davidtvs / PyTorch-ENet

PyTorch implementation of ENet
MIT License
389 stars 129 forks source link
camvid cityscape enet pytorch

PyTorch-ENet

PyTorch (v1.1.0) implementation of ENet: A Deep Neural Network Architecture for Real-Time Semantic Segmentation, ported from the lua-torch implementation ENet-training created by the authors.

This implementation has been tested on the CamVid and Cityscapes datasets. Currently, a pre-trained version of the model trained in CamVid and Cityscapes is available here.

Dataset Classes 1 Input resolution Batch size Epochs Mean IoU (%) GPU memory (GiB) Training time (hours)2
CamVid 11 480x360 10 300 52.13 4.2 1
Cityscapes 19 1024x512 4 300 59.54 5.4 20

1 When referring to the number of classes, the void/unlabeled class is always excluded.
2 These are just for reference. Implementation, datasets, and hardware changes can lead to very different results. Reference hardware: Nvidia GTX 1070 and an AMD Ryzen 5 3600 3.6GHz. You can also train for 100 epochs or so and get similar mean IoU (± 2%).
3 Test set.
4 Validation set.

Installation

Local pip

  1. Python 3 and pip
  2. Set up a virtual environment (optional, but recommended)
  3. Install dependencies using pip: pip install -r requirements.txt

Docker image

  1. Build the image: docker build -t enet .
  2. Run: docker run -it --gpus all --ipc host enet

Usage

Run main.py, the main script file used for training and/or testing the model. The following options are supported:

python main.py [-h] [--mode {train,test,full}] [--resume]
               [--batch-size BATCH_SIZE] [--epochs EPOCHS]
               [--learning-rate LEARNING_RATE] [--lr-decay LR_DECAY]
               [--lr-decay-epochs LR_DECAY_EPOCHS]
               [--weight-decay WEIGHT_DECAY] [--dataset {camvid,cityscapes}]
               [--dataset-dir DATASET_DIR] [--height HEIGHT] [--width WIDTH]
               [--weighing {enet,mfb,none}] [--with-unlabeled]
               [--workers WORKERS] [--print-step] [--imshow-batch]
               [--device DEVICE] [--name NAME] [--save-dir SAVE_DIR]

For help on the optional arguments run: python main.py -h

Examples: Training

python main.py -m train --save-dir save/folder/ --name model_name --dataset name --dataset-dir path/root_directory/

Examples: Resuming training

python main.py -m train --resume True --save-dir save/folder/ --name model_name --dataset name --dataset-dir path/root_directory/

Examples: Testing

python main.py -m test --save-dir save/folder/ --name model_name --dataset name --dataset-dir path/root_directory/

Project structure

Folders

Files