app.py

from pathlib import Path
import torch
import gradio as gr
from torch import nn
import cv2
import numpy as np

BASE_DIR = Path(__file__).parent
LABELS = (BASE_DIR / "class_names.txt").read_text().splitlines()

model = nn.Sequential(
    nn.Conv2d(1, 32, 3, padding="same"),
    nn.ReLU(),
    nn.MaxPool2d(2),
    nn.Conv2d(32, 64, 3, padding="same"),
    nn.ReLU(),
    nn.MaxPool2d(2),
    nn.Conv2d(64, 128, 3, padding="same"),
    nn.ReLU(),
    nn.MaxPool2d(2),
    nn.Flatten(),
    nn.Linear(1152, 256),
    nn.ReLU(),
    nn.Linear(256, len(LABELS)),
)
state_dict = torch.load(BASE_DIR /"pytorch_model.bin", map_location="cpu")
model.load_state_dict(state_dict, strict=False)
model.eval()


def predict(im):
    # 检查是否为空画布
    if isinstance(im, dict):
        # 空画布
        if "layers" in im and (not im["layers"]):
            return {"请先在画板上画图": 1.0}
        # 优先用composite
        if "composite" in im and im["composite"] is not None:
            im = im["composite"]
        # 否则用layers[0]
        elif "layers" in im and isinstance(im["layers"], list) and len(im["layers"]) > 0:
            im = im["layers"][0]
        else:
            return {"无法识别输入": 1.0}
    # 转为numpy数组
    im = np.array(im)
    # 只取第一个通道（灰度）
    if im.ndim == 3:
        im = im[..., 0]
        # resize 到模型训练时的尺寸（假设24x24）
    im = cv2.resize(im, (24, 24), interpolation=cv2.INTER_AREA)
    x = torch.tensor(im, dtype=torch.float32).unsqueeze(0).unsqueeze(0) / 255.0
    with torch.no_grad():
        out = model(x)
    probabilities = torch.nn.functional.softmax(out[0], dim=0)
    values, indices = torch.topk(probabilities, 5)
    return {LABELS[i]: v.item() for i, v in zip(indices, values)}

interface = gr.Interface(
    predict,
    inputs="sketchpad",
    outputs="label",
    # theme="huggingface",
    title="Sketch Recognition",
    description="Who wants to play Pictionary? Draw a common object like a shovel or a laptop, and the algorithm will guess in real time!",
    article="<p style='text-align: center'>Sketch Recognition | Demo Model</p>",
    # live=True,
)
interface.launch(share=True)