QuAR checkpoints

Quantized checkpoints for QuAR, the code for Test-Time Adaptation of Quantized ViTs via Single-Pass Quantizer-Aligned Recalibration.

23 files: PTQ4ViT and AdaLog quantizations of ViT-B/16, DeiT-B, ViT-S and ResNet-50 from W3A3 to W8A8, on ImageNet, CIFAR-10/100 and DomainNet-126, plus the full-precision DomainNet-126 model the last of those are built from.

Use

git clone https://github.com/chahh9808/QuAR && cd QuAR
bash tools/fetch_artifacts.sh

The script puts them in checkpoints/ and checks each one against the SHA-256 in artifacts/MANIFEST.tsv, which also records what produced it. AdaLog's own ImageNet checkpoints and the two source CIFAR ViTs are not mirrored here; the script prints where to get them.

Provenance

Every file is a quantization, that is a modification, of an Apache-2.0 model:

Source From
ViT-B/16, ViT-S timm augreg, google/vit-base-patch16-224-in21k
DeiT-B facebook/deit-base-patch16-224
CIFAR-10 ViT-B nateraw/vit-base-patch16-224-cifar10
CIFAR-100 ViT-B pkr7098/cifar100-vit-base-patch16-224-in21k
ResNet-50 (BatchNorm, GroupNorm) torchvision and timm weights
DomainNet-126 ViT-B/16 a timm ViT-B/16 fine-tuned on the real domain

Quantized with PTQ4ViT, AdaLog and MQBench.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support