app.py

import time
import gradio as gr
from transformers import pipeline
from huggingface_hub import InferenceClient
from typing import List, Dict, Tuple, Any, Optional
from diffusers import AutoPipelineForText2Image
import torch

# Article Analysis Constants
MAX_CHAR = 8000
NER_NUM_ROWS = 10

# Model Constants
SUMM_MODEL_ID      = "sshleifer/distilbart-cnn-12-6"
SENTIMENT_MODEL_ID = "ahmedrachid/FinancialBERT-Sentiment-Analysis"
FINCLS_MODEL_ID    = "nickmuchi/distilroberta-finetuned-financial-text-classification"
NER_MODEL_ID       = "dslim/bert-base-NER"
CHAT_MODEL_ID = "meta-llama/Llama-3.2-3B-Instruct"
IMAGE_MODEL_ID = "stabilityai/sd-turbo"

_summ_pipe = None
_sentiment_pipe = None
_fincls_pipe = None
_ner_pipe = None
_img_pipe_cpu = None

# Image Constants
IMG_STEPS = 2
IMG_GUIDANCE = 0.5
IMG_WIDTH = 512
IMG_HEIGHT = 512

# Chat Constants
CHAT_MAX_TOKENS = 512
CHAT_TEMPERATURE = 0.7
CHAT_TOP_P = 0.95
CHAT_SYSTEM_PROMPT = ("\nYou are assisting with analysis of a financial news article."
        + "\nBe clear, cite facts from context, and avoid investment advice."
        + "\nUse the provided ARTICLE as your primary context."
        + "\nIf the user asks about something not in context, say what you do/don't know."
)

DEVICE_CPU = -1

# Assignment 4 Pipelines
def _get_summ_pipe():
    global _summ_pipe
    if _summ_pipe is None:
        _summ_pipe = pipeline(
            "summarization",
            model=SUMM_MODEL_ID,
            device=DEVICE_CPU,
        )
    return _summ_pipe

def _get_sentiment_pipe():
    global _sentiment_pipe
    if _sentiment_pipe is None:
        _sentiment_pipe = pipeline(
            "text-classification",
            model=SENTIMENT_MODEL_ID,
            truncation=True,
            device=DEVICE_CPU,
        )
    return _sentiment_pipe

def _get_fincls_pipe():
    global _fincls_pipe
    if _fincls_pipe is None:
        _fincls_pipe = pipeline(
            "text-classification",
            model=FINCLS_MODEL_ID,
            truncation=True,
            return_all_scores=True,
            device=DEVICE_CPU,
        )
    return _fincls_pipe

def _get_ner_pipe():
    global _ner_pipe
    if _ner_pipe is None:
        _ner_pipe = pipeline(
            "token-classification",
            model=NER_MODEL_ID,
            aggregation_strategy="simple",
            device=DEVICE_CPU,
        )
    return _ner_pipe

# Image Generation
# Return a plain string token from LoginButton value.
def _hf_token_str(hf_token):
    if hf_token is None:
        return None
    if isinstance(hf_token, str):
        return hf_token or None
    # gr.OAuthToken-like object
    if hasattr(hf_token, "token"):
        return hf_token.token
    # dict {"token": "..."}
    if isinstance(hf_token, dict):
        return hf_token.get("token")
    return None

def _get_img_pipe_cpu():
    global _img_pipe_cpu
    if _img_pipe_cpu is None:
        pipe = AutoPipelineForText2Image.from_pretrained(
            IMAGE_MODEL_ID,
            torch_dtype=torch.float32,
            use_safetensors=True,
        )
        pipe.to("cpu")
        for fn in ("enable_attention_slicing", "enable_vae_slicing"):
            try:
                getattr(pipe, fn)()
            except Exception:
                pass
        _img_pipe_cpu = pipe
    return _img_pipe_cpu

def _try_cloud_text2image(prompt: str, hf_token: Optional[gr.OAuthToken]):
    tok = getattr(hf_token, "token", None) if hf_token else None
    if not tok:
        return None
    try:
        client = InferenceClient(token=tok)
        return client.text_to_image(prompt, model=IMAGE_MODEL_ID)
    except Exception:
        return None

# Analysis helpers
def _normalize_text(text: str, max_len: int = MAX_CHAR) -> str:
    return (text or "").strip()[:max_len]

def run_summary(text: str) -> str:
    try:
        txt = _normalize_text(text, MAX_CHAR)
        if not txt:
            return ""
        sp = _get_summ_pipe()
        out = sp(txt[:3000], max_length=160, min_length=48, do_sample=False)
        return out[0]["summary_text"].strip() if out else ""
    except Exception as e:
        print("Summary error:", e)
        return ""

def run_text_nlp(text: str) -> Tuple[str, float, str, float]:
    try:
        txt = _normalize_text(text)
        if not txt:
            return "", 0.0, "", 0.0
        sp = _get_sentiment_pipe()
        fp = _get_fincls_pipe()
        s_pred = sp(txt)[0]
        dist = fp(txt)[0]
        top = max(dist, key=lambda d: d["score"]) if dist else {"label": "", "score": 0.0}
        return (
            s_pred.get("label", ""),
            float(s_pred.get("score", 0.0)),
            top.get("label", ""),
            float(top.get("score", 0.0)),
        )
    except Exception as e:
        print("Text NLP error:", e)
        return "Error", 0.0, "Error", 0.0

def run_ner_rows(text: str, limit: int = NER_NUM_ROWS) -> List[List[str]]:
    try:
        txt = _normalize_text(text, MAX_CHAR)
        if not txt:
            return []
        ner = _get_ner_pipe()
        ents = ner(txt)
        rows = [
            [e.get("entity_group", ""), e.get("word", ""), f"{float(e.get('score', 0.0)):.2f}"]
            for e in ents
        ]
        return rows[:limit]
    except Exception as e:
        print("NER error:", e)
        return [["Error", str(e), "0.00"]]

# Chat helpers
def build_context_block(article: str, analysis: Dict[str, Any]) -> str:
    parts = []
    if article:
        parts.append(f"ARTICLE (truncated):\n{article[:MAX_CHAR]}")
    if analysis:
        parts.append(
            "ANALYSIS SUMMARY:\n"
            f"- Sentiment: {analysis.get('sentiment')} ({analysis.get('sentiment_score'):.2f})\n"
            f"- Financial stance: {analysis.get('category')} ({analysis.get('category_score'):.2f})"
        )
        if analysis.get("summary"):
            parts.append(f"- Auto Summary: {analysis['summary']}")
        ents = analysis.get("entities", [])
        if ents:
            ent_str = ", ".join({r[1] for r in ents[:40]})
            parts.append(f"- Top entities: {ent_str}")
    return "\n\n".join(parts)

def _warn_if_no_token(hf_token: Optional[gr.OAuthToken]) -> str:
    if not hf_token or not getattr(hf_token, "token", None):
        return "\nYou are not logged in to Hugging Face. Click **Login** (left sidebar) for better reliability.\n\n"
    return ""

def respond_chat(
    message: str,
    history: List[Dict[str, str]],
    article_text: str,
    analysis: Dict[str, Any],
    hf_token: gr.OAuthToken,
    _profile,
):
    tok = _hf_token_str(hf_token)

    login_notice = _warn_if_no_token(hf_token)

    client = InferenceClient(
        token=tok,
        model=CHAT_MODEL_ID
    )

    context_block = build_context_block(article_text or "", analysis or {})
    sys = (CHAT_SYSTEM_PROMPT)

    messages = [
        {"role": "system", "content": sys},
        {"role": "system", "content": context_block},
        *history,
        {"role": "user", "content": message},
    ]

    response = login_notice
    try:
        stream = client.chat_completion(
            messages,
            max_tokens=CHAT_MAX_TOKENS,
            stream=True,
            temperature=CHAT_TEMPERATURE,
            top_p=CHAT_TOP_P,
        )
        for chunk in stream:
            choices = getattr(chunk, "choices", [])
            piece = ""
            if choices and getattr(choices[0], "delta", None) and choices[0].delta.content:
                piece = choices[0].delta.content
            response += piece
            yield response
    except Exception as e:
        response += (
            f"\nChat request failed for model `{CHAT_MODEL_ID}`.\n"
            f"Error: {e}\n"
        )
        yield response

# Image helpers
def generate_image(prompt, width, height, hf_token, *args):
    import traceback
    t0 = time.time()
    prompt = (prompt or "").strip()
    if not prompt:
        return None, "Provide a prompt."

    # 1) Cloud first (shared GPU)
    try:
        img = _try_cloud_text2image(prompt, hf_token)
        if img is not None:
            return img, f"{time.time()-t0:.2f}s"
    except Exception as e:
        print("Cloud image error:", e)
        traceback.print_exc()

    # 2) CPU fallback
    try:
        pipe = _get_img_pipe_cpu()
        width, height = int(width), int(height)
        out = pipe(
            prompt=prompt,
            num_inference_steps=IMG_STEPS,
            guidance_scale=IMG_GUIDANCE,
            width=width,
            height=height,
        )
        img = out.images[0]
        return img, f"{time.time()-t0:.2f}s | steps={IMG_STEPS}, g={IMG_GUIDANCE}"
    except Exception as e:
        print("CPU image error:", e)
        traceback.print_exc()
        return None, f"Generation failed: {e}"

# Gradio UI
with gr.Blocks(fill_height=True) as demo:
    gr.Markdown("**ARIN 460 Final — Financial News Multi-Model**")

    article_state = gr.State("")
    analysis_state = gr.State({})

    with gr.Sidebar():
        login_btn = gr.LoginButton()
        gr.Markdown("**Workflow**\n1) Input\n2) Analysis (Assignment 4)\n3) Chat\n4) Image")

    with gr.Tabs():
        with gr.Tab("Input"):
            txt_in = gr.Textbox(lines=12, label="Article text")
            analyze_btn = gr.Button("Analyze", variant="primary")
            run_status = gr.Markdown()

        with gr.Tab("Text Analysis"):
            summary_box = gr.Textbox(label="Summary", lines=4, interactive=False)
            sent_lbl = gr.Textbox(label="Sentiment", interactive=False)
            sent_score = gr.Number(label="Sentiment score", precision=3, interactive=False)
            fin_lbl = gr.Textbox(label="Financial Category", interactive=False)
            fin_score = gr.Number(label="Category score", precision=3, interactive=False)
            ta_status = gr.Markdown()

        with gr.Tab("NER"):
            ner_out = gr.Dataframe(headers=["entity", "text", "score"],
                                   datatype=["str", "str", "str"], interactive=False)
            ner_status = gr.Markdown()

        with gr.Tab("Chat"):
            chat = gr.ChatInterface(
                respond_chat,
                type="messages",
                additional_inputs=[
                    article_state, analysis_state, login_btn
                ],
            )
            chat.chatbot.height = 400

        with gr.Tab("Image"):
            img_prompt = gr.Textbox(label="Prompt", lines=3)
            width_slider = gr.Slider(256, 768, value=IMG_WIDTH, step=64, label="Width")
            height_slider = gr.Slider(256, 768, value=IMG_HEIGHT, step=64, label="Height")
            gen_btn = gr.Button("Generate Image", variant="primary")
            image_out = gr.Image(label="Result", type="pil")
            gen_status = gr.Markdown()
            gen_btn.click(
                generate_image,
                inputs=[img_prompt, width_slider, height_slider, login_btn],
                outputs=[image_out, gen_status]
            )

    def _analyze_all(text):
        t0 = time.time()
        summ = run_summary(text)
        s_lbl, s_score, c_lbl, c_score = run_text_nlp(text)
        ner_rows = run_ner_rows(text)
        dt = time.time() - t0
        analysis = {
            "summary": summ,
            "sentiment": s_lbl,
            "sentiment_score": s_score,
            "category": c_lbl,
            "category_score": c_score,
            "entities": ner_rows,
        }
        return (
            f"Processed in **{dt:.2f}s**.",
            summ, s_lbl, s_score, c_lbl, c_score, f"Updated at {time.strftime('%H:%M:%S')}",
            ner_rows, f"Extracted {len(ner_rows)} entities.",
            text, analysis
        )

    # Analyze button
    analyze_btn.click(lambda: gr.update(value="Analyzing...", interactive=False), [], [analyze_btn]) \
               .then(_analyze_all, inputs=[txt_in],
                     outputs=[run_status, summary_box, sent_lbl, sent_score, fin_lbl, fin_score,
                              ta_status, ner_out, ner_status, article_state, analysis_state]) \
               .then(lambda: gr.update(value="Analyze", interactive=True), [], [analyze_btn])

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