convnext_xxlarge.clip_laion2b_soup_ft_in12k timm模型库

Model card for convnext_xxlarge.clip_laion2b_soup_ft_in12k

A ConvNeXt image classification model. CLIP image tower weights pretrained in OpenCLIP on LAION and fine-tuned on ImageNet-12k by Ross Wightman.

Please see related OpenCLIP model cards for more details on pretrain:

• https://huggingface.co/laion/CLIP-convnext_xxlarge-laion2B-s34B-b82K-augreg-soup
• https://huggingface.co/laion/CLIP-convnext_large_d.laion2B-s26B-b102K-augreg
• https://huggingface.co/laion/CLIP-convnext_base_w-laion2B-s13B-b82K-augreg
• https://huggingface.co/laion/CLIP-convnext_base_w_320-laion_aesthetic-s13B-b82K

Model Details

• Model Type: Image classification / feature backbone
• Model Stats:
  • Params (M): 879.7
  • GMACs: 198.1
  • Activations (M): 124.5
  • Image size: 256 x 256
• Papers:
  • OpenCLIP: https://github.com/mlfoundations/open_clip
  • LAION-5B: An open large-scale dataset for training next generation image-text models: https://arxiv.org/abs/2210.08402
  • @: m
  • Learning Transferable Visual Models From Natural Language Supervision: https://arxiv.org/abs/2103.00020
  • A ConvNet for the 2020s: https://arxiv.org/abs/2201.03545
• Original: https://github.com/mlfoundations/open_clip
• Pretrain Dataset: LAION-2B
• Dataset: ImageNet-12k

Model Usage

Image Classification

from urllib.request import urlopen
from PIL import Image
import timm

img = Image.open(urlopen(
    'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
))

model = timm.create_model('convnext_xxlarge.clip_laion2b_soup_ft_in12k', pretrained=True)
model = model.eval()

# get model specific transforms (normalization, resize)
data_config = timm.data.resolve_model_data_config(model)
transforms = timm.data.create_transform(**data_config, is_training=False)

output = model(transforms(img).unsqueeze(0))  # unsqueeze single image into batch of 1

top5_probabilities, top5_class_indices = torch.topk(output.softmax(dim=1) * 100, k=5)

Feature Map Extraction

from urllib.request import urlopen
from PIL import Image
import timm

img = Image.open(urlopen(
    'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
))

model = timm.create_model(
    'convnext_xxlarge.clip_laion2b_soup_ft_in12k',
    pretrained=True,
    features_only=True,
)
model = model.eval()

# get model specific transforms (normalization, resize)
data_config = timm.data.resolve_model_data_config(model)
transforms = timm.data.create_transform(**data_config, is_training=False)

output = model(transforms(img).unsqueeze(0))  # unsqueeze single image into batch of 1

for o in output:
    # print shape of each feature map in output
    # e.g.:
    #  torch.Size([1, 384, 64, 64])
    #  torch.Size([1, 768, 32, 32])
    #  torch.Size([1, 1536, 16, 16])
    #  torch.Size([1, 3072, 8, 8])

    print(o.shape)

Image Embeddings

from urllib.request import urlopen
from PIL import Image
import timm

img = Image.open(urlopen(
    'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/beignets-task-guide.png'
))

model = timm.create_model(
    'convnext_xxlarge.clip_laion2b_soup_ft_in12k',
    pretrained=True,
    num_classes=0,  # remove classifier nn.Linear
)
model = model.eval()

# get model specific transforms (normalization, resize)
data_config = timm.data.resolve_model_data_config(model)
transforms = timm.data.create_transform(**data_config, is_training=False)

output = model(transforms(img).unsqueeze(0))  # output is (batch_size, num_features) shaped tensor

# or equivalently (without needing to set num_classes=0)

output = model.forward_features(transforms(img).unsqueeze(0))
# output is unpooled, a (1, 3072, 8, 8) shaped tensor

output = model.forward_head(output, pre_logits=True)
# output is a (1, num_features) shaped tensor

Model Comparison

Explore the dataset and runtime metrics of this model in timm model results.

All timing numbers from eager model PyTorch 1.13 on RTX 3090 w/ AMP.

model	top1	top5	img_size	param_count	gmacs	macts	samples_per_sec	batch_size
convnextv2_huge.fcmae_ft_in22k_in1k_512	88.848	98.742	512	660.29	600.81	413.07	28.58	48
convnextv2_huge.fcmae_ft_in22k_in1k_384	88.668	98.738	384	660.29	337.96	232.35	50.56	64
convnext_xxlarge.clip_laion2b_soup_ft_in1k	88.612	98.704	256	846.47	198.09	124.45	122.45	256
convnext_large_mlp.clip_laion2b_soup_ft_in12k_in1k_384	88.312	98.578	384	200.13	101.11	126.74	196.84	256
convnextv2_large.fcmae_ft_in22k_in1k_384	88.196	98.532	384	197.96	101.1	126.74	128.94	128
convnext_large_mlp.clip_laion2b_soup_ft_in12k_in1k_320	87.968	98.47	320	200.13	70.21	88.02	283.42	256
convnext_xlarge.fb_in22k_ft_in1k_384	87.75	98.556	384	350.2	179.2	168.99	124.85	192
convnextv2_base.fcmae_ft_in22k_in1k_384	87.646	98.422	384	88.72	45.21	84.49	209.51	256
convnext_large.fb_in22k_ft_in1k_384	87.476	98.382	384	197.77	101.1	126.74	194.66	256
convnext_large_mlp.clip_laion2b_augreg_ft_in1k	87.344	98.218	256	200.13	44.94	56.33	438.08	256
convnextv2_large.fcmae_ft_in22k_in1k	87.26	98.248	224	197.96	34.4	43.13	376.84	256
convnext_base.clip_laion2b_augreg_ft_in12k_in1k_384	87.138	98.212	384	88.59	45.21	84.49	365.47	256
convnext_xlarge.fb_in22k_ft_in1k	87.002	98.208	224	350.2	60.98	57.5	368.01	256
convnext_base.fb_in22k_ft_in1k_384	86.796	98.264	384	88.59	45.21	84.49	366.54	256
convnextv2_base.fcmae_ft_in22k_in1k	86.74	98.022	224	88.72	15.38	28.75	624.23	256
convnext_large.fb_in22k_ft_in1k	86.636	98.028	224	197.77	34.4	43.13	581.43	256
convnext_base.clip_laiona_augreg_ft_in1k_384	86.504	97.97	384	88.59	45.21	84.49	368.14	256
convnext_base.clip_laion2b_augreg_ft_in12k_in1k	86.344	97.97	256	88.59	20.09	37.55	816.14	256
convnextv2_huge.fcmae_ft_in1k	86.256	97.75	224	660.29	115.0	79.07	154.72	256
convnext_small.in12k_ft_in1k_384	86.182	97.92	384	50.22	25.58	63.37	516.19	256
convnext_base.clip_laion2b_augreg_ft_in1k	86.154	97.68	256	88.59	20.09	37.55	819.86	256
convnext_base.fb_in22k_ft_in1k	85.822	97.866	224	88.59	15.38	28.75	1037.66	256
convnext_small.fb_in22k_ft_in1k_384	85.778	97.886	384	50.22	25.58	63.37	518.95	256
convnextv2_large.fcmae_ft_in1k	85.742	97.584	224	197.96	34.4	43.13	375.23	256
convnext_small.in12k_ft_in1k	85.174	97.506	224	50.22	8.71	21.56	1474.31	256
convnext_tiny.in12k_ft_in1k_384	85.118	97.608	384	28.59	13.14	39.48	856.76	256
convnextv2_tiny.fcmae_ft_in22k_in1k_384	85.112	97.63	384	28.64	13.14	39.48	491.32	256
convnextv2_base.fcmae_ft_in1k	84.874	97.09	224	88.72	15.38	28.75	625.33	256
convnext_small.fb_in22k_ft_in1k	84.562	97.394	224	50.22	8.71	21.56	1478.29	256
convnext_large.fb_in1k	84.282	96.892	224	197.77	34.4	43.13	584.28	256
convnext_tiny.in12k_ft_in1k	84.186	97.124	224	28.59	4.47	13.44	2433.7	256
convnext_tiny.fb_in22k_ft_in1k_384	84.084	97.14	384	28.59	13.14	39.48	862.95	256
convnextv2_tiny.fcmae_ft_in22k_in1k	83.894	96.964	224	28.64	4.47	13.44	1452.72	256
convnext_base.fb_in1k	83.82	96.746	224	88.59	15.38	28.75	1054.0	256
convnextv2_nano.fcmae_ft_in22k_in1k_384	83.37	96.742	384	15.62	7.22	24.61	801.72	256
convnext_small.fb_in1k	83.142	96.434	224	50.22	8.71	21.56	1464.0	256
convnextv2_tiny.fcmae_ft_in1k	82.92	96.284	224	28.64	4.47	13.44	1425.62	256
convnext_tiny.fb_in22k_ft_in1k	82.898	96.616	224	28.59	4.47	13.44	2480.88	256
convnext_nano.in12k_ft_in1k	82.282	96.344	224	15.59	2.46	8.37	3926.52	256
convnext_tiny_hnf.a2h_in1k	82.216	95.852	224	28.59	4.47	13.44	2529.75	256
convnext_tiny.fb_in1k	82.066	95.854	224	28.59	4.47	13.44	2346.26	256
convnextv2_nano.fcmae_ft_in22k_in1k	82.03	96.166	224	15.62	2.46	8.37	2300.18	256
convnextv2_nano.fcmae_ft_in1k	81.83	95.738	224	15.62	2.46	8.37	2321.48	256
convnext_nano_ols.d1h_in1k	80.866	95.246	224	15.65	2.65	9.38	3523.85	256
convnext_nano.d1h_in1k	80.768	95.334	224	15.59	2.46	8.37	3915.58	256
convnextv2_pico.fcmae_ft_in1k	80.304	95.072	224	9.07	1.37	6.1	3274.57	256
convnext_pico.d1_in1k	79.526	94.558	224	9.05	1.37	6.1	5686.88	256
convnext_pico_ols.d1_in1k	79.522	94.692	224	9.06	1.43	6.5	5422.46	256
convnextv2_femto.fcmae_ft_in1k	78.488	93.98	224	5.23	0.79	4.57	4264.2	256
convnext_femto_ols.d1_in1k	77.86	93.83	224	5.23	0.82	4.87	6910.6	256
convnext_femto.d1_in1k	77.454	93.68	224	5.22	0.79	4.57	7189.92	256
convnextv2_atto.fcmae_ft_in1k	76.664	93.044	224	3.71	0.55	3.81	4728.91	256
convnext_atto_ols.a2_in1k	75.88	92.846	224	3.7	0.58	4.11	7963.16	256
convnext_atto.d2_in1k	75.664	92.9	224	3.7	0.55	3.81	8439.22	256

Citation

@software{ilharco_gabriel_2021_5143773,
  author       = {Ilharco, Gabriel and
                  Wortsman, Mitchell and
                  Wightman, Ross and
                  Gordon, Cade and
                  Carlini, Nicholas and
                  Taori, Rohan and
                  Dave, Achal and
                  Shankar, Vaishaal and
                  Namkoong, Hongseok and
                  Miller, John and
                  Hajishirzi, Hannaneh and
                  Farhadi, Ali and
                  Schmidt, Ludwig},
  title        = {OpenCLIP},
  month        = jul,
  year         = 2021,
  note         = {If you use this software, please cite it as below.},
  publisher    = {Zenodo},
  version      = {0.1},
  doi          = {10.5281/zenodo.5143773},
  url          = {https://doi.org/10.5281/zenodo.5143773}
}

@inproceedings{schuhmann2022laionb,
  title={{LAION}-5B: An open large-scale dataset for training next generation image-text models},
  author={Christoph Schuhmann and
          Romain Beaumont and
          Richard Vencu and
          Cade W Gordon and
          Ross Wightman and
          Mehdi Cherti and
          Theo Coombes and
          Aarush Katta and
          Clayton Mullis and
          Mitchell Wortsman and
          Patrick Schramowski and
          Srivatsa R Kundurthy and
          Katherine Crowson and
          Ludwig Schmidt and
          Robert Kaczmarczyk and
          Jenia Jitsev},
  booktitle={Thirty-sixth Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
  year={2022},
  url={https://openreview.net/forum?id=M3Y74vmsMcY}
}

i

@inproceedings{Radford2021LearningTV,
  title={Learning Transferable Visual Models From Natural Language Supervision},
  author={Alec Radford and Jong Wook Kim and Chris Hallacy and A. Ramesh and Gabriel Goh and Sandhini Agarwal and Girish Sastry and Amanda Askell and Pamela Mishkin and Jack Clark and Gretchen Krueger and Ilya Sutskever},
  booktitle={ICML}, 
  year={2021}
}

@article{liu2022convnet,
  author  = {Zhuang Liu and Hanzi Mao and Chao-Yuan Wu and Christoph Feichtenhofer and Trevor Darrell and Saining Xie},
  title   = {A ConvNet for the 2020s},
  journal = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year    = {2022},
}