See our collection for all versions of ResNetV2 (BiT).

Run ResNetV2 (BiT) with Keras 3: JAX, PyTorch, or TensorFlow

zeromodels/resnetv2_152x2_bit_goog_in21k

Paper: Big Transfer (BiT): General Visual Representation Learning (arXiv:1912.11370) · HF Papers

ResNetV2 / BiT checkpoints use pre-activation ResNet trained at large scale (often ImageNet-21k). Same ImageClassify / Model split as other classification backbones.

For more details on the model, please go to the upstream model card.

Pure-Keras 3 conversion of timm/resnetv2_152x2_bit.goog_in21k for zeromodels. One implementation runs unmodified on TensorFlow / Torch / JAX.

This is an image-classification / backbone checkpoint (ResNetV2ImageClassify / ResNetV2Model).

✨ Quick start

import os

os.environ["KERAS_BACKEND"] = "torch"  # or "jax" / "tensorflow"

from PIL import Image
from zeromodels.models.resnetv2 import ResNetV2ImageClassify, ResNetV2Model, ResNetV2ImageProcessor

model = ResNetV2ImageClassify.from_weights("zeromodels/resnetv2_152x2_bit_goog_in21k")
processor = ResNetV2ImageProcessor.from_weights("zeromodels/resnetv2_152x2_bit_goog_in21k")

image = Image.open("your_image.jpg").convert("RGB")
pixels = processor(image)  # resize + normalize (normalization lives in the processor)
logits = model(pixels, training=False)
print(logits.shape)  # (1, num_classes)

# Feature extraction: the backbone without the classifier head
backbone = ResNetV2Model.from_weights("zeromodels/resnetv2_152x2_bit_goog_in21k", as_backbone=True)
features = backbone(pixels, training=False)

Load any ResNetV2 (BiT) variant the same way with from_weights("zeromodels/<variant>"):

Variant	Hub
`resnetv2_101x1_bit_goog_in21k`	`zeromodels/resnetv2_101x1_bit_goog_in21k`
`resnetv2_101x1_bit_goog_in21k_ft_in1k`	`zeromodels/resnetv2_101x1_bit_goog_in21k_ft_in1k`
`resnetv2_101x3_bit_goog_in21k`	`zeromodels/resnetv2_101x3_bit_goog_in21k`
`resnetv2_101x3_bit_goog_in21k_ft_in1k`	`zeromodels/resnetv2_101x3_bit_goog_in21k_ft_in1k`
`resnetv2_152x2_bit_goog_in21k`	`zeromodels/resnetv2_152x2_bit_goog_in21k`
`resnetv2_152x2_bit_goog_in21k_ft_in1k`	`zeromodels/resnetv2_152x2_bit_goog_in21k_ft_in1k`
`resnetv2_152x4_bit_goog_in21k`	`zeromodels/resnetv2_152x4_bit_goog_in21k`
`resnetv2_152x4_bit_goog_in21k_ft_in1k`	`zeromodels/resnetv2_152x4_bit_goog_in21k_ft_in1k`
`resnetv2_50x1_bit_goog_in21k`	`zeromodels/resnetv2_50x1_bit_goog_in21k`
`resnetv2_50x1_bit_goog_in21k_ft_in1k`	`zeromodels/resnetv2_50x1_bit_goog_in21k_ft_in1k`
`resnetv2_50x3_bit_goog_in21k`	`zeromodels/resnetv2_50x3_bit_goog_in21k`
`resnetv2_50x3_bit_goog_in21k_ft_in1k`	`zeromodels/resnetv2_50x3_bit_goog_in21k_ft_in1k`

Tips

Set KERAS_BACKEND before importing Keras / zeromodels.
ResNetV2ImageClassify returns class logits; ResNetV2Model returns features (as_backbone=True for multi-scale stages).
See docs and Loading Weights.
Upstream / timm checkpoints: ResNetV2ImageClassify.from_weights("hf:timm/resnetv2_152x2_bit.goog_in21k").