app.py

import gradio as gr
import numpy as np
import os
import PIL
from PIL import ImageOps
import test_1
import torch
model = test_1.load_trained_model()



def convert_to_28x28_grid(canvas_image):
    #将750x750画布转换为28x28数值矩阵
    canvas_array = np.array(canvas_image)

    GRID_SIZE = 28
    CELL_SIZE = 750 // GRID_SIZE

    # 转换为灰度
    if canvas_array.ndim == 3:
        gray = np.dot(canvas_array[..., :3], [0, 0, 0])  # 标准灰度公式
    else:
        gray = canvas_array

    grid = np.zeros((GRID_SIZE, GRID_SIZE), dtype=np.float32)

    for i in range(GRID_SIZE):
        for j in range(GRID_SIZE):
            # 提取每个单元格
            x_start = i * CELL_SIZE
            x_end = (i + 1) * CELL_SIZE
            y_start = j * CELL_SIZE
            y_end = (j + 1) * CELL_SIZE

            cell = gray[x_start:x_end, y_start:y_end]

            # 注意：所有以下操作应在循环内部
            activation = np.mean(255 - cell)  # 反色处理
            grid[i, j] = min(255, activation * 2.5)  # 提高对比度

    return grid.astype(np.uint8)  # 0-255整型


def predict_digit(input_image):
    if input_image is None:
        return "请先绘制数字", {}

    device = torch.device("cuda" if torch.cuda.is_available() else "cpu") # 使用 GPU/CPU
    model.to(device)

    img=ImageOps.invert(input_image)

    # 执行网格转换
    digit_matrix = convert_to_28x28_grid(img)

    # 模型预测
    pred_class, probabilities = test_1.predict_user_image(digit_matrix, model,device)

    probs_dict = {str(i): float(p) for i, p in enumerate(probabilities)}
    return f"结果: {pred_class}", probs_dict



# 界面构建
with gr.Blocks() as demo:
    gr.Markdown("""# digit-recognition""")

    with gr.Row():
        # 本地使用 shape 参数
        input_image = gr.Sketchpad(
            shape=(750, 750),
            brush_color="black",
            image_mode="L",  # 确保输出是单通道
            type="pil"
        )


    with gr.Row():
        output_label = gr.Label(label="识别结果")
        output_prob = gr.Label(label="概率分布")

    with gr.Row():
        recognize_btn = gr.Button("识别", variant="primary")

    recognize_btn.click(
        predict_digit,
        inputs=input_image,
        outputs=[output_label, output_prob]
    )

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