"""A small Gradio interface backed by one persistent image classifier.""" from __future__ import annotations from collections.abc import Callable from pathlib import Path import gradio as gr from PIL import Image import torch from .checkpoint import load_checkpoint from .inference import prediction_transform, resolve_device WAITING_MESSAGE = "### 等待识别\n\n上传图片后,点击左侧的“开始识别”。" def _readable_label(name: str) -> str: if name == "Background_without_leaves": return "无叶片背景" return name.replace("___", " · ").replace("_", " ") class ImageClassifier: """Load weights once and reuse the CLI's deterministic image preprocessing.""" def __init__(self, checkpoint_path: str, device: str = "cpu") -> None: path = Path(checkpoint_path) if not path.is_file(): raise FileNotFoundError(f"找不到模型文件:{path}。请先完成训练或指定已有权重。") self.device = resolve_device(device) self.model, checkpoint = load_checkpoint(path, self.device) self.class_names = list(checkpoint["class_names"]) self.transform = prediction_transform(checkpoint) def classify(self, image: Image.Image | None) -> tuple[str, dict[str, float], list[list[str]]]: if image is None: raise gr.Error("请先上传一张叶片图片,再点击“开始识别”。") if not isinstance(image, Image.Image): raise gr.Error("无法读取这张图片,请上传 JPG、PNG 或其他常见图片格式。") try: tensor = self.transform(image).unsqueeze(0).to(self.device) except (OSError, ValueError) as error: raise gr.Error("无法读取这张图片,请换一张图片重试。") from error with torch.inference_mode(): logits = self.model(tensor) if logits.shape != (1, len(self.class_names)): raise RuntimeError("Model output does not match the saved class mapping") probabilities = logits.float().softmax(dim=1)[0] if not torch.isfinite(probabilities).all(): raise gr.Error("本次识别未能完成,请重新上传图片后重试。") scores, indices = probabilities.topk(min(5, len(self.class_names))) ranking = { self.class_names[index]: float(score) for index, score in zip(indices.cpu().tolist(), scores.cpu().tolist()) } best_label, confidence = next(iter(ranking.items())) summary = f"### {_readable_label(best_label)}\n\n预测概率 **{confidence:.2%}**" rows = [[name, f"{score:.2%}"] for name, score in ranking.items()] return summary, ranking, rows def create_app( checkpoint_path: str, device: str = "cpu", example_path: str | None = None, *, prediction_decorator: Callable[[Callable], Callable] | None = None, ) -> gr.Blocks: """Build the page and load its classifier without starting a web server.""" classifier = ImageClassifier(checkpoint_path, device) classify = classifier.classify if prediction_decorator is not None: classify = prediction_decorator(classify) with gr.Blocks(title="植物病害识别", analytics_enabled=False, fill_width=True) as app: gr.Markdown("# 植物病害识别\n上传一张叶片照片,查看最可能的分类结果。") gr.Markdown(f"支持 **{len(classifier.class_names)} 个类别**,包含健康叶片、病害叶片和无叶片背景。") with gr.Row(equal_height=True, elem_id="classifier-layout"): with gr.Column(scale=1, min_width=0): image = gr.Image( label="叶片照片", type="pil", sources=["upload"], image_mode="RGB", height=360, ) with gr.Row(): submit = gr.Button("开始识别", variant="primary") clear = gr.ClearButton(value="清空", variant="secondary") if example_path is not None and Path(example_path).is_file(): gr.Examples( examples=[[str(Path(example_path).resolve())]], inputs=[image], label="试用示例图片", cache_examples=False, ) with gr.Column(scale=1, min_width=0): summary = gr.Markdown(WAITING_MESSAGE) probabilities = gr.Label(label="最可能的 5 个类别", num_top_classes=5) table = gr.Dataframe( headers=["类别", "预测概率"], datatype=["str", "str"], row_count=5, column_count=2, interactive=False, label="分类结果明细", wrap=True, ) gr.Markdown("建议选择叶片清晰、光照均匀的照片。概率表示模型对已有类别的判断。") prediction = submit.click( classify, inputs=[image], outputs=[summary, probabilities, table], api_name="classify", concurrency_limit=1, ) image.change( lambda: (WAITING_MESSAGE, None, []), inputs=None, outputs=[summary, probabilities, table], queue=False, api_visibility="private", cancels=[prediction], ) clear.add([image, summary, probabilities, table]) return app def launch_app( checkpoint_path: str, device: str = "cpu", host: str = "0.0.0.0", port: int = 7860, example_path: str | None = None, *, prediction_decorator: Callable[[Callable], Callable] | None = None, ) -> None: app = create_app( checkpoint_path, device, example_path, prediction_decorator=prediction_decorator, ) app.queue(max_size=16, default_concurrency_limit=1).launch( server_name=host, server_port=port, share=False, show_error=False, footer_links=[], theme=gr.themes.Soft( primary_hue="emerald", secondary_hue="green", neutral_hue="slate", font=["system-ui", "sans-serif"], font_mono=["ui-monospace", "monospace"], ), css=""" .gradio-container { width: 100% !important; max-width: 1120px !important; min-width: 0 !important; margin: 0 auto; } .gradio-container .main {width: 100%; min-width: 0; box-sizing: border-box;} #classifier-layout { display: grid !important; grid-template-columns: minmax(0, 1fr) minmax(0, 1fr); gap: 24px; } @media (max-width: 760px) { .gradio-container .main {padding-left: 16px !important; padding-right: 16px !important;} #classifier-layout {grid-template-columns: minmax(0, 1fr); gap: 16px;} } """, )