weaver-pytorch-rnx0dvmdxk 0.0.4

Small utilities for PyTorch

Weaver PyTorch 🧶🧵

from weaver import get_classifier, get_optimizer, get_scheduler, get_transforms
from torchvision.transforms import Compose
model = get_classifier('torchvision', 'resnet50')
optim = get_optimizer(model.parameters(), name='SGD', lr=1e-3)
sched = get_scheduler(optim, name='CosineAnnealingLR', T_max=10)
transform = Compose(get_transforms([
    {'name': 'RandAugment', 'num_ops': 2, 'magnitude': 10},
    {"name": "ToTensor"},
    {"name": "Normalize", "mean": "cifar10", "std": "cifar10"}
]))

Install

pip install weaver-pytorch-rnx0dvmdxk

---

API

get_classifier(src, name, **kwargs)

• weaver: wide_resnet{depth}_{width}, preact_resnet{depth}
• torchvision: https://pytorch.org/vision/stable/models.html

get_optimizer(params, name, **kwargs)

• PyTorch: https://pytorch.org/docs/stable/optim.html#algorithms
• AdaBelief: https://github.com/juntang-zhuang/Adabelief-Optimizer

get_scheduler(optim, name, **kwargs)

• PyTorch: https://pytorch.org/docs/stable/optim.html#how-to-adjust-learning-rate
• Custom: HalfCosineAnnealingLR

get_transform(name, **kwargs)

• PyTorch: https://pytorch.org/vision/stable/transforms.html
• Custom: AllRandAugment, Cutout, Contain

get_transforms(kwargs_list)

• get list of transforms

Others

• weaver.optimizers.exclude_wd(module: Module, skip_list=['bias', 'bn'])
• weaver.optimizers.EMAModel(model: Module, alpha: float)
• weaver.datasets.IndexedDataset
• weaver.datasets.RandomSubset