Prerequisites

PyTorch compatible GPU with CUDA 12.1 Conda (Python 3.9) Other packages are listed in requirements.txt.

Data

The data is not big. It contains $1600$ images in the train folder and $400$ images in the test folder. Firstly, download and unzip the data.zip.

Usage

Create a new conda environment:

conda create --name culvert python=3.9
activate culvert       # Windows
conda activate culvert # Linux

Clone this repo:

git clone https://github.com/SHUs-Lab/SHDA23YZ.git
cd SHDA23YZ

Install required packages by typing

pip install -r requirements.txt

Download data.zip and unzip it.

Then, run object_detection.py to train SPPNet. Also, you can run sppnet_search.py to search for neural architecture.

SPPNet parameters are defined below:

Model Summary

Layer (type:depth-idx)	Param #
SPPNet
├─Sequential: 1-1
│ └─Conv2d: 2-1	1,792
│ └─ReLU: 2-2
│ └─MaxPool2d: 2-3
│ └─Conv2d: 2-4	73,856
│ └─ReLU: 2-5
│ └─MaxPool2d: 2-6
│ └─Conv2d: 2-7	295,168
│ └─ReLU: 2-8
│ └─MaxPool2d: 2-9
├─ModuleList: 1-2
│ └─AdaptiveMaxPool2d: 2-10
│ └─AdaptiveMaxPool2d: 2-11
│ └─AdaptiveMaxPool2d: 2-12
├─Sequential: 1-3
│ └─Linear: 2-13	22,283,264
│ └─ReLU: 2-14
│ └─Linear: 2-15	2,050
---------------------------------------	-----------------
Total params:	22,656,130
Trainable params:	22,656,130
Non-trainable params:	0

For NAS search space

We explore the following search spaces for all three components:

Feature Engineering:

We define the search space for the filter size of the first convolutional layer as ranging from 1 to 9 (1, 3, 5, 7, 9).

SPP Layer:

We experiment with five different filter sizes for the first SPP (Spatial Pyramid Pooling) layer, spanning from 1 to 5 (1, 2, 3, 4, 5).

Fully-Connected Layers:

We customize the feature size for two fully-connected layers within the following ranges: 128, 256, 512, 1024, 2048, 4096, and 8192.