torchelie.models¶

We provide trained models. Use the argument pretrained=task in order to use them. Example: torchelie.models.resnet18(1000, pretrained='classification/imagenet').

Alternatively: torchelie.models.get_model('resnet18', 1000, pretrained='classification/imagenet')

Pretrained models¶
model	task	notes	source
resnet18	classification/imagenet	top1: 69.75%	torchvision
resnet34	classification/imagenet	top1: 73.31%	torchvision
resnet50	classification/imagenet	top1: 76.13%	torchvision
resnet101	classification/imagenet	top1: 77.37%	torchvision
resnet152	classification/imagenet	top1: 78.31%	torchvision
preact_resnet18	classification/imagenet	top1: 68.41% (192x192 crop)	torchelie
vgg11	classification/imagenet	top1: 69.02%	torchvision
vgg13	classification/imagenet	top1: 69.92%	torchvision
vgg16	classification/imagenet	top1: 71.59%	torchvision
vgg19	classification/imagenet	top1: 72.37%	torchvision
vgg11_bn	classification/imagenet	top1: 70.37%	torchvision
vgg13_bn	classification/imagenet	top1: 71.58%	torchvision
vgg16_bn	classification/imagenet	top1: 73.36%	torchvision
vgg19_bn	classification/imagenet	top1: 74.21%	torchvision
vgg19	perceptual/imagenet	activations normalized to have mean 1 for perceptual losses	torchelie

VGG¶

`VGG`	Construct a VGG-like model.
`vgg11`
`vgg13`
`vgg16`
`vgg19`
`vgg11_bn`
`vgg13_bn`
`vgg16_bn`
`vgg19_bn`

Pix2Pix¶

`Pix2PixGenerator`	UNet generator from Pix2Pix.
`pix2pix_256`	The architecture used in Pix2Pix, able to train on 256x256 or 512x512 images.
`pix2pix_dev`	A version of pix2pix_256 with less filter to use less memory and compute.
`PatchDiscriminator`
`patch16`	Patch Discriminator from pix2pix
`patch34`	Patch Discriminator from pix2pix
`patch70`	Patch Discriminator from pix2pix
`patch286`	Patch Discriminator from pix2pix
`residual_patch34`
`residual_patch70`
`residual_patch142`
`residual_patch286`

Pix2PixHD¶

Pix2PixHDGlobalGenerator

Residual generator used in Pix2PixHD

StyleGAN2¶

`StyleGAN2Generator`	Generator from StyleGAN2
`StyleGAN2Discriminator`	Experimental: Build the discriminator for StyleGAN2

Other GANs¶

`AutoGAN`	Experimental: Generator discovered in AutoGAN: Neural Architecture Search for Generative Adversarial Networks.
`autogan_32`	Experimental
`autogan_64`	Experimental
`autogan_128`	Experimental
`PerceptualNet`	Make a VGG16 with appropriately named layers that records intermediate activations.
`ResidualDiscriminator`
`res_discr_3l`
`res_discr_4l`
`res_discr_5l`
`res_discr_6l`
`res_discr_7l`
`snres_discr_4l`
`snres_discr_5l`
`snres_discr_6l`
`snres_discr_7l`
`snres_projdiscr_4l`
`snres_projdiscr_5l`
`snres_projdiscr_6l`
`snres_projdiscr_7l`

Image classifiers¶

`Attention56Bone`	Attention56 bone
`attention56`	Experimental: Build a attention56 network
`EfficientNet`
`Hourglass`	Experimental: Hourglass model from Deep Image Prior. .. warning:: Hourglass() is experimental, and may change or be deleted soon if not already broken.

Image Transformer¶

`UNet`	Experimental: U-Net from U-Net: Convolutional Networks for Biomedical Image Segmentation.
`Hourglass`	Experimental: Hourglass model from Deep Image Prior. .. warning:: Hourglass() is experimental, and may change or be deleted soon if not already broken.

Classification heads¶

`ClassificationHead`	A one layer classification head, turning activations / features into class log probabilities.
`ProjectionDiscr`	Experimental: A classification head for conditional GANs discriminators using a projection discriminator .

PixelCNN¶

PixelCNN

A PixelCNN model with 6 blocks

`ResNet`
`ResNetInput`
`preact_resnet101`
`preact_resnet152`
`preact_resnet18`
`preact_resnet20_cifar`
`preact_resnet34`
`preact_resnet50`
`preact_resnext101_32x4d`
`preact_resnext152_32x4d`
`preact_resnext50_32x4d`
`preact_wide_resnet101`
`preact_wide_resnet50`
`resnet101`
`resnet152`
`resnet18`
`resnet20_cifar`
`resnext101_32x4d`
`resnext152_32x4d`
`wide_resnet101`
`wide_resnet50`