Skip to content

Latest commit

 

History

History
executable file
·
98 lines (79 loc) · 3.69 KB

README.md

File metadata and controls

executable file
·
98 lines (79 loc) · 3.69 KB
Raw

Installation

Setup a tensorflow environment,

conda create -n yoloret python=3.9
conda activate yoloret
conda install -c anaconda tensorflow-gpu

Install additional packages

conda install -c anaconda pillow
conda install -c conda-forge opencv
conda install tqdm
pip install --upgrade neural_structured_learning

Preparing Dataset

VOC Dataset

Download and extract the VOC 2007 and 2012 dataset as shown below

wget https://pjreddie.com/media/files/VOCtrainval_11-May-2012.tar
wget https://pjreddie.com/media/files/VOCtrainval_06-Nov-2007.tar
wget https://pjreddie.com/media/files/VOCtest_06-Nov-2007.tar
tar -xf VOCtrainval_11-May-2012.tar
tar -xf VOCtrainval_06-Nov-2007.tar
tar -xf VOCtest_06-Nov-2007.tar

We have provided label lists in the folder data_paths which can be directly adapted to prepare data for training and evaluation.

python update_voc_path.py "/path/to/VOC/dataset/"

COCO Dataset

COCO dataset needs to be downloaded directly from the official website : https://cocodataset.org/#download

Download the 2017 Train Images, 2017 Val Images and 2017 Train/Val annotations.

We have provided label lists in the folder data_paths which can be directly adapted to prepare data for training and evaluation.

python update_coco_path.py "/path/to/COCO/dataset/"

Training a Model

Step 1 : Train with frozen transfer learning weights

python main.py --mode=TRAIN --train_dataset=<train list> --val_dataset=<val list> --freeze --classes_path=<class names list> --backbone=<MOBILENETV2x75|MOBILENETV2x14|EFFICIENTNETB3>

Step 2 : Fine tune the complete model

python main.py --mode=TRAIN --train_dataset=<train list> --val_dataset=<val list> --train_unfreeze=<last checkpoint> --classes_path=<class names list> --backbone=<MOBILENETV2x75|MOBILENETV2x14|EFFICIENTNETB3>

Example,

python main.py --mode=TRAIN --train_dataset=voc_train_14910.txt --val_dataset=voc_val_1641.txt --freeze --classes_path=model_data/voc_classes.txt --backbone=EFFICIENTNETB3
python main.py --mode=TRAIN --train_dataset=voc_train_14910.txt --val_dataset=voc_val_1641.txt --train_unfreeze=<last checkpoint> --classes_path=model_data/voc_classes.txt --backbone=EFFICIENTNETB3

Evaluating a Model:

A trained detection model can be evaluated using the following code,

python main.py --mode=MAP --model=<final checkpoint> --test_dataset=<test list> --classes_path=<class names list> --backbone=<MOBILENETV2x75|MOBILENETV2x14|EFFICIENTNETB3>

We have provided a few model checkpoints in the repo for testing. They can be used directly as follows

MobileNetV2x0.75 (320x320) on VOC Dataset

python main.py --mode=MAP --model=checkpoints/mobilenetv2x75_320_voc.h5  --test_dataset=voc_test_4952.txt --backbone=MOBILENETV2x75 --classes_path=model_data/voc_classes.txt --input_size=320 --input_size=320

EfficientNet-B3 (416x416) on COCO Dataset

python main.py --mode=MAP --model=checkpoints/efficientnetb3_416_coco.h5  --test_dataset=coco_test_4952.txt --backbone=EFFICIENTNETB3 --classes_path=model_data/coco_classes.txt --input_size=416 --input_size=416

EfficientNet-B3 (224x224) on COCO Dataset

python main.py --mode=MAP --model=checkpoints/efficientnetb3_224_coco.h5  --test_dataset=coco_test_4952.txt --backbone=EFFICIENTNETB3 --classes_path=model_data/coco_classes.txt --input_size=224 --input_size=224

Acknowledgement

This repository is adapted from https://github.com/fsx950223/mobilenetv2-yolov3