Riptide is a collection of functions, scripts, and tools that enable ultra low-bitwidth neural networks to be easily trained and deployed at high speeds. Riptide is built on top of Tensorflow for training and TVM for deployment. Riptide models uses a novel operator called Fused Glue to replace all floating point operations inside of a binary neural network. By combining Fused Glue layers with other optimizations such as Bitpack Fusion, Riptide is able to generate models that run 4-12X faster than floating point equivalents on the Raspberry Pi. For full implementation details see our paper, which was presented at MLSys 2020 and can be found here.

Getting Started

Requirements

Riptide works best with Tensorflow 2.1. If you don't have tensorflow, you can install it with pip install tensorflow.

Next, we can install the prerequisites to TVM

sudo apt-get update
sudo apt-get install -y python3 \\
                        python3-dev \\
                        python3-setuptools gcc  \\
                        libtinfo-dev zlib1g-dev \\
                        build-essential \\
                        cmake \\
                        libedit-dev \\
                        libxml2-dev

pip3 install --user numpy \\
                    decorator \\
                    attrs \\
                    tornado \\
                    psutil \\
                    xgboost                        

Installing Riptide

First, recursively clone Riptide (which has a custom fork of TVM as a submodule).

git clone --recursive git@github.com:jwfromm/riptide.git

Next we need to build the TVM submodule and set up our environment variables to properly detect it.

cd Riptide/tvm && mkdir build && cp cmake/config.cmake build && cd build
cmake ..
make -j4
export TVM_HOME={RiptideLocation}/tvm
export PYTHONPATH=$TVM_HOME/python:$TVM_HOME/topi/python:$PYTHONPATH

Note that if you want to compile a model for an embedded platform like the Raspberry Pi, you'll need to install llvm-dev and set USE_LLVM to ON in config.cmake.

Consider adding the above environment variables to your .bashrc to save time later.

You should now be able to import Riptide in Python and are ready to train and deploy a binary model!

We also provide a prebuilt docker image in Riptide/docker to make deployment of Riptide across environments easier.

Training a Binary Model

Riptide provides implementations of various binary layers and functions in binary_layers.py and binary_funcs.py respectively. Although you don't need to examine or change these files to train a model, they are written to be easy to read and adjust for other low-bit algorithms. We provide a selection of binary models and their floating point equivalents in riptide/models. These include Alexnet, VGGNet, various Resnets, and SqueezeNet. You can create your own binary models by using BinaryConv2D, BinaryDense, and BatchNormalization imported from binary_layers.

To train a model, navigate to scripts and take a look at train_imagenet.py. This script provides a simple and efficient interface for training models on the ImageNet Dataset. We use Tensorflow Datasets to prepare and load images so you'll first need to download ImageNet and have tfds.load generate tfrecords.

Once ImageNet is ready, you can start a training job as follows:

python train_imagenet.py --model alexnet --experiment 2A1W --gpus 0,1,2,3 --binary --bits 2 --model_dir ~/models

This will start training an alexnet binarized with 2 bit activations and 1 bit weights on 4 GPUs. Checkpoints and tensorboard logs will be saved to model_dir/alexnet_2A1W. Riptide automatically logs quite a bit of useful information during training including binary histograms that are pretty neat. To look at these logs run.

tensorboard --logdir ~/models/

Then open a browser and navigate to localhost:6006.

Training should take somewhere between a day or two to a few weeks depending on your model and number of GPUs. Once finished, you can load the trained model as follows:

First recreate the model architecture.

import tensorflow as tf
from riptide.binary.binary_funcs import *
from riptide.binary.binary_layers import Config
from riptide.get_models import get_model
actQ = DQuantize
weightQ = XQuantize
config = Config(actQ=actQ, weightQ=weightQ, bits=2.0)
with config:
  model = get_model('alexnet')

Then we can load the checkpoint weights after initializing shapes.

dummy_in = tf.keras.layers.Input(shape=[224, 224, 3], batch_size=1)
dummy_out = model(dummy_in)
model.load_weights('~/models/alexnet_2A2W/model.ckpt-xxxxx)

Where xxxxx is the checkpoint identifier you want to load.

Deploying a Binary Model

Once you've trained a binary model and are ready to run it on something like a Raspberry Pi, it's quite simple to convert the keras model and weights into a Relay representation.

import tvm
from tvm import relay
mod, params = relay.frontend.from_keras(
  model, 
  shape={'input_1': [1, 224, 224, 3]}, 
  layout='NHWC')

Then, we can compile the relay graph for a specific hardware platform, in this case an ARM cpu.

target = tvm.garget.arm_cpu("rasp3b")
with relay.build_config(opt_level=3):
  graph, lib, params = relay.build(mod, target=target, params=params)

The output of relay.build is a set of artifacts that can be used to run our network on an ARM CPU using the TVM runtime. One simple way to do that is through a TVM RPC server.

from tvm import autotvm
from tvm.contrib import util
import tvm.contrib.graph_runtime as runtime

# Export the runtime library
tmp = util.tempdir()
lib_fname = tmp.relpath('net.tar')
lib.export_library(lib_fname)

# Connect to the RPC server.
remote = autotvm.measure.request_remote(
  'rasp3b', 'tracker', 9191, timeout=10000)

# Upload library and prepare to run.
remote.upload(lib_fname)
rlib = remote.load_module('net.tar')
# Create TVM runtime.
module = runtime.create(graph, rlib, ctx)
# Upload model parameters.
module.set_input(**params)

# Set input and run the model
module.set_input(0, np.random.uniform(size=(1, 224, 224, 3)))
module.run()
print(module.get_output(0))
)

We can also easily measure runtime using a TVM time_evaluator.

ftimer = module.module.time_evaluator(
  "run", remote.cpu(), number=10, repeat=1)
prof_res = np.array(ftimer().results) * 1000  # Convert to milliseconds
print("Mean inference time (std dev): %.2f ms (%.2f ms)" %
      (np.mean(prof_res), np.std(prof_res)))

Other Useful Things

There are a bunch of potentially useful Jupyter notebooks located in notebooks. Digging through some might help find useful examples depending on what you're trying to do.

We also provide an implementation of Riptide inside of the LARQ library, which is a similar binary network training framework. If you're interested in this implementation please see the riptide branch of our fork and the corresponding riptide branch of the LARQ model zoo.