Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/dmarnerides/pydlt
PyTorch based Deep Learning Toolbox
https://github.com/dmarnerides/pydlt
computer-vision deep-learning pytorch toolbox
Last synced: about 2 months ago
JSON representation
PyTorch based Deep Learning Toolbox
- Host: GitHub
- URL: https://github.com/dmarnerides/pydlt
- Owner: dmarnerides
- License: bsd-3-clause-clear
- Created: 2018-01-28T01:16:03.000Z (almost 7 years ago)
- Default Branch: master
- Last Pushed: 2018-07-27T11:39:46.000Z (over 6 years ago)
- Last Synced: 2024-08-04T00:13:11.274Z (5 months ago)
- Topics: computer-vision, deep-learning, pytorch, toolbox
- Language: Python
- Size: 1.96 MB
- Stars: 204
- Watchers: 9
- Forks: 15
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
- Awesome-pytorch-list - pydlt
- Awesome-pytorch-list-CNVersion - pydlt
README
[](http://pydlt.readthedocs.io/en/latest/?badge=latest)
[](https://travis-ci.org/dmarnerides/pydlt)
# PyTorch Deep Learning Toolbox
PyDLT is a set of tools aimed to make experimenting with [PyTorch](http://pytorch.org/)
easier (than it already is).
[Full Documentation here.](http://pydlt.readthedocs.io/).
## Features
- **Trainers** (currently supporting *Vanilla*, *VanillaGAN*, *WGAN-GP*, *BEGAN*, *FisherGAN*)
```python
trainer = dlt.train.VanillaGANTrainer(generator, discriminator, g_optim, d_optim)
for batch, (prediction, losses) in trainer(data_loader):
# Training happens in the iterator and relevant results are returned for each step
```
- Built in configurable **parser** with arguments.
```python
opt = dlt.config.parse() # Has built in options (can add extra)
print('Some Settings: ', opt.experiment_name, opt.batch_size, opt.lr)
```
- **Configuration files** support and parser compatible functions.
```bash
$ python main.py @settings.cfg
Some Settings: config_test 32 0.0001
```
- **HDR imaging** support (.hdr, .exr, and .pfm formats)
```python
img = dlt.hdr.imread('test.pfm')
dlt.hdr.imwrite('test.exr', img)
```
- **Checkpointing** of (torch serializable) objects; Network state dicts supported.
```python
data_chkp = Checkpointer('data')
data_chkp.save(np.array([1,2,3]))
a = data_chkp.load()
```
- **Image operations** and easy conversions between multiple library views (torch, cv, plt)
```python
img = cv2.imread('image.jpg') # Height x Width x Channels - BGR
dlt.viz.imshow(img, view='cv') # Height x Width x Channels - RGB
tensor_with_torch_view = cv2torch(img) # Channels x Height x Width - RGB
```
- Easy **visualization** (and make_grid supporting Arrays, Tensors, Variables and lists)
```python
for batch, (prediction, loss) in trainer(loader):
grid = dlt.util.make_grid([ batch[0], batch[1], prediction], size(3, opt.batch_size))
dlt.viz.imshow(grid, pause=0.01, title='Training Progress')
```
- Parameter and input/outputs/gradients **layer visualization**.
```python
net = nn.Sequential(nn.Linear(10, 10))
dlt.viz.modules.forward_hook(net, [nn.Linear], tag='layer_outputs', histogram=False)
net(Variable(torch.Tensor(3,10)))
```
- CSV **Logger**.
```python
log = dlt.util.Logger('losses', ['train_loss', 'val_loss'])
log({'train_loss': 10, 'val_loss':20})
```
- Command line tool for easy **plotting** of CSV files (with live updating).
```bash
$ dlt-plot --file losses.csv train_loss val_loss --refresh 5 --loglog True --tail 100
```
- A minimal **Progress bar** (with global on/off switch).
```python
from dlt.util import barit
barit.silent = False # Default is False
for batch in barit(loader, start='Loading'):
pass
```
## Installation
Make sure you have [PyTorch](http://pytorch.org/) installed. OpenCV is also required:
```bash
conda install -c menpo opencv
```
### conda install (recommended):
```bash
conda install -c demetris pydlt
```
### From source:
```bash
git clone https://github.com/dmarnerides/pydlt.git
cd pydlt
python setup.py install
```
## About
I created this toolbox while learning Python and PyTorch, after working with
(Lua) Torch, to help speed up experiment prototyping.
If you notice something is wrong or missing please do a pull request or
open up an issue.
## Contact
Demetris Marnerides: [email protected]