🧠 What is MTensor?

MTensor is a CPU-based deep learning library written in C++, inspired by PyTorch's design. It supports dynamic graphs, training, autograd, and a wide range of tensor and neural network operations.

⚠️ Beta Notice: MTensor 2.0 is tested on Windows (MSVC 17) and Linux (GCC 11+/Clang 14+). macOS support is coming soon.

---

🚀 Technologies Used

• oneDNN – High-performance tensor primitives
• OpenMP – Threaded parallelism
• AVX512 – Custom fused operations for optimizers, reductions
• CImg + libpng – Image handling support
• nlohmann/json – Exporting graphs and manual tensor initialization
• vcpkg – Dependency manager

---

🧱 Features

🔹 Core Tensor API

• Internal TensorImpl with smart pointers
• Tensor methods: shape(), stride(), numel(), randn(), ones(), arange(), fill(), ...
• Torch-style formatted printing
• Broadcasting support (binary ops, matmul)

🔹 Tensor Operations

• Arithmetic & Logic: +, -, *, /, <, >, ==, min, max
• Elementwise: abs, clip, exp, log, pow, relu, sigmoid, ...
• Reduction: sum, mean, min, max, mul, norm_lp_*, ...
• Normalization: batchnorm, groupnorm, layernorm, rmsnorm
• Convolution: conv1d/2d/3d, transposedConv1d/2d/3d
• Pooling: avgpool, maxpool (1d, 2d, 3d)
• Memory & Views: clone, reshape, permute, transpose, squeeze, contiguous
• Matrix ops: matmul, A @ B + C
• Other: embedding, softmax, logsoftmax, cat(), stack()

🔹 Autograd & Graphing

• Dynamic computation graph with backward()
• Tensor gradient helpers: requires_grad, detach, grad, ...
• Visualize autograd graph as an interactive HTML (build_graph())

---

🧠 Neural Networks

• Layers: Linear, Conv, TransposeConv, Embedding, Dropout, Flatten, Unflatten, Sequential
• Activations: ReLU, LeakyReLU, Sigmoid, Tanh, Softmax, LogSoftmax
• Losses: MSELoss, L1Loss, CrossEntropy, KLDiv, BCEWithLogits
• Optimizers (AVX512-fused): SGD, MomentumSGD, Adam, AdamW
• Initializers: Kaiming and Xavier (Uniform/Normal)

---

🗂️ Data Handling

• Datasets:

  • MNISTDataset (reads MNIST binary format)
  • ImageFolderDataset (uses CImg for loading images)

• Dataloader: For training batches

• Image I/O: image_to_tensor(), tensor_to_image() (via CImg)

---

⚙️ Installation

🧪 MTensor 2.0 is tested on Windows (MSVC 17) and Linux (GCC 11+/Clang 14+). All dependencies are managed with vcpkg.

---

Prerequisites

Windows

• Visual Studio 2022 (MSVC 17)
• CMake 3.31.6+

Linux (Ubuntu/Debian)

sudo apt-get install -y build-essential ninja-build cmake libomp-dev

Linux (Fedora/RHEL)

sudo dnf install -y gcc-c++ ninja-build cmake libomp-devel

Linux (Arch Linux / CachyOS)

sudo pacman -S gcc ninja cmake openmp

---

1. Clone the Repository

git clone --recurse-submodules https://github.com/your-username/MTensor.git
cd MTensor

✅ vcpkg is included as a submodule under external/vcpkg.

---

2. Set Up vcpkg

If needed:

git submodule update --init --recursive

Bootstrap vcpkg:

cd external/vcpkg
./bootstrap-vcpkg.bat     # Windows
./bootstrap-vcpkg.sh      # Linux

Install dependencies:

./vcpkg install

📦 Dependencies are defined in vcpkg.json.

🔧 On Linux, vcpkg will automatically use the x64-linux triplet.

---

3. Configure the Project

Windows (Visual Studio 2022)

Use the default CMake preset:

cmake --preset windows-vs2022

Linux (Ninja)

cmake --preset linux-ninja

🛠️ To change generator or toolchain, edit CMakePresets.json.

---

4. Build the Project

Windows

Build in release mode:

cmake --build --preset release

Or in debug mode:

cmake --build --preset debug

Linux

Build in release mode:

cmake --build --preset linux-release

Or in debug mode:

cmake --build --preset linux-debug

🧪 The build produces a shared library (MTensor.dll on Windows, libMTensor.so on Linux) in the build/ directory.

---

📌 Roadmap

• ✅ Tensor & autograd engine
• ✅ Neural network layers and optimizers
• ✅ Dataset/dataloader system
• ✅ Linux support
• 🔜 External project integration
• 🔜 macOS support
• 🔜 Python bindings
• 🔜 Training UI (browser-based)