Dead wood image tile classification on KV260

some intro

• four classes: debris, forest, water, other

Experimental Environment Setup

also some intro

Host Machine

only for CUDA-enabled host machine

docker pull tumbgd/vai-pt-cuda

Target Machine

Xilinx Kria KV260

Usage

On host machine

Download dataset here, then unzip.

0. Generate h5 dataset

1. Train model

2. Pruning the trained model

3. Quantizing the pruned trained model

4. Compiling the quantized pruned trained model (for on-board depolyment):

   vai_c_xir -x ./quantize_result/Model_int.xmodel -a /opt/vitis_ai/compiler/arch/DPUCZDX8G/KV260/arch.json -o dwc_ob -n dwc_ob

On target machine

Please follow the README in onboard.

Performances

Model size/accuracy trade-off

Model	Acc.	# param.	Size
non-opt	86.0%	205.28k	806kB
opt	85.2%	76.65k (62.6% smaller)	334kB (58.6% smaller)

Inference speed (FPS)

Device & Model	Inference Speed (FPS)
opt on KV260	211.07
non-opt on KV260	87.71
non-opt on laptop (CPU)	255.59
non-opt on laptop (GPU)	308.75

Remaining Questions

1. setup VART on Ubuntu 22.04 (currently on pre-built images with shabby GUI)

   • lower version VART seems okay, but much less functions supported.

2. bias_corr is None. Seems no error here (accuray hardly drops), but why not 0 rather than None

3. ReLU should be supported as stated in the Xilinx document but not in practice.

   • Maybe torch.nn.ReLU is not supported, but toerch.nn.functional.relu is. Need a try here.
   • Or just not supported. In this case, try all possible activation functions to avoid multiple subgraphs. Our aim is to run a model fully on a single DPU graph to aviod data copy between DPU and CPU.