Real-time PyTorch tensor visualization via Vulkan external memory.
Zero-copy from CUDA to screen — for training, debugging, and research.
通过 Vulkan 外部内存实现 PyTorch 张量实时可视化。
CUDA 到屏幕零拷贝 — 用于训练、调试与科研。
// real-time tensor visualization at 60 fps
// 60 fps 实时张量可视化
Anything that lives in a torch.Tensor can be visualized — neural rendering, RL, physics, signals.
任何存在于 torch.Tensor 中的数据都能可视化 — 神经渲染、强化学习、物理仿真、信号处理。
Inspect rendered views frame-by-frame as your radiance field trains. Adjust hyperparams with live sliders.
逐帧检查辐射场训练中的渲染视图，通过滑条实时调整超参数。
Display rendered images, gradient maps, and error heatmaps side-by-side. Data never leaves the GPU.
并排显示渲染图、梯度图与误差热力图。数据全程不离开 GPU。
Watch DQN agents learn Snake in real time. Visualize Q-value heatmaps, reward curves, exploration.
实时观看 DQN 智能体学习贪吃蛇。可视化 Q 值热力图、奖励曲线、探索策略。
Watch textures converge and spot artifacts during training without saving intermediate PNGs.
训练过程中实时观察纹理收敛、发现伪影，无需保存中间图片。
Change LR, optimizer, loss function, and network architecture at runtime — see effects instantly.
运行时修改学习率、优化器、损失函数、网络结构 — 即时看到效果。
SDF fields, cellular automata, simulations — use built-in 3D viewer or pixel-perfect 2D grids.
SDF 场、元胞自动机、仿真 — 使用内置 3D 查看器或像素级 2D 网格。
torch.Tensor. No staging, no CPU round-trip.torch.Tensor，无中转，无 CPU 回传。import torch, vultorch
texture = torch.rand(512, 512, 4, device="cuda")
view = vultorch.View("Viewer", 800, 600)
panel = view.panel("Output")
panel.canvas("main").bind(texture)
view.run() # opens window, renders at 60 fps# Vulkan shared memory → torch.Tensor — zero memcpy
tensor = vultorch.create_tensor(512, 512, channels=4)
tensor[:] = model(input) # write directly, visible on screenview = vultorch.View("Training", 1000, 700)
# ... set up panels, canvases, controls ...
while view.step():
loss = train_one_step()
update_display(output_tensor)
view.end_step() # render frame13 runnable examples — from hello-world to a full neural rendering workstation and RL agents.
13 个可运行示例 — 从 hello-world 到完整的神经渲染工作站与强化学习智能体。
Requires the Vulkan SDK and CMake 3.25+.
需要 Vulkan SDK 和 CMake 3.25+。
git clone --recursive https://github.com/ChenlizheMe/Vultorch.git
cd Vultorch
# Windows
cmake --preset release-windows
cmake --build --preset release-windows
# Linux / WSL2
cmake --preset release-linux
cmake --build --preset release-linux
pip install dist/vultorch-*.whl