hls4ml_dense_relu_merge

Testing the hls4ml optimizer pass for merging the ReLU layer into the Dense/Conv2D layers when ReLU immediately follows them---a frequently encountered pattern in neural networks (NNs).

Setup Dev Environment

We use venv to set up our virtual environment.

If you don't already have venv, install:

python3 -m pip install --user --upgrade pip

python3 -m pip install --user virtualenv

Create a virtual environment in this repo

python3 -m venv env

and then activate it (before install and using packages in the virtual environment)

source env/bin/activate

Confirm you're in the virtual environment by checking the path of the Python interpreter:

which python

as it should be in the env directory:

env/bin/python

If you wish to leave the virtual environment, deactivate it:

deactivate

I found that I needed to separately install graphviz/dot on my system using sudo apt install graphviz. pip installing graphviz was not sufficient.

Install the requirements:

python3 -m pip install -r requirements.txt

To upgrade a specific requirement:

python3 -m pip install --upgrade REQUIREMENT_NAME

To synthesize the models, I use Vivado HLS 2020.1.

Train and synthesize a model

I provide two models: 1) a fully connected NN and 2) a convolutional NN. Both are to be trained on the MNIST dataset. Note: Currently the SkipOptimizers config option doesn't work in hls4ml, so there isn't a way to directly test turning off the relu merge optimization.

Fully connected NN

To train the fully-connected network, run

python3 train.py -c dense.yaml

To convert the model into an hls4ml model and Vivado HLS project, run:

python3 convert.py -c dense.yaml

Convolutional NN

To train the CNN, run

python3 train.py -c conv.yaml