pytorch_tiramisu

pytorch_tiramisu is a python package that adds Tiramisu Compiler as a compiler backend to PyTorch Deep Learning Framework.

pytorch_tiramisu offers two modes of utilization:

1. Mode 1: The first mode is dedicated to non expert users that want to benefit from directly laveraging the compiler stack. The user will be using pre-compiled operators. (This mode is only available for CPUs). This mode does not require the installation of Tiramisu.

2. Mode 2: If you want to register more operators or apply some optimizations that are not yet supported by the package, you can install tiramisu and fully pass all the compiler stack, as the figure above illustrates.

Build & Install

1. Install the latest Nightly build of PyTorch. You can choose to install it from source for more efficient acceleration follow these instructions.
2. Install pytorch_tiramisu: We recommend to install the package from source, since it is still a research project in its infancy.

Using pypi

pip install pytorch_tiramisu

Install from Source

• Building pytorch_tiramisu from source requires the following packages :

  • cmake
  • automake
  • libtool

• You can check the Dockerfile which shows the environment we used for testing and building pytorch_tiramisu.

• Clone this repository:

git clone --recursive https://github.com/IHIaadj/tiramisu_pytorch.git 
cd tiramisu_pytorch

• Install the package: This file sets the paths to both PyTorch and pybind11.

# If mode 2 uncomment the fourth line from build.sh
./build.sh

Tests & Tutorials

You can test the installation by running the following code:

import torch 
import pytorch_tiramisu as pt 
pt.enable(jit=True) 

Take a look at one of our Jupyter notebooks to quickly try different features and deep learning models:

• Getting Started with pytorch_tiramisu
• ResNet in Tiramisu
• XLNet

Usage

• pytorch_tiramisu can hook into PyTorch JIT and compile the model operators that it supports. (if jit parameter is set to True)

import torch 
import torch.nn.functional as F
import pytorch_tiramisu as pt 
pt.enable(jit=True) 

# The following function will be compiled with Tiramisu
@torch.jit.script
def relu_(a):
    return F.relu(a) 

• Otherwise, pytorch_tiramisu.compile(model) can be used to perform the compilation of the deep learning model prior to running the final graph execution.

import torch 
import torch.nn.functional as F
import pytorch_tiramisu as pt 
pt.enable(jit=False) 

class Net(nn.Module):
    def __init__(self):
        super(Net, self).__init__()
        self.fc = nn.Linear(256, 10)

    def forward(self, x):
        x = self.fc3(x)
        return x
        
model = Net() 
a = torch.randn(1, 256)
generated = pt.compile(model(a)) # Execute an optimization pass and generate the operators.
pt.execute(generated)