позитив", 0.0)
if neg_thr is not None or pos_thr is not None:
_neg_thr = neg_thr if neg_thr is not None else 1.0
_pos_thr = pos_thr if pos_thr is not None else 1.0
if p_neg >= _neg_thr:
return "негатив"
if p_pos >= _pos_thr:
return "позитив"
return "нейтрально"
# fallback argmax
return choose_tone_from_probs(tone_probs)
# ── SentimentService ──────────────────────────────────────────────────────────
class SentimentService:
def __init__(self, sentiment_model_dir: str, use_cuda: bool = True):
import torch
device = torch.device("cuda" if use_cuda and torch.cuda.is_available() else "cpu")
print(f"[INFO] SentimentService device: {device}")
print(f"[INFO] Загружаем модель: {sentiment_model_dir}")
self.model, self.tokenizer, self.max_length = load_sentiment_model(
sentiment_model_dir, device
)
self.device = device
def analyze_text(
self,
text: str,
negative_threshold: Optional[float] = None,
positive_threshold: Optional[float] = None,
) -> Tuple[str, Dict[str, float]]:
"""Анализирует тональность одного текста.
Использует neg_thr/pos_thr нативно через infer_sentiment_single.
Дополнительно применяет _apply_thresholds() для согласованности
с SL-sentiment и HFBatchSender.
"""
text_clean = clean_text(text)
if not text_clean:
return "—", {}
_neg_thr = negative_threshold if negative_threshold is not None else _SENTIMENT_NEGATIVE_THRESHOLD
_pos_thr = positive_threshold if positive_threshold is not None else _SENTIMENT_POSITIVE_THRESHOLD
try:
res: Any = infer_sentiment_single(
text_clean,
self.model,
self.tokenizer,
self.max_length,
self.device,
negative_threshold=_neg_thr,
positive_threshold=_pos_thr,
)
print(f"[DEBUG] infer_sentiment_single raw result: {res!r}")
except Exception as e:
print(f"[ERROR] infer_sentiment_single exception: {e}")
traceback.print_exc()
return "—", {}
if not isinstance(res, dict):
print(f"[WARN] unexpected type from infer_sentiment_single: {type(res)}")
return "—", {}
tone_probs = normalize_prob_dict(res.get("tone_probs"))
# Применяем _apply_thresholds для явного контроля (консистентно с SL-sentiment)
tone_str = _apply_thresholds(tone_probs, neg_thr=_neg_thr, pos_thr=_pos_thr)
if tone_str == "—":
tone_str = normalize_tone_label(res.get("tone_str"))
print(f"[DEBUG] final tone_str: {tone_str!r}, probs: {tone_probs}")
return tone_str, tone_probs
_service: Optional[SentimentService] = None
def get_service() -> SentimentService:
global _service
if _service is None:
print("[INFO] Инициализация SentimentService...")
_service = SentimentService(SENTIMENT_REPO, use_cuda=USE_CUDA)
return _service
# ── UI functions ──────────────────────────────────────────────────────────────
def analyze_single_text(text: str, neg_threshold: float, pos_threshold: float):
text = text or ""
if not text.strip():
return "—", "—"
try:
svc = get_service()
tone_str, tone_probs = svc.analyze_text(
text,
negative_threshold=float(neg_threshold),
positive_threshold=float(pos_threshold),
)
except Exception as e:
print(f"[ERROR] analyze_single_text / A-sentiment: {e}")
traceback.print_exc()
return "—", "—"
tone_short = normalize_tone_label(tone_str)
if tone_short == "—" and tone_probs:
tone_short = choose_tone_from_probs(tone_probs)
probs_block = build_probs_block(tone_probs)
print(f"[DEBUG] tone_short: {tone_short!r}")
print(f"[DEBUG] tone_probs: {tone_probs}")
print(f"[DEBUG] probs_block:\n{probs_block}")
return tone_short, probs_block
def analyze_single_text_alias(text: str, neg_threshold: float, pos_threshold: float):
return analyze_single_text(text, neg_threshold, pos_threshold)
# ── Batch endpoint (Issue #222 / Шаг 8) ──────────────────────────────────────
def analyze_batch(
texts: List[str],
neg_thr: float = _SENTIMENT_NEGATIVE_THRESHOLD,
pos_thr: float = _SENTIMENT_POSITIVE_THRESHOLD,
) -> List[List[str]]:
"""Батч-анализ тональности (A).
A-модель нативно работает с neg_thr/pos_thr — tone_thr не используется.
HFBatchSender вызывает:
client.predict(texts, neg_thr, pos_thr, api_name='/analyze_batch')
Returns:
[[tone_label, probs_block], ...] той же длины что и texts.
"""
svc = get_service()
results: List[List[str]] = []
for text in texts:
if not text or not str(text).strip():
results.append(["нейтрально", "—"])
continue
try:
tone_str, tone_probs = svc.analyze_text(
str(text),
negative_threshold=float(neg_thr),
positive_threshold=float(pos_thr),
)
tone_short = normalize_tone_label(tone_str)
if tone_short == "—" and tone_probs:
tone_short = choose_tone_from_probs(tone_probs)
probs_block = build_probs_block(tone_probs)
results.append([tone_short, probs_block])
except Exception as e:
print(f"[ERROR] analyze_batch / A-sentiment (item): {e}")
results.append(["ошибка", "—"])
return results
# ─────────────────────────────────────────────────────────────────────────────
with gr.Blocks(title="Тональность сообщения") as demo:
gr.Markdown("# Определение тональности сообщения")
gr.Markdown(
"Модель: `aiforever/ruroberta-large` (fine-tuned на датасете Авито, tbsa-формат) \n"
f"Дефолтный порог негатива: `{_SENTIMENT_NEGATIVE_THRESHOLD}` · "
f"Дефолтный порог позитива: `{_SENTIMENT_POSITIVE_THRESHOLD}`"
)
inp_text = gr.Textbox(
label="Текст сообщения",
placeholder="Вставьте сообщение...",
lines=8,
)
with gr.Row():
neg_threshold_slider = gr.Slider(
minimum=0.0,
maximum=1.0,
value=_SENTIMENT_NEGATIVE_THRESHOLD,
step=0.01,
label=f"Порог негатива (neg_thr, default={_SENTIMENT_NEGATIVE_THRESHOLD})",
info="P(neg) >= neg_thr → негатив",
)
pos_threshold_slider = gr.Slider(
minimum=0.0,
maximum=1.0,
value=_SENTIMENT_POSITIVE_THRESHOLD,
step=0.01,
label=f"Порог позитива (pos_thr, default={_SENTIMENT_POSITIVE_THRESHOLD})",
info="P(pos) >= pos_thr → позитив",
)
btn = gr.Button("Анализировать")
out_tone = gr.Textbox(label="Тональность", interactive=False)
out_details = gr.Textbox(
label="Подробные вероятности",
interactive=False,
lines=6,
)
btn.click(
fn=analyze_single_text,
inputs=[inp_text, neg_threshold_slider, pos_threshold_slider],
outputs=[out_tone, out_details],
api_name="/analyze_single_text",
)
gr.Button(visible=False).click(
fn=analyze_single_text_alias,
inputs=[inp_text, neg_threshold_slider, pos_threshold_slider],
outputs=[out_tone, out_details],
api_name="analyze_single_text",
)
# Issue #222 / Шаг 8: батч-endpoint
# HFBatchSender: client.predict(texts, neg_thr, pos_thr, api_name='/analyze_batch')
gr.api(
fn=analyze_batch,
api_name="analyze_batch",
)
if __name__ == "__main__":
demo.launch(server_name="0.0.0.0", server_port=7860, show_error=True)