app.py

import torch
import json
import torch.nn.functional as F
from transformers import BertTokenizer
import gradio as gr
from typing import List, Dict

from model import CommentMTLModel   # your class

# ------------ Device optimisation -----------------------------------------------------------------
if torch.backends.mps.is_available():
    device = torch.device("mps")
elif torch.cuda.is_available():
    device = torch.device("cuda")
else:
    device = torch.device("cpu")

# ------------ Model / tokenizer ------------------------------------------------------
TOKENIZER_DIR = "/app/bert-base-uncased"     # 新增
tokenizer = BertTokenizer.from_pretrained(
    TOKENIZER_DIR,
    local_files_only=True                    # 强制离线
)

with open("config.json") as f:
    cfg = json.load(f)

model = CommentMTLModel(
    model_name="bert-base-uncased",
    num_sentiment_labels=cfg["num_sentiment_labels"],
    num_toxicity_labels=cfg["num_toxicity_labels"],
    dropout_prob=cfg.get("dropout_prob", 0.1)
)
model.load_state_dict(torch.load("pytorch_model.bin", map_location=device))
model.to(device).eval()

sentiment_labels = ["Negative", "Neutral", "Positive"]
toxicity_labels  = ["Toxic", "Severe Toxic", "Obscene", "Threat", "Insult", "Identity Hate"]

# ------------ Core inference function ------------------------------------------------
@torch.inference_mode()
def analyse_batch(comments_text: str) -> Dict:
    """
    comments_text: multiline string, each line is a comment (≤100 lines)
    returns: aggregated statistics dict
    """
    # Split input into list of comments, remove blank lines
    comments: List[str] = [line for line in comments_text.splitlines() if line.strip()]

    # Ensure we have at most 100 comments
    comments = comments[:100]

    # ---- encode all comments (batched) ----------
    enc = tokenizer(
        comments,
        return_tensors="pt",
        padding=True,
        truncation=True,
        max_length=512
    )
    enc = {k: v.to(device) for k, v in enc.items()}

    # ---- forward pass (split to mini-batches in case 100 is too big) ----
    batch_size = 32
    n = enc["input_ids"].shape[0]

    # counters
    sent_counts = {lab: 0 for lab in sentiment_labels}
    tox_counts  = {lab: 0 for lab in toxicity_labels}
    comments_with_any_tox = 0

    for i in range(0, n, batch_size):
        sl = slice(i, i + batch_size)
        out = model(
            input_ids      = enc["input_ids"][sl],
            attention_mask = enc["attention_mask"][sl],
            token_type_ids = enc.get("token_type_ids", None)[sl] if "token_type_ids" in enc else None
        )

        # ----- sentiment (softmax, pick max) ----------------------------
        sent_logits = out["sentiment_logits"]          # (b, 3)
        sent_pred   = sent_logits.softmax(dim=1).argmax(dim=1)  # (b,)
        for idx in sent_pred.tolist():
            sent_counts[sentiment_labels[idx]] += 1

        # ----- toxicity (sigmoid, multi-label) --------------------------
        tox_probs = out["toxicity_logits"].sigmoid()   # (b, 6)
        toxic_mask = tox_probs > 0.30                  # boolean mask
        comments_with_any_tox += toxic_mask.any(dim=1).sum().item()

        # add per-label counts
        for lab_idx, lab in enumerate(toxicity_labels):
            tox_counts[lab] += toxic_mask[:, lab_idx].sum().item()

    return {
        "sentiment_counts": sent_counts,
        "toxicity_counts": tox_counts,
        "comments_with_any_toxicity": int(comments_with_any_tox)
    }

# ------------ Gradio interface -------------------------------------------------------
iface = gr.Interface(
    fn=analyse_batch,
    inputs=gr.Textbox(
        label="YouTube comments (max 100, one per line)",
        placeholder="Paste up to 100 comments, each on its own line.",
        lines=20,
        max_lines=100
    ),
    outputs=gr.JSON(label="Aggregated statistics"),
    title="YouTube Comment Sentiment & Toxicity Batch API",
    description=(
        "Paste up to 100 raw comment strings, each on a new line, "
        "then click Analyze to receive counts of Positive/Neutral/Negative comments "
        "plus counts of toxicity labels where probability > 0.30."
    ),
    allow_flagging="never"
)

if __name__ == "__main__":
    iface.launch()