app.py

import json
import numpy as np
import onnxruntime as ort
from PIL import Image

# --- Fix gradio_client bug: json_schema_to_python_type crashes on bool schemas
#     (TypeError: argument of type 'bool' is not iterable) -> breaks /info & API ---
import gradio_client.utils as _gcu

_orig_j2p = _gcu._json_schema_to_python_type


def _safe_j2p(schema, defs=None):
    if isinstance(schema, bool):
        return "Any"
    return _orig_j2p(schema, defs)


_gcu._json_schema_to_python_type = _safe_j2p

_orig_get_type = _gcu.get_type


def _safe_get_type(schema):
    if not isinstance(schema, dict):
        return "Any"
    return _orig_get_type(schema)


_gcu.get_type = _safe_get_type

import gradio as gr
from huggingface_hub import hf_hub_download

MODEL_REPO = "Boos4721/cpdaily-ocr"

# 从 HF Hub 拉取模型与字符表
model_path = hf_hub_download(MODEL_REPO, "cpdaily_captcha_ocr_fp16.onnx")
charset_path = hf_hub_download(MODEL_REPO, "charset.json")

CHARS = json.load(open(charset_path))  # ["<blank>", "A", "B", ...]
IMG_H, IMG_W = 32, 160

_opts = ort.SessionOptions()
_opts.log_severity_level = 3
_sess = ort.InferenceSession(model_path, sess_options=_opts, providers=["CPUExecutionProvider"])
_inp = _sess.get_inputs()[0].name


PLACEHOLDER = (
    "<div class='result-card empty'>"
    "<div class='result-text'>—</div>"
    "<div class='result-hint'>上传或拖入一张验证码图片开始识别</div>"
    "</div>"
)


def _result_html(text, avg):
    pct = avg * 100
    color = "#16a34a" if pct >= 95 else ("#f59e0b" if pct >= 80 else "#dc2626")
    chars = "".join(f"<span class='ch'>{c}</span>" for c in text)
    return (
        "<div class='result-card'>"
        f"<div class='result-text'>{chars}</div>"
        "<div class='conf-row'>"
        "<span class='conf-label'>平均置信度 / Confidence</span>"
        f"<span class='conf-val' style='color:{color}'>{pct:.1f}%</span>"
        "</div>"
        "<div class='conf-bar'>"
        f"<div class='conf-fill' style='width:{pct:.1f}%;background:{color}'></div>"
        "</div>"
        "</div>"
    )


def _infer(image):
    """核心推理：返回 (text, avg_confidence, per_char_confidences)。"""
    img = Image.fromarray(image).convert("L").resize((IMG_W, IMG_H), Image.BILINEAR)
    x = (np.asarray(img, dtype=np.float32) / 255.0)[None, None, :, :]
    logits = _sess.run(None, {_inp: x})[0][0]      # [T, 63] log-softmax
    prob = np.exp(logits)
    idx = logits.argmax(-1)
    conf = prob.max(-1)
    out, confs, prev = [], [], -1
    for t, p in enumerate(idx):
        if p != prev and p != 0:                    # CTC greedy: 去重 + 去 blank
            out.append(CHARS[p]); confs.append(float(conf[t]))
        prev = p
    text = "".join(out)
    avg = float(np.mean(confs)) if confs else 0.0
    return text, avg, confs


def recognize(image):
    if image is None:
        return PLACEHOLDER
    text, avg, _ = _infer(image)
    return _result_html(text, avg)


def recognize_json(image):
    """JSON API 端点：返回 {text, confidence, chars}。"""
    if image is None:
        return {"text": "", "confidence": 0.0, "chars": []}
    text, avg, confs = _infer(image)
    return {
        "text": text,
        "confidence": round(avg, 4),
        "chars": [{"char": c, "confidence": round(cf, 4)} for c, cf in zip(text, confs)],
    }


CSS = """
.gradio-container {max-width: 880px !important; margin: 0 auto !important;}
#title-block h1 {font-size: 1.7rem; margin-bottom: .2rem;}
#title-block p, #title-block li {color: var(--body-text-color-subdued); font-size: .9rem; line-height: 1.5;}
.result-card {min-height: 180px; display: flex; flex-direction: column;
  align-items: center; justify-content: center; gap: 14px;
  border: 1px solid var(--border-color-primary); border-radius: 14px;
  padding: 24px; background: var(--block-background-fill);}
.result-card.empty {opacity: .7;}
.result-text {font-size: 2.6rem; font-weight: 700; letter-spacing: .35rem;
  font-family: ui-monospace, "SF Mono", Menlo, monospace; line-height: 1;}
.result-text .ch {display: inline-block; padding: 0 2px;}
.result-hint {font-size: .9rem; color: var(--body-text-color-subdued); letter-spacing: normal; font-weight: 400;}
.conf-row {display: flex; align-items: baseline; gap: 10px;}
.conf-label {font-size: .85rem; color: var(--body-text-color-subdued);}
.conf-val {font-size: 1.25rem; font-weight: 700;}
.conf-bar {width: 70%; height: 8px; border-radius: 999px;
  background: var(--neutral-200); overflow: hidden;}
.conf-fill {height: 100%; border-radius: 999px; transition: width .3s ease;}
footer {display: none !important;}
"""

with gr.Blocks(title="cpdaily-ocr · 今日校园验证码识别", theme=gr.themes.Soft(
        primary_hue="orange", secondary_hue="blue"), css=CSS) as demo:
    with gr.Column(elem_id="title-block"):
        gr.Markdown(
            "# 🔠 cpdaily-ocr · 今日校园验证码识别\n"
            "上传一张今日校园 / Cpdaily 的 5 位验证码图片,自动识别其中字符。\n\n"
            "轻量 **CRNN + CTC** 模型 · 纯 ONNX · **1.07 MB** · **99.37%** 准确率"
        )
    with gr.Row(equal_height=True):
        with gr.Column(scale=1):
            inp = gr.Image(label="验证码图片 / Captcha", type="numpy",
                           height=200, sources=["upload", "clipboard"])
            btn = gr.Button("识别 / Recognize", variant="primary", size="lg")
        with gr.Column(scale=1):
            out_html = gr.HTML(value=PLACEHOLDER)
    gr.Markdown("### 示例 / Examples")
    gr.Examples(
        examples=[["examples/sample1.png"], ["examples/sample2.png"], ["examples/sample3.png"]],
        inputs=inp,
    )
    gr.Markdown(
        "字符集 A-Z / a-z / 0-9(62 类)· 输入灰度 32×160 · CTC 贪心解码 · "
        "模型 [`Boos4721/cpdaily-ocr`](https://huggingface.co/Boos4721/cpdaily-ocr)"
    )
    # UI 交互（不暴露为可编程 API，避免重复端点）
    inp.change(recognize, inputs=inp, outputs=out_html, api_name=False)
    btn.click(recognize, inputs=inp, outputs=out_html, api_name="recognize")

    # ---- 纯 JSON API 端点（不渲染 UI）：api_name="/predict_json" ----
    api_in = gr.Image(type="numpy", visible=False)
    api_out = gr.JSON(visible=False)
    api_in.change(recognize_json, inputs=api_in, outputs=api_out,
                  api_name="predict_json", show_progress="hidden")

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