https://github.com/zeroact/vae-study-project
vae study project
https://github.com/zeroact/vae-study-project
Last synced: 2 months ago
JSON representation
vae study project
- Host: GitHub
- URL: https://github.com/zeroact/vae-study-project
- Owner: ZeroAct
- Created: 2021-01-02T15:16:11.000Z (over 4 years ago)
- Default Branch: master
- Last Pushed: 2021-01-21T08:22:08.000Z (over 4 years ago)
- Last Synced: 2025-02-01T21:33:22.656Z (4 months ago)
- Language: Python
- Size: 177 KB
- Stars: 3
- Watchers: 2
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
## Pytorch VAE
#### My Vae Study Project
This repo contains a simple implementation of VAE (Variational AutoEncoder) in pytorch. I refered this repo ([AntixK/PyTorch-VAE: A Collection of Variational Autoencoders (VAE) in PyTorch. (github.com)](https://github.com/AntixK/PyTorch-VAE)) vanilla_vae.py and other blogs. Thanks for everyone who posts nice codes.
### Requirements
- python 3.7
- pytorch >= 1.3
- opencv-python
- others in requirements.txt ...
### Installation
- git clone https://github.com/ZeroAct/vae-study-project.git
- cd vae-study-project
- pip install -r requirements.txt
### Usage
#### 1. Make .ini file
In `conifgs` directory you will see `default.ini` file like this.
```python
[model]
input_size = [256, 256]
in_channel = 3
latent_dim = 100
hidden_dims = [64, 64, 128, 128, 256, 256]
```
This section determines the model.
```python
[hyperparameters]
epochs = 5
batch_size = 2
optimizer = adam
learning_rate = 1e-2
loss = L1
```
You can change hyperparameters here.
`adam` is the only supported optimizer for now.
You can use `L1`, `MSE`, `SSIM` loss functions. If you want to use other loss functions, you can implement them in `vae_pytorch/loss.py`.
```python
[data]
train_img_path = "datasets/train"
val_img_path = "datasets/val"
```
Set the image path. It will automatically get images recursively.
#### 2. Train
After making `.ini` file, you train the model by script :
```python
python train.py -c configs/my.ini # your config file (require)
-v 1 # validation each 1 epoch (default)
-s 1 # save model each 1 epoch (default)
-n 8 # num workers for DataLoader (default)
-o True # show reconstruction result while training (default)
```
The results(reconstruction, weights) will be saved under `logs/{.ini file name}/`.
#### 3. Predict
After training model, you can see the encoded latent vectors and reconstructed images.
The results will be saved under `results/{.ini file name}/`.
### Results
I used cifar-10 datasets. I trained 3 models with different loss function. Each model trained for 50 epochs.
| L1 Loss | MSE Loss | SSIM (Filter Size 3) |
| :-------------------------: | :-----------------------: | :------------------------: |
| 
| 
| 
|
It seems like L1 loss is better than MSE loss. Using SSIM loss was the best at reconstructing the details, but worst at reconstructing color information. (I don't know why...)
### Time Line
#### 202101-03 : initial commit
#### 2021-01-03 : Loss Class / Add Predict Code / Remove Some Errors
#### 2021-01-14 : Modify Loss functions / Remove trash files
#### 2021-01-15 : bug fix / update readme.md