特性：

• 支持：Conv2d、ResNet BasicBlock、Linear、MaxPool2d、BatchNorm2d、GAP、Softmax等神经网络模块（参考infras/myresnet.h）
• 基于C++实现；兼容PyTorch训练框架
• 高效的张量（Tensor，参考infras/myresnet.h）计算；支持张量级优化
• 基于OpenBLAS线性运算库的底层自主实现
• 预置ResNet32网络，并在GTSRB数据集上完成测试

Features:

• Support: Conv2d, ResNet BasicBlock, Linear, MaxPool2d, BatchNorm2d, GAP, Softmax, etc. (refer to infras/myresnet.h)
• Implemented in C++; Compatible with PyTorch training framework
• Efficient Tensor computations (refer to infras/tensor.h); Tensor-level optimization supported
• Low-level customization: OpenBLAS-based
• ResNet32 implemented, tested on GTSRB dataset

make

或者 or

make openblas # install OpenBLAS and opencv (~10min)
make -C ./data/traffic/ # download GTSRB dataset (depends on your network)
make -C ./infras/ poc.out # compile the inference framework (<1min)

导出PyTorch训练的模型

cd model
python train.py
python export_model.py

使用MLInfras完成模型预测

make -C ./infras/