DORY examples and tutorial

Topology present in tutorial

• Mobilenets V1/V2 - 4/8 bits - 128/224
• Dronet
• Feed-Forward Neural networks

Requirements

DORY

See DORY folder for all the requirements. gap_sdk and python packages

Installation

The execution of the dory example requires the following folders:

1. dory_example: contains examples to launch DORY.
2. dory: repository with the framework (submodule of dory_exampe)

Execute the following commands to clone all

git clone https://github.com/pulp-platform/dory_examples

Update modules

Before continuing, update the modules to the latest working version. Refer to the following branches as the most updated:

• dory_example --> master
• dory --> master

Execution

network_generate.py: generate files for a whole network. The file can be directly executed with:

python3 network_generate.py

Input example networks are included in the Test_suite_DORY folder. Folder application contains the whole application to be executed on the GAP8 platform.

There are many parameters that you can specify in the network_generate.py. Use

python3 network_generate.py -h

to see all the parameters which can be specified.

Logs files are produced inside logs folder, with network parsed and layers tiling.

Install and build a new network from scratch

First, source in the gap_sdk the platform you prefer

source sourceme.sh

Execute the following commands to generate the files and test a network on gvsoc with the latest version of GAP8 SDK.

python3 network_generate.py Quantlab GAP8 ./dory_examples/config_files/config_Quantlab_MV2_4bits.json --app_dir ./dory_examples/application
cd ${application_dir}
make clean all run CORE=8 platform=gvsoc

The variable CORE define the number of the cores of the cluster used for parallelization. The default is 8 (the maximum number of cores inside the cluster of GAP8).

Some examples of files generation networks:

python3 network_generate.py Quantlab GAP8 ./dory_examples/config_files/config_Quantlab_MV2_4bits.json --app_dir ./dory_examples/application
python3 network_generate.py Quantlab GAP8 ./dory_examples/config_files/config_Quantlab_MV2_8bits.json --app_dir ./dory_examples/application
python3 network_generate.py NEMO GAP8 ./dory_examples/config_files/config_NEMO_dronet.json --app_dir ./dory_examples/application

Note that for Quantlab MV2 networks you should use mixed-sw backend that supports signed input/output activations.