- pytorch
- torchvision
- pytorch lighting
- wandb
- numpy
- scipy
- tqdm
- PIL
- https://github.com/alibaba/TinyNeuralNetwork
The dataset used for this project is the Cityscapes Image Pairs dataset available at https://www.kaggle.com/datasets/dansbecker/cityscapes-image-pairs. After the the dataset is downloaded, simply extract the dataset and then place it in data
directory in the root directory of this repository.
In this project we train a semantic segmentation model
The model is of the UNet architecture. It is a combination of EfficientNetB0 as the encoder and FFNet as the decoder.
To train the model from scratch simply run:
python train.py
A trained model checkpoint is also provided - model.ckpt
Model quantization is performed and evaluated in compression.ipynb. The quantized model is also provided in the quant_output directory.