折腾笔记[39]-使用Scala3的Storch计算

news/2025/12/8 0:38:26/文章来源:https://www.cnblogs.com/qsbye/p/19319227

摘要

使用Scala3的Storch(对标PyTorch)简单计算张量.

前言

本文目的是分享人工踩坑经验, AI搜索引擎可以更快给出正确结果(用于投喂AI😂).

简介

bytedeco系列javacpp库

[https://github.com/bytedeco]
在 AI 产业里，Java 平台主要承担“生产部署”的角色。因此，让 Java 能够一键调用深度模型部署所需的全部工具与框架，就成了顺理成章的需求。然而，最常用的框架之一 Apache TVM，过去对 Java 并不友好——直到最近 JavaCPP Presets for TVM 出现，才补齐了这块短板。

不过，TVM 的 GitHub 仓库里仍有超过 50% 的代码是 Python。这意味着，大多数场景下，部署链路里免不了要跑几段 Python 模块；再加上预处理脚本往往只有 Python 实现。与其硬搬，不如让 Java 直接无痛调用这些 Python 代码。于是，JavaCPP Presets for CPython、LLVM、MKL、oneDNN（DNNL/MKL-DNN）、OpenBLAS、NumPy、SciPy 应运而生。它们被打包成普通 JAR，像依赖任何第三方库一样直接引入；同时，这些 JAR 也能自动满足 TVM 的全部原生依赖——无需本地安装任何组件！

例如，在下面的 OpenCV 示例项目里，我们把所有依赖（含 Python 运行时）一并打包进 GraalVM Native Image，配合 Quarkus 即可“开箱即用”：
[https://github.com/bytedeco/sample-projects/tree/master/opencv-stitching-native]

In the AI industry, the Java platform is used mainly when deploying to production. Consequently, it makes sense to provide easy access to all the tools and frameworks necessary to deploy deep learning models for Java. However, one of the most widely used frameworks, Apache TVM, did not have very good support for Java, until recently. With the JavaCPP Presets for TVM, we now have proper support for the Java platform.

Still, according to the GitHub repository of TVM, over 50% of the code is written in Python. This means that, for most use cases, we probably need to run some of those modules as part of the deployment pipeline. Moreover, there are usually preprocessing algorithms among other things that have been implemented only as Python scripts. For these reasons, we might as well make it easy as possible to execute this Python code from Java. This is where the JavaCPP Presets for CPython, LLVM, MKL, oneDNN (aka DNNL, MKL-DNN), OpenBLAS, NumPy, and SciPy come in. They are available as normal JAR files, which can be used as any other JAR files, but they can also be used to satisfy the dependencies of TVM—without having to install anything! For example, as shown in the following sample project for OpenCV, we can bundle everything with GraalVM Native Image as usual and it simply just works as expected with Quarkus:
[https://github.com/bytedeco/sample-projects/tree/master/opencv-stitching-native]

<dependency><groupId>org.bytedeco</groupId><artifactId>javacpp-platform</artifactId><version>1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>opencv-platform</artifactId><version>4.10.0-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>ffmpeg-platform</artifactId><version>7.1-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>flycapture-platform</artifactId><version>2.13.3.31-1.5.9</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>spinnaker-platform</artifactId><version>4.0.0.116-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>libdc1394-platform</artifactId><version>2.2.6-1.5.9</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>libfreenect-platform</artifactId><version>0.5.7-1.5.9</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>libfreenect2-platform</artifactId><version>0.2.0-1.5.9</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>librealsense-platform</artifactId><version>1.12.4-1.5.9</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>librealsense2-platform</artifactId><version>2.53.1-1.5.9</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>videoinput-platform</artifactId><version>0.200-1.5.9</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>artoolkitplus-platform</artifactId><version>2.3.1-1.5.9</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>chilitags-platform</artifactId><version>master-1.5.8</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>flandmark-platform</artifactId><version>1.07-1.5.8</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>arrow-platform</artifactId><version>6.0.1-1.5.8</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>hdf5-platform</artifactId><version>1.14.3-1.5.10</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>hyperscan-platform</artifactId><version>5.4.2-1.5.10</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>lz4-platform</artifactId><version>1.9.4-1.5.8</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>mkl-platform</artifactId><version>2025.0-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>mkl-dnn-platform</artifactId><version>0.21.5-1.5.8</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>dnnl-platform</artifactId><version>3.6.1-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>openblas-platform</artifactId><version>0.3.28-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>arpack-ng-platform</artifactId><version>3.9.1-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>cminpack-platform</artifactId><version>1.3.11-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>fftw-platform</artifactId><version>3.3.10-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>gsl-platform</artifactId><version>2.8-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>cpython-platform</artifactId><version>3.13.0-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>numpy-platform</artifactId><version>2.1.3-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>scipy-platform</artifactId><version>1.14.1-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>gym-platform</artifactId><version>0.26.2-1.5.8</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>llvm-platform</artifactId><version>19.1.3-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>libffi-platform</artifactId><version>3.4.6-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>libpostal-platform</artifactId><version>1.1-1.5.8</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>libraw-platform</artifactId><version>0.21.2-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>leptonica-platform</artifactId><version>1.85.0-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>tesseract-platform</artifactId><version>5.5.0-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>caffe-platform</artifactId><version>1.0-1.5.8</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>openpose-platform</artifactId><version>1.7.0-1.5.8</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>cuda-platform</artifactId><version>12.6-9.5-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>nvcodec-platform</artifactId><version>12.2.72-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>opencl-platform</artifactId><version>3.0-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>mxnet-platform</artifactId><version>1.9.1-1.5.8</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>pytorch-platform</artifactId><version>2.5.1-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>sentencepiece-platform</artifactId><version>0.2.0-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>tensorflow-platform</artifactId><version>1.15.5-1.5.8</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>tensorflow-lite-platform</artifactId><version>2.18.0-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>tensorrt-platform</artifactId><version>10.6-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>tritonserver-platform</artifactId><version>2.51.0-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>ale-platform</artifactId><version>0.8.0-1.5.8</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>depthai-platform</artifactId><version>2.24.0-1.5.10</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>onnx-platform</artifactId><version>1.17.0-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>ngraph-platform</artifactId><version>0.26.0-1.5.8</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>onnxruntime-platform</artifactId><version>1.20.0-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>tvm-platform</artifactId><version>0.18.0-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>bullet-platform</artifactId><version>3.25-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>liquidfun-platform</artifactId><version>master-1.5.8</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>qt-platform</artifactId><version>5.15.2-1.5.8</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>skia-platform</artifactId><version>2.88.3-1.5.8</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>cpu_features-platform</artifactId><version>0.7.0-1.5.8</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>modsecurity-platform</artifactId><version>3.0.8-1.5.8</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>systems-platform</artifactId><version>1.5.11</version></dependency><!-- Additional dependencies required to use CUDA, cuDNN, NCCL, and TensorRT --><dependency><groupId>org.bytedeco</groupId><artifactId>opencv-platform-gpu</artifactId><version>4.10.0-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>caffe-platform-gpu</artifactId><version>1.0-1.5.8</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>mxnet-platform-gpu</artifactId><version>1.9.1-1.5.8</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>pytorch-platform-gpu</artifactId><version>2.5.1-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>tensorflow-platform-gpu</artifactId><version>1.15.5-1.5.8</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>onnxruntime-platform-gpu</artifactId><version>1.20.0-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>tvm-platform-gpu</artifactId><version>0.18.0-1.5.11</version></dependency><!-- Optional dependencies to load Python-enabled builds --><dependency><groupId>org.bytedeco</groupId><artifactId>tensorflow-platform-python</artifactId><version>1.15.5-1.5.8</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>tensorflow-platform-python-gpu</artifactId><version>1.15.5-1.5.8</version></dependency><!-- Additional dependencies to use bundled CUDA, cuDNN, NCCL, and TensorRT --><dependency><groupId>org.bytedeco</groupId><artifactId>cuda-platform-redist</artifactId><version>12.6-9.5-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>tensorrt-platform-redist</artifactId><version>10.6-1.5.11</version></dependency><!-- Additional dependencies to use bundled full version of MKL --><dependency><groupId>org.bytedeco</groupId><artifactId>mkl-platform-redist</artifactId><version>2025.0-1.5.11</version></dependency><dependency><groupId>org.bytedeco</groupId><artifactId>javacv-platform</artifactId><version>1.5.11</version></dependency>

清单:

Prebuilt Java Bindings to C/C++ Libraries
These are part of a project that we call the JavaCPP Presets. Many coexist in the same GitHub repository, and all use JavaCPP to wrap predefined C/C++ libraries from open-source land. The bindings expose almost all of the relevant APIs and make them available in a portable and user-friendly fashion to any Java virtual machine (including Android), as if they were like any other normal Java libraries. We have presets for the following C/C++ libraries:
OpenCV – [sample usage] [API] – More than 2500 optimized computer vision and machine learning algorithms
FFmpeg – [sample usage] [API] – A complete, cross-platform solution to record, convert and stream audio and video
FlyCapture – [sample usage] [API] – Image acquisition and camera control software from PGR
Spinnaker – [sample usage] [API] – Image acquisition and camera control software from FLIR
libdc1394 – [sample usage] [API] – A high-level API for DCAM/IIDC cameras
OpenKinect – [sample usage] [API] [API 2] – Open source library to use Kinect for Xbox and for Windows sensors
librealsense – [sample usage] [API] [API 2] – Cross-platform library for Intel RealSense depth and tracking cameras
videoInput – [sample usage] [API] – A free Windows video capture library
ARToolKitPlus – [sample usage] [API] – Marker-based augmented reality tracking library
Chilitags – [sample usage] [API] – Robust fiducial markers for augmented reality and robotics
flandmark – [sample usage] [API] – Open-source implementation of facial landmark detector
Arrow – [sample usage] [API] – A cross-language development platform for in-memory data
HDF5 – [sample usage] [API] – Makes possible the management of extremely large and complex data collections
Hyperscan – [sample usage] [API] – High-performance regular expression matching library
LZ4 – [sample usage] [API] – Extremely fast compression algorithm
MKL – [sample usage] [API] – The fastest and most-used math library for Intel-based systems
oneDNN – [sample usage] [API] [API 2] – Intel Math Kernel Library for Deep Neural Networks (DNNL)
OpenBLAS – [sample usage] [API] – An optimized BLAS library based on GotoBLAS2 1.13 BSD version, plus LAPACK
ARPACK-NG – [sample usage] [API] – Collection of subroutines designed to solve large scale eigenvalue problems
CMINPACK – [sample usage] [API] – For solving nonlinear equations and nonlinear least squares problems
FFTW – [sample usage] [API] – Fast computing of the discrete Fourier transform (DFT) in one or more dimensions
GSL – [sample usage] [API] – The GNU Scientific Library, a numerical library for C and C++ programmers
CPython – [sample usage] [API] – The standard runtime of the Python programming language
NumPy – [sample usage] [API] – Base N-dimensional array package
SciPy – [sample usage] [API] – Fundamental library for scientific computing
Gym – [sample usage] [API] – A toolkit for developing and comparing reinforcement learning algorithms
LLVM – [sample usage] [API] – A collection of modular and reusable compiler and toolchain technologies
libffi – [sample usage] [API] – A portable foreign-function interface library
libpostal – [sample usage] [API] – For parsing/normalizing street addresses around the world
LibRaw – [sample usage] [API] – A simple and unified interface for RAW files generated by digital photo cameras
Leptonica – [sample usage] [API] – Software useful for image processing and image analysis applications
Tesseract – [sample usage] [API] – Probably the most accurate open source OCR engine available
Caffe – [sample usage] [API] – A fast open framework for deep learning
OpenPose – [sample usage] [API] – Real-time multi-person keypoint detection for body, face, hands, and foot estimation
CUDA – [sample usage] [API] – Arguably the most popular parallel computing platform for GPUs
NVIDIA Video Codec SDK – [sample usage] [API] – An API for hardware accelerated video encode and decode
OpenCL – [sample usage] [API] – Open standard for parallel programming of heterogeneous systems
MXNet – [sample usage] [API] – Flexible and efficient library for deep learning
PyTorch – [sample usage] [API] – Tensors and dynamic neural networks with strong GPU acceleration
SentencePiece – [sample usage] [API] – Unsupervised text tokenizer for neural-network-based text generation
TensorFlow – [sample usage] [API] – Computation using data flow graphs for scalable machine learning
TensorFlow Lite – [sample usage] [API] – An open source deep learning framework for on-device inference
TensorRT – [sample usage] [API] – High-performance deep learning inference optimizer and runtime
Triton Inference Server – [sample usage] [API] – An optimized cloud and edge inferencing solution
ALE – [sample usage] [API] – The Arcade Learning Environment to develop AI agents for Atari 2600 games
DepthAI – [sample usage] [API] – An embedded spatial AI platform built around Intel Myriad X
ONNX – [sample usage] [API] – Open Neural Network Exchange, an open source format for AI models
nGraph – [sample usage] [API] – An open source C++ library, compiler, and runtime for deep learning frameworks
ONNX Runtime – [sample usage] [API] – Cross-platform, high performance scoring engine for ML models
TVM – [sample usage] [API] – An end to end machine learning compiler framework for CPUs, GPUs and accelerators
Bullet Physics SDK – [sample usage] [API] – Real-time collision detection and multi-physics simulation
LiquidFun – [sample usage] [API] – 2D physics engine for games
Qt – [sample usage] [API] – A cross-platform framework that is usually used as a graphical toolkit
Skia – [sample usage] [API] – A complete 2D graphic library for drawing text, geometries, and images
cpu_features – [sample usage] [API] – A cross platform C99 library to get cpu features at runtime
ModSecurity – [sample usage] [API] – A cross platform web application firewall (WAF) engine for Apache, IIS and Nginx
Systems – [sample usage] [API] – To call native functions of operating systems (glibc, XNU libc, Win32, etc)
Add here your favorite C/C++ library, for example: Caffe2, OpenNI, OpenMesh, PCL, etc. Read about how to do that.

Scala3简介

[https://scala-lang.org.cn]
[https://scala-lang.org/]
[https://www.scala-js.org/]
一门与你共同成长、从几行脚本到跨平台大型应用的编程语言。
Scala 让你用更少代码，做更多事情。作为一门高级语言，它的现代特性显著提升生产力，让代码更易读。在 Scala 中，你可以将函数式与面向对象两种范式自由组合，优雅地组织程序结构。
凭借 JVM、JavaScript 与 Native 三大运行时，Scala 天生适合构建高速、并发、分布式系统。它把“互操作”放在首位，轻松调用无数经产业验证的库，坐拥庞大生态。
Scala 的静态类型系统默认帮你打造安全可靠的系统：智能的内置检查、可操作的报错信息、线程安全的数据结构与集合，让许多棘手 bug 在程序首次运行前就被消灭。

我该选哪个版本？
Scala Next 当前 3.7.4——默认推荐，功能最新、修复最全，适合大多数用户。
Scala LTS 当前 3.3.7——长期支持版，建议用于发布库。

安装 Scala，最省心的是用 cs setup（由 Coursier 驱动的官方安装器）。一条命令，即可装好使用最新 Scala 所需的一切工具。

A programming language that scales with you: from small scripts to large multiplatform applications.
Scala lets you write less to do more. As a high-level language, its modern features increase productivity and lead to more readable code. With Scala, you can combine both functional and object-oriented programming styles to help structure programs.
Scala is well suited to building fast, concurrent, and distributed systems with its JVM, JavaScript and Native runtimes. Scala prioritizes interoperability, giving easy access to many ecosystems of industry-proven libraries.
Scala's static types help you to build safe systems by default. Smart built-in checks and actionable error messages, combined with thread-safe data structures and collections, prevent many tricky bugs before the program first runs.
Which version of Scala should I choose?
Scala Next currently 3.7.4 - The default to be used by most users, containing the latest features, bug fixes and improvements.
Scala LTS currently 3.3.7 - Advised to be used for publishing libraries.
To install Scala, it is recommended to use cs setup, the Scala installer powered by Coursier. It installs everything necessary to use the latest Scala release from a command line.

Storch简介

GPU accelerated deep learning and numeric computing for Scala 3.
STorch is a Scala library for fast tensor computations and deep learning, same as PyTorch.
Like PyTorch, STorch provides 📇 🏠 🪟 🍎 🐧
Both Compile from Cpp LibTorch, STorch has Same Equal Position with Pytorch
A NumPy like API for working with tensors
GPU support [Future will support HuaWei CANN NPU]
Automatic differentiation on Linux | Mac | Windows | IOS | Andriod
A neural network API for building and training neural networks.
What All you need Pytorch features could find in STorch !!!
Minimal time and learning cost migrate from python pytorch developer
Maximal ROI for Big Data developer and Data Scientist
Support LLM | Recommend System | Compute Vision | NLP | Financial Tech | RL |Big Data | Data Science Research
Big Deep Learning Environment Chains boost you concentrate on development
Support Java SMID Vector API, Please use JDK Version >= 17
Could Use NCCL | GLOO | RAY | SPARK | FLINK for DISTRIBUTE TRAINING and inference [Future will support UCC | MPI]
Best Data Driven Process and Monad Feature Engineer Model Pipeline Partner for Apache Spark | Flink | Polar
STorch aims to close to the Python API to make porting existing models and the life of people already familiar with PyTorch easier.

实现

1. 安装Scala3

Scala 3.7.2

# macOS
brew install coursier/formulas/coursier && cs setup# Linux
curl -fL https://github.com/coursier/coursier/releases/latest/download/cs-x86_64-pc-linux.gz | gzip -d > cs && chmod +x cs && ./cs setup
curl -fL https://github.com/VirtusLab/coursier-m1/releases/latest/download/cs-aarch64-pc-linux.gz | gzip -d > cs && chmod +x cs && ./cs setup# Windows
# 使用安装包: [https://github.com/coursier/coursier/releases/latest/download/cs-x86_64-pc-win32.zip]

2. 手动编译STorch的jar包

原因: io.github.mullerhai链接重定向有问题

# 这个版本更新一点
git clone https://github.com/bytedeco/pytorch.git
cd pytorch
sbt publishLocal

3. 构建工程

初始化工程:

sbt new scala/scala3.g8

build.sbt

val scala3Version = "3.7.2"lazy val root = project.in(file(".")).settings(name         := "hello_storch",version      := "0.1.0-SNAPSHOT",scalaVersion := scala3Version,/* ---------- 仓库配置 ---------- */resolvers ++= Seq(Resolver.mavenCentral,Resolver.sonatypeCentralSnapshots,// 存放Storch仓库jar包Resolver.file("local", file(System.getProperty("user.home") + "/.ivy2/local"))(Resolver.ivyStylePatterns)),/* ---------- 3. 依赖 ---------- */libraryDependencies ++= Seq("io.github.mullerhai" % "storch_core_3"  % "0.7.6-1.5.12","org.bytedeco"        % "pytorch-platform" % "2.5.1-1.5.11","org.scalameta"       %% "munit"         % "1.0.0" % Test),/* 原生库需要 fork */fork := true,/* 设置编码以支持中文 */javaOptions += "-Dfile.encoding=UTF-8",Compile / compile / javacOptions ++= Seq("-encoding", "UTF-8"),Compile / scalacOptions ++= Seq("-encoding", "UTF-8"),/* 设置环境变量以支持中文输出 */envVars := Map("JAVA_TOOL_OPTIONS" -> "-Dfile.encoding=UTF-8 -Dsun.stdout.encoding=UTF-8 -Dsun.stderr.encoding=UTF-8","file.encoding" -> "UTF-8","sun.jnu.encoding" -> "UTF-8","sun.stdout.encoding" -> "UTF-8","sun.stderr.encoding" -> "UTF-8"),/* 确保控制台输出使用正确编码 */run / javaOptions ++= Seq("-Dfile.encoding=UTF-8","-Dsun.stdout.encoding=UTF-8", "-Dsun.stderr.encoding=UTF-8"))

src/main/scala/Main.scala

import torch._
import torch.ops._@main def hello(): Unit =println("Hello world!")// 创建张量的基本示例println("=== 创建张量示例 ===")// 从序列创建张量val data = Seq(0, 1, 2, 3)// data: Seq[Int] = List(0, 1, 2, 3)val t1 = Tensor(data)// t1: Tensor[Int32] = dtype=int32, shape=[4], device=CPU // [0, 1, 2, 3]println(s"t1: $t1")// 检查张量是否等于arange生成的张量val arangeTensor = arange(0, 4, dtype = int32)println(s"t1.equal(arange(0,4)): ${t1.equal(arangeTensor)}")// 数据类型转换val t2 = t1.to(dtype = float32)// t2: Tensor[Float32] = dtype=float32, shape=[4], device=CPU // [0.0000, 1.0000, 2.0000, 3.0000]println(s"t2 (转换为float32): $t2")// 张量加法val t3 = t1 + t2// t3: Tensor[Float32] = dtype=float32, shape=[4], device=CPU // [0.0000, 2.0000, 4.0000, 6.0000]println(s"t3 = t1 + t2: $t3")println("\n=== 随机张量和特殊张量 ===")// 创建随机张量val shape = Seq(2, 3)// shape: Seq[Int] = List(2, 3)val randTensor = rand(shape)// randTensor: Tensor[Float32] = dtype=float32, shape=[2, 3], device=CPU // [[0.4341, 0.9738, 0.9305],//  [0.8987, 0.1122, 0.3912]]println(s"随机张量: $randTensor")// 创建零张量val zerosTensor = zeros(shape, dtype = int64)// zerosTensor: Tensor[Int64] = dtype=int64, shape=[2, 3], device=CPU // [[0, 0, 0],//  [0, 0, 0]]println(s"零张量: $zerosTensor")println("\n=== 简单神经网络示例 ===")// 创建输入张量val x = ones(Seq(5))// x: Tensor[Float32] = dtype=float32, shape=[5], device=CPU // [1.0000, 1.0000, 1.0000, 1.0000, 1.0000]println(s"输入x: $x")// 创建权重矩阵（需要梯度）val w = randn(Seq(5, 3), dtype = float32, requires_grad = true)// w: Tensor[Float32] = dtype=float32, shape=[5, 3], device=CPU // 随机初始化的权重矩阵println(s"权重矩阵w: $w")// 创建偏置向量（需要梯度）val b = randn(Seq(3), dtype = float32, requires_grad = true)// b: Tensor[Float32] = dtype=float32, shape=[3], device=CPU // 随机初始化的偏置println(s"偏置b: $b")// 前向传播计算val z = (x.matmul(w)) + b// z: Tensor[Float32] = dtype=float32, shape=[3], device=CPU // 线性变换结果println(s"前向传播结果z = x·w + b: $z")println("\nStorch 张量操作演示完成！")

编译:

sbt update
sbt compile
sbt run

4. 效果

现在中文可以正常显示了：=== 创建张量示例 ===t2 (转换为float32): tensor dtype=float32, shape=[4], device=CPU=== 随机张量和特殊张量 ===随机张量: tensor dtype=float32, shape=[2, 3], device=CPU零张量: tensor dtype=int64, shape=[2, 3], device=CPU=== 简单神经网络示例 ===输入x: tensor dtype=float32, shape=[5], device=CPU权重矩阵w: tensor dtype=float32, shape=[5, 3], device=CPU偏置b: tensor dtype=float32, shape=[3], device=CPU前向传播结果z = x·w + b: tensor dtype=float32, shape=[3], device=CPUStorch 张量操作演示完成！