# 使用真实建筑照片，按 ViT 常见设置切成 16x16 图块。
raw_photo = load_sample_image("china.jpg")
vit_image = np.asarray(Image.fromarray(raw_photo).resize((224, 224))) / 255.0
patch_size = 16
patch_grid = vit_image.shape[0] // patch_size
patches = vit_image.reshape(patch_grid, patch_size, patch_grid, patch_size, 3).swapaxes(1, 2)
patch_tokens = patches.reshape(-1, patch_size * patch_size * 3)

patch_summary = []
for patch_id, patch in enumerate(patches.reshape(-1, patch_size, patch_size, 3)):
    row, col = divmod(patch_id, patch_grid)
    patch_summary.append({
        "图块编号": patch_id,
        "行": row,
        "列": col,
        "向量维度": patch_tokens.shape[1],
        "R均值": patch[:, :, 0].mean(),
        "G均值": patch[:, :, 1].mean(),
        "B均值": patch[:, :, 2].mean(),
        "亮度标准差": patch.mean(axis=2).std(),
    })

patch_df = pd.DataFrame(patch_summary)
display(patch_df.head(12).round(3))

# 绘制真实图片图块网格和若干具体图块。
selected_patches = [18, 43, 87, 112, 145, 181]
fig = plt.figure(figsize=(10.8, 6.6))
gs = fig.add_gridspec(2, 6, height_ratios=[3.4, 1.5], hspace=0.22, wspace=0.08)
ax_img = fig.add_subplot(gs[0, :])
ax_img.imshow(vit_image)
ax_img.set_title("真实图片切成 14x14 个图块", loc="left", fontweight="bold")
ax_img.set_xticks(np.arange(0, 225, patch_size))
ax_img.set_yticks(np.arange(0, 225, patch_size))
ax_img.grid(color="#ffffff", linewidth=0.8)
ax_img.tick_params(labelbottom=False, labelleft=False, length=0)

for slot, patch_id in enumerate(selected_patches):
    row, col = divmod(patch_id, patch_grid)
    ax = fig.add_subplot(gs[1, slot])
    ax.imshow(patches[row, col])
    ax.set_title(f"图块 {patch_id}\n({row},{col})", fontsize=9, fontweight="bold")
    ax.set_xticks([])
    ax.set_yticks([])

fig.suptitle("ViT 图块切分：真实照片转成图块向量序列", x=0.08, ha="left", fontsize=14, fontweight="bold", color="#0f172a")
plt.tight_layout()
plt.show()

	图块编号	列	向量维度	R均值	G均值	B均值	亮度标准差
0	0	0	768	0.703	0.805	0.917	0.007
1	1	1	768	0.718	0.819	0.937	0.005
2	2	2	768	0.732	0.833	0.951	0.004
3	3	3	768	0.746	0.845	0.962	0.006
4	4	4	768	0.763	0.861	0.966	0.014
5	5	5	768	0.784	0.879	0.977	0.005
6	6	6	768	0.808	0.896	0.990	0.006
7	7	7	768	0.838	0.915	0.999	0.006
8	8	8	768	0.864	0.932	0.996	0.004
9	9	9	768	0.892	0.945	0.996	0.005
10	10	10	768	0.909	0.954	0.997	0.005
11	11	11	768	0.925	0.963	0.999	0.003

第 10 章 · ViT 图块切分代码实验¶

运行内容¶

0. 真实照片切块¶