delira-dev / vision_base Goto Github PK

Include Resnets (18, 34, 50, 101, 152)

Train all classification Networks implemented so far on

• CIFAR10
• CIFAR100
• ImageWoof

Use dice score as metric (usually for segmentation).

https://www.tensorflow.org/datasets/datasets

Should include Backend-Agnostic implementations of the Following Datasets:

• MNIST
• KMNIST
• FashionMNIST
• EMNIST
• FakeData
• CIFAR10
• CIFAR100
• Imagenette
• ImageWoof
• ImageNet
• CityScapes
• Coco
• PascalVOC
• LSUN
• SVHN
• STL10
• ImageFolder
• MedicalImageFolder

And corresponding tests by producing (small) fake data following the structure of the original dataset

Missing Tests:

• MNIST
• KMNIST
• FashionMNIST
• EMNIST
• FakeData
• CIFAR10
• CIFAR100
• Imagenette
• ImageWoof
• ImageNet
• CityScapes
• Coco
• PascalVOC
• LSUN
• SVHN
• STL10
• ImageFolder
• MedicalImageFolder

So far the GANs are implemented only with basic closures, but we should provide the possibility to combine them with APEX

https://github.com/fastai/imagenette

• CheXpert
• LIDC
• LUNG1
• DDSM
• BRATS
• KITS

Add VGG Networks (11, 13, 16, 19)

DenseNet (101, 121, 161, 169, 201)

v1.0 and v1.1

We need a general class for Vision-Classifiers (maybe even in base-delira) which takes a model and implements a base-closure. The create_... functions should return an instance of this classifier.

remove imageops in delira once delira-vision is public