https://github.com/enhuiz/segnn
https://github.com/enhuiz/segnn
Last synced: 4 months ago
JSON representation
- Host: GitHub
- URL: https://github.com/enhuiz/segnn
- Owner: enhuiz
- Created: 2019-03-28T16:40:29.000Z (about 6 years ago)
- Default Branch: master
- Last Pushed: 2019-04-09T13:04:43.000Z (about 6 years ago)
- Last Synced: 2025-01-07T18:28:38.311Z (5 months ago)
- Language: Python
- Size: 4.26 MB
- Stars: 0
- Watchers: 3
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# SegNN
This is the code for comp5421 program assignment 1.
## Install Dependencies
```bash
$ pip3 install -r requirements.txt
```
## Run Inference
Run the following commands under the project directory.
```bash
./eval_best_model.sh
```
## Run Training
```bash
./run_dilated_resnet18_upernet.sh
```
## Create Your Own Model
Steps to add a model (let's call it MyNN):
1. Create you `MyNNEncoder` in `segnn/models/encoder.py` and `MyNNDecoder` in `segnn/models/decoder.py`.
2. Change `scripts/nn_make.py`, create you way to initialize the model.
3. Create `run_my_nn.sh`.
Then run the `.sh`, you result will be automatically output to the exp/ and stdout.
## Directories
```bash
.
├── data
│ └── comp5421_TASK2 <- data folder
├── exp <- experiment result, output of the scripts
│ └── mynn
│ ├── test <- predicted labels for test
│ ├── ckpt
│ │ ├── ckpt <- pytorch .pth checkpoints, will be used to predict test/val
│ │ └── config.json <- backup training configuration in json format
│ │ └── iou.txt <- iou of val/
│ └── val <- predicted labels for val
├── requirements.txt
├── run_my_nn.sh <- executor, unix-based system required
├── scripts
│ ├── nn_forward.py <- create labels
│ ├── nn_make.py <- make proto .pth model to zoo/, so that nn_train.py can take it and train
│ ├── nn_train.py <- train .pth
│ ├── parse_options.sh <- shell utils to parse options, like argparse in python
│ ├── run.sh <- run train, feedforward and evaluation
│ └── visualize_mask.sh <- visualize results
├── segnn
│ ├── data.py <- data set
│ ├── __init__.py
│ ├── models
│ │ ├── decoder.py <- ppm, upernet etc.
│ │ └── encoder.py <- resnet etc.
│ │ └── __init__.py
│ ├── transforms.py <- overload pytorch transforms
│ └── utils.py <- where you add your utils
└── zoo <- .pth created by nn_make.py
└── my_nn.pth
```
## Result
### Validation
#### Example
```bash
$ python3 scripts/visualize_mask.py data/comp5421_TASK2/val/images exp/dilated_resnet18_upernet_refine/val visual/refine/val
```
where
```json
{
"black": "unknown",
"blue": "water",
"white": "sky",
"grey": "road",
"orange": "building",
"green": "vegetation",
"brown: "ground"
}
```
#### mIoU
```python
{'meanIoU': 0.5122276750329822, 'IoU_array': array([0.24294965, 0.87237296, 0. , 0.53663633, 0.83179963, 0.71088077, 0.39095439])}
```
### Test
#### Example
```bash
$ python3 scripts/visualize_mask.py data/comp5421_TASK2/test/images exp/dilated_resnet18_upernet_refine/test visual/refine/test
```
