PyTorch Deep Learning Toolbox

PyDLT is a set of tools aimed to make experimenting with PyTorch easier (than it already is).

Full Documentation here..

Features

• Trainers (currently supporting Vanilla, VanillaGAN, WGAN-GP, BEGAN, FisherGAN)

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.

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.

$ python main.py @settings.cfg
Some Settings:  config_test 32 0.0001

• HDR imaging support (.hdr, .exr, and .pfm formats)

img = dlt.hdr.imread('test.pfm')
dlt.hdr.imwrite('test.exr', img)

• Checkpointing of (torch serializable) objects; Network state dicts supported.

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)

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)

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.

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.

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).

$ dlt-plot --file losses.csv train_loss val_loss --refresh 5 --loglog True --tail 100

• A minimal Progress bar (with global on/off switch).

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 installed. OpenCV is also required:

conda install -c menpo opencv

conda install (recommended):

conda install -c demetris pydlt

From source:

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]