from homura import Registry
SEMI_DATASET_REGISTRY = Registry('semi_dataset')
while True:
from .datasets import *
break
from homura import Registry, get_environ
from .classification import ExtraSVHN, ImageNet, OriginalSVHN
from .detection import VOCDetection, det_collate_fn
from .segmentation import ExtendedVOCSegmentation, seg_collate_fn
from .visionset import VisionSet
DATASET_REGISTRY = Registry('vision_datasets', type=VisionSet)
from torchvision import datasets, transforms
from .. import transforms as homura_transforms
DATASET_REGISTRY.register_from_dict({
'cifar10':
VisionSet(datasets.CIFAR10, "~/.torch/data/cifar10", 10, [
transforms.ToTensor(),
transforms.Normalize((0.4914, 0.4822, 0.4465),
(0.2023, 0.1994, 0.2010))
], [
transforms.RandomCrop(32, padding=4, padding_mode='reflect'),
transforms.RandomHorizontalFlip()
]),
'cifar100':
VisionSet(datasets.CIFAR100, "~/.torch/data/cifar100", 100, [
transforms.ToTensor(),
transforms.Normalize((0.5071, 0.4867, 0.4408),
(0.2675, 0.2565, 0.2761))
], [
transforms.RandomCrop(32, padding=4, padding_mode='reflect'),
transforms.RandomHorizontalFlip()
]),
'svhn':
from homura import Registry
MODEL_REGISTRY = Registry('vision_model')
from torchvision import models
MODEL_REGISTRY.register(models.resnet18)
MODEL_REGISTRY.register(models.resnet50)
MODEL_REGISTRY.register(models.wide_resnet50_2)
MODEL_REGISTRY.register(models.densenet121)
MODEL_REGISTRY.register(models.vgg19_bn)
from .densenet import densenet40, densenet100
from .cifar_resnet import wrn28_2, wrn40_2, wrn28_10, resnet20, resnet56, resnext29_32x4d
from .unet import unet
from homura import Registry
SEMI_TRAINER_REGISTRY = Registry('semi_trainer')
while True:
from .utils import unroll
from .Ladder import Ladder
from .MeanTeacher import MeanTeacher
from .InterpolationConsistency import InterpolationConsistency
from .AdversariallyLearnedInference import AdversariallyLearnedInference
from .MixMatch import MixMatch
break
from homura import Registry
SCHEDULER_REGISTRY = Registry('scheduler')
while True:
from .scheduler import Linear
break
from homura import Registry
TRANSFORM_REGISTRY = Registry('transform')
while True:
from .transform_many_times import *
from .randaugment import RandAugment
break
from homura import Registry
MODEL_REGISTRY = Registry('vision_model')
from torchvision import models
MODEL_REGISTRY.register(models.resnet18)
MODEL_REGISTRY.register(models.resnet50)
MODEL_REGISTRY.register(models.resnet101)
MODEL_REGISTRY.register(models.resnet152)
MODEL_REGISTRY.register(models.wide_resnet50_2)
MODEL_REGISTRY.register(models.densenet121)
MODEL_REGISTRY.register(models.vgg19_bn)
if hasattr(models, "efficientnet_b0"):
# >=1.10
for v in range(8):
MODEL_REGISTRY.register(getattr(models, f"efficientnet_b{v}"))
from .densenet import densenet40, densenet100
from .cifar_resnet import wrn28_2, wrn40_2, wrn28_10, resnet20, resnet56, resnext29_32x4d
from .unet import unet
self.ln = nn.LayerNorm(emb_dim)
self.fc = nn.Linear(emb_dim, num_classes)
self._init_weights()
def _init_weights(self):
nn.init.zeros_(self.fc.weight)
def forward(self, input: torch.Tensor) -> torch.Tensor:
x = self.patch_emb(input) # BxNxC
x = self.blocks(x)
return self.fc(self.ln(x).mean(dim=1))
from homura import Registry
MLPMixers = Registry("MLPMixer")
@MLPMixers.register
def mixer_s32(num_classes, **kwargs) -> MLPMixer:
return MLPMixer(num_classes, 512, 256, 2048, 32, 8, **kwargs)
@MLPMixers.register
def mixer_s16(num_classes, **kwargs) -> MLPMixer:
return MLPMixer(num_classes, 512, 256, 2048, 16, 8, **kwargs)
@MLPMixers.register
def mixer_b32(num_classes, **kwargs) -> MLPMixer:
return MLPMixer(num_classes, 768, 384, 3072, 32, 12, **kwargs)