MobilenetV1 Imagenet Classification on GAP8

This reference design aims at demonstrating the end-to-end deployment of a state-of-the-art deep netowork, such as a quantized INT8 MobilenetV1, on the GAP8 platfom. Throuought this project we show:

• How to generate C code for the infernce task with quantized network network
• How to run a NN inference task on GAP8 chip
• How to train and dump a INT8 quantized network from Pytorch

Getting Started

Requirements

To run this example code, the SDK tag > 3.1 must be installed from here, optionally including the [Autotiler] (https://github.com/GreenWaves-Technologies/gap_sdk#getting-the-autotiler) tool. For training or dumping the network model, Pytorch==1.1 with GPU support must be installed.

Running the inference task

You can run the compile the code and run the inference task on a GAPUINO board ith the command:

make clean all run

To avoid reloading data on the FLASH memory at every run:

make clean all run ALREADY_FLASHED=1

Alternatively, the code can run on the [virtual platform] (https://github.com/GreenWaves-Technologies/gap_sdk#using-the-virtual-platform):

make run platform=gvsoc

If the Autotiler tool is installed, the MobileNetV1 C code can be regenereted by:

make clean model

Dumping MobilenetV1 graph description from Pytorch

python3 dump_mobilenet_to_GAP8.py --ckp_file './fake_quant_models/imagenet/mobilenetv1_224_1.0/checkpoint.pth.tar' --img_file './image144.jpeg'

Quantization-aware retraining

To replicate the quantization-aware retraining flow, you can type:

cd training/
python3 main.py -a mobilenet --mobilenet_width 1.0 --mobilenet_input 224 --save Imagenet/mobilenet_224_1.0 --dataset imagenet --type_quant 'PerLayerSymPACT' --weight_bits 8 --activ_bits 7 --activ_type learned --gpus 0,1,2,3 -j 32 --epochs 12 -b 128 --save_check --quantizer --batch_fold_delay 1 --batch_fold_type folding_weights

More details can be found [here].

Performance

The table below details the performance when running a MobilenetV1 model scoring 70% on Imagenet on a GAPUINO board.