mxbox 0.0.3

image and video datasets and models for mxnet deep learning

MXbox: The missing toolbox for mxnet.

MXbox is a toolbox aiming to provide a general and simple interface for vision tasks. This project is greatly inspired by PyTorch and torchvision. Detailed copyright files will be attached later. Improvements and suggestions are welcome.

Installation

pip install mxbox

Features

1. Define preprocess as a flow

img_transform = transforms.Compose([
    transforms.RandomSizedCrop(224),
    transforms.RandomHorizontalFlip(),
    transforms.mx.ToNdArray(),
    transforms.mx.Normalize(mean = [ 0.485, 0.456, 0.406 ],
                            std  = [ 0.229, 0.224, 0.225 ]),
])

PS: By default, mxbox use PIL to read and transform images. But it also supports other backends like skimage_and Numpy.

More examples can be found in XXX.

2. Build DataLoader in several lines

feedin_shapes = {
    'batch_size': 8,
    'data': [mx.io.DataDesc(name='data', shape=(32, 3, 128, 128), layout='NCHW')],
    'label': [mx.io.DataDesc(name='softmax_label', shape=(32,), layout='N')]
}

dst = TestDataset(root='../../data', transform=img_transform, label_transform=label_transform)
loader = BoxLoader(dst, feedin_shapes, collate_fn=mx_collate, num_workers=1)

3. Load popular model and pretrained weights

vgg = mxbox.models.vgg(num_classes=10, pretrained=True)
resnet = mxbox.models.resnet50(num_classes=10, pretrained=True)

Documentation

Under construction, coming soon.

TODO list

1. Random shuffle

2. Efficient multi thread reading

3. Common Models