import torch
import open_clip
from PIL import Image
from transformers import AutoModel, CLIPImageProcessor
device = "cuda" if torch.cuda.is_available() else "cpu"

# obtained 69.1% zero-shot ImageNet score
model = AutoModel.from_pretrained(
    'jienengchen/ViTamin-B',
    trust_remote_code=True).to(device).eval()

image = Image.open('./image.png').convert('RGB')
image_processor = CLIPImageProcessor.from_pretrained('jienengchen/ViTamin-B')

pixel_values = image_processor(images=image, return_tensors='pt').pixel_values
pixel_values = pixel_values.to(torch.bfloat16).cuda()

tokenizer = open_clip.get_tokenizer('hf-hub:laion/CLIP-ViT-L-14-DataComp.XL-s13B-b90K')
text = tokenizer(["a photo of vitamin", "a dog", "a cat"]).to(device)

with torch.no_grad(), torch.cuda.amp.autocast():
    image_features, text_features, logit_scale = model(pixel_values, text)
    text_probs = (100.0 * image_features @ text_features.to(torch.float).T).softmax(dim=-1)

print("Label probs:", text_probs)

Main Results with CLIP Pre-training on DataComp-1B

image encoder	image size	num patches	text encoder depth/width	seen samples (B)	trainable params Image+Text (M)	MACs Image+Text (G)	ImageNet Acc.	avg. 38 datasets	ImageNet dist. shift.	VTAB	retrieval
ViTamin-S	224	196	12/384	1.28	22.0+40.4	5.50+1.64	62.2	53.2	51.3	51.7	50.0
ViTamin-S-LTT	224	196	12/384	1.28	22.0+40.4	5.50+1.64	63.4	54.6	51.6	54.9	52.9
ViTamin-B	224	196	12/512	1.28	87.5+63.4	21.8+2.9	68.9	57.7	58.3	56.4	54.1
ViTamin-B-LTT	224	196	12/512	1.28	87.5+63.4	21.8+2.9	70.8	59.4	59.3	56.6	59.4
ViTamin-L	224	196	12/768	12.8	333.3+123.7	72.6+6.6	80.8	66.7	69.8	65.3	60.3
ViTamin-L	256	256	12/768	12.8+0.2	333.4+123.7	94.8+6.6	81.2	67.0	71.1	65.3	61.2
ViTamin-L	336	441	12/768	12.8+0.2	333.6+123.7	163.4+6.6	81.6	67.0	72.1	64.4	61.6
ViTamin-L	384	576	12/768	12.8+0.2	333.7+123.7	213.4+6.6	81.8	67.2	72.4	64.7	61.8
ViTamin-L2	224	196	24/1024	12.8	333.6+354.0	72.6+23.3	80.9	66.4	70.6	63.4	61.5
ViTamin-L2	256	256	24/1024	12.8+0.5	333.6+354.0	94.8+23.3	81.5	67.4	71.9	64.1	63.1
ViTamin-L2	336	441	24/1024	12.8+0.5	333.8+354.0	163.4+23.3	81.8	67.8	73.0	64.5	63.6
ViTamin-L2	384	576	24/1024	12.8+0.5	334.0+354.0	213.4+23.3	82.1	68.1	73.4	64.8	63.7
ViTamin-XL	256	256	27/1152	12.8+0.5	436.1+488.7	125.3+33.1	82.1	67.6	72.3	65.4	62.7
ViTamin-XL	384	576	27/1152	12.8+0.5	436.1+488.7	281.9+33.1	82.6	68.1	73.6	65.6	63.8
ViTamin-XL	256	256	27/1152	40	436.1+488.7	125.3+33.1	82.3	67.5	72.8	64.0	62.1
ViTamin-XL	336	441	27/1152	40+1	436.1+488.7	215.9+33.1	82.7	68.0	73.9	64.1	62.6
ViTamin-XL	384	576	27/1152	40+1	436.1+488.7	281.9+33.1	82.9	68.1	74.1	64.0	62.5

Main Results on Downstream tasks

Open-Vocab Detection

image encoder	detector	OV-COCO (AP₅₀^novel)	OV-LVIS (AP_r)
ViT-L/14	Sliding F-ViT	36.1	32.5
ViTamin-L	Sliding F-ViT	37.5	35.6

Open-Vocab Segmentation

image encoder	segmentor	ADE	Cityscapes	MV	A-150	A-847	PC-459	PC-59	PAS-21
ViT-L/14	Sliding FC-CLIP	24.6	40.7	16.5	31.8	14.3	18.3	55.1	81.5
ViTamin-L	Sliding FC-CLIP	27.3	44.0	18.2	35.6	16.1	20.4	58.4	83.4

Note: Panoptic dataset (ADE, CityScapes, MV) are with the metric of PQ. Semantic dataset (A-150, A-847, PC-459, PC-59, PAS-21) are with the metric of mIoU.

Large Multi-modal Models

image encoder	image size	VQAv2	GQA	VizWiz	SQA	T-VQA	POPE	MME	MM-Bench	MM-B-CN	SEED	LLaVA-Wild	MM-Vet
ViTamin-L	336	78.4	61.6	51.1	66.9	58.7	84.6	1421	65.4	58.4	57.7	64.5	33.6
ViTamin-L	384	78.9	61.6	55.4	67.6	59.8	85.5	1447	64.5	58.3	57.9	66.1	33.6

Citing ViTamin

@inproceedings{chen2024vitamin,
  title={ViTamin: Design Scalable Vision Models in the Vision-language Era},
  author={Chen, Jieneng and Yu, Qihang and Shen, Xiaohui and Yuille, ALan and Chen, Liang-Chieh},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year={2024}
}

jienengchen/ViTamin-S-LTT

Model card for ViTamin-S-LTT

Main Results with CLIP Pre-training on DataComp-1B

Main Results on Downstream tasks

Citing ViTamin

jienengchen/
ViTamin-S-LTT