CaptchBreaker/inference/export_onnx.py

"""
ONNX 导出脚本

从 checkpoints/ 加载训练好的 PyTorch 模型，导出为 ONNX 格式到 onnx_models/。
支持逐个导出或一次导出全部。
"""

import torch
import torch.nn as nn

from config import (
    CHECKPOINTS_DIR,
    ONNX_DIR,
    ONNX_CONFIG,
    IMAGE_SIZE,
    NORMAL_CHARS,
    MATH_CHARS,
    THREED_CHARS,
    NUM_CAPTCHA_TYPES,
    REGRESSION_RANGE,
)
from models.classifier import CaptchaClassifier
from models.lite_crnn import LiteCRNN
from models.threed_cnn import ThreeDCNN
from models.regression_cnn import RegressionCNN


def export_model(
    model: nn.Module,
    model_name: str,
    input_shape: tuple,
    onnx_dir: str | None = None,
):
    """
    导出单个模型为 ONNX。

    Args:
        model:       已加载权重的 PyTorch 模型
        model_name:  模型名 (classifier / normal / math / threed_text / threed_rotate / threed_slider)
        input_shape: 输入形状 (C, H, W)
        onnx_dir:    输出目录 (默认使用 config.ONNX_DIR)
    """
    from pathlib import Path

    out_dir = Path(onnx_dir) if onnx_dir else ONNX_DIR
    out_dir.mkdir(parents=True, exist_ok=True)
    onnx_path = out_dir / f"{model_name}.onnx"

    model.eval()
    model.cpu()

    dummy = torch.randn(1, *input_shape)

    # 分类器和识别器的 dynamic_axes 不同
    if model_name == "classifier" or model_name in ("threed_rotate", "threed_slider"):
        dynamic_axes = {"input": {0: "batch"}, "output": {0: "batch"}}
    else:
        # CTC 模型: output shape = (T, B, C)
        dynamic_axes = {"input": {0: "batch"}, "output": {1: "batch"}}

    torch.onnx.export(
        model,
        dummy,
        str(onnx_path),
        opset_version=ONNX_CONFIG["opset_version"],
        input_names=["input"],
        output_names=["output"],
        dynamic_axes=dynamic_axes if ONNX_CONFIG["dynamic_batch"] else None,
    )

    size_kb = onnx_path.stat().st_size / 1024
    print(f"[ONNX] 导出完成: {onnx_path}  ({size_kb:.1f} KB)")


def _load_and_export(model_name: str):
    """从 checkpoint 加载模型并导出 ONNX。"""
    ckpt_path = CHECKPOINTS_DIR / f"{model_name}.pth"
    if not ckpt_path.exists():
        print(f"[跳过] {model_name}: checkpoint 不存在 ({ckpt_path})")
        return

    ckpt = torch.load(ckpt_path, map_location="cpu", weights_only=True)
    acc_info = ckpt.get('best_acc') or ckpt.get('best_tol_acc', '?')
    print(f"[加载] {model_name}: epoch={ckpt.get('epoch', '?')}  acc={acc_info}")

    if model_name == "classifier":
        model = CaptchaClassifier(num_types=NUM_CAPTCHA_TYPES)
        h, w = IMAGE_SIZE["classifier"]
        input_shape = (1, h, w)
    elif model_name == "normal":
        chars = ckpt.get("chars", NORMAL_CHARS)
        h, w = IMAGE_SIZE["normal"]
        model = LiteCRNN(chars=chars, img_h=h, img_w=w)
        input_shape = (1, h, w)
    elif model_name == "math":
        chars = ckpt.get("chars", MATH_CHARS)
        h, w = IMAGE_SIZE["math"]
        model = LiteCRNN(chars=chars, img_h=h, img_w=w)
        input_shape = (1, h, w)
    elif model_name == "threed_text":
        chars = ckpt.get("chars", THREED_CHARS)
        h, w = IMAGE_SIZE["3d_text"]
        model = ThreeDCNN(chars=chars, img_h=h, img_w=w)
        input_shape = (1, h, w)
    elif model_name == "threed_rotate":
        h, w = IMAGE_SIZE["3d_rotate"]
        model = RegressionCNN(img_h=h, img_w=w)
        input_shape = (1, h, w)
    elif model_name == "threed_slider":
        h, w = IMAGE_SIZE["3d_slider"]
        model = RegressionCNN(img_h=h, img_w=w)
        input_shape = (1, h, w)
    else:
        print(f"[错误] 未知模型: {model_name}")
        return

    model.load_state_dict(ckpt["model_state_dict"])
    export_model(model, model_name, input_shape)


def export_all():
    """依次导出 classifier, normal, math, threed_text, threed_rotate, threed_slider 六个模型。"""
    print("=" * 50)
    print("导出全部 ONNX 模型")
    print("=" * 50)
    for name in ["classifier", "normal", "math", "threed_text", "threed_rotate", "threed_slider"]:
        _load_and_export(name)
    print("\n全部导出完成。")


if __name__ == "__main__":
    export_all()