SSD300 Model in PyTorch

Team

Shiming Luo, Jonathan Lam, Nathan Apolonio, Eric Megrabov

Dataset

PASCAL VOC2012

Usage

We need to update torchvision to the latest version.

pip install --user --upgrade torch torchvision

Get a Demo

Since the trained weigt is too big (300MB), please download it from the google drive: https://drive.google.com/drive/folders/1vGomCtIbWX_ah9_FE-ciN5P-wqLJeJ2M?usp=sharing and save the entire demo directory under folder weight/

Jupyter Notebook

Open code/demo.ipynb and follow run the cells in it.

If you wanna do predictions on your own images, please change the variable img_path in code/demo.ipynb to your own.

On terminal

If you wanna do predictions on your own images, please change the variable img_path in code/demo.ipynb to your own.

cd code
python3 demo.py

The result images and ground truth images would be saved under directory called img/

Train the Model

Set the download=True for train_dataset and val_dataset for the first time training.

cd code
python3 train_ssd.py

Since the pre-trained VGG_16 weight is not uploaded (too big), this may call some error. But you could download it by following bash command.

cd weight
wget https://s3.amazonaws.com/amdegroot-models/vgg16_reducedfc.pth

This would make the program run properly.

Evaluate the Model

cd code
python3 evaluation_ssd.py

tip: we need to train the model at least 1 epoch to save a trained weights and then evaluate the trained model.

Result

Training Loss:

At around 12000 global step, the training loss increased rapidlly, so we change the learning rate from 1e-3 to 1e-4.

On training set:

mAP Matrix:

[0.91019469 0.84114809 0.88723437 0.67804573 0.37342244 0.91621779 0.73382705 0.97321562 0.80776674 0.84021185 0.9088941 0.96271324 0.92898749 0.95671725 0.66330431 0.69221699 0.78723723 0.95487637 0.95664115 0.89042396]

mAP = 0.8331648228692401

On validation set:

mAP Matrix:

[0.72079278 0.54627386 0.47994215 0.38346232 0.11459754 0.70997109 0.49702614 0.56876101 0.30975806 0.49377826 0.45675858 0.55071475 0.60932779 0.60878767 0.41897762 0.21055238 0.47257278 0.39755191 0.5924096 0.50312271]

mAP = 0.4822569492169384

The mAP on validation set may not as high as it on training set. It may because we did not do data augmentation.

To Do

1. Data Augmentation