Model card for hgnet_base.ssld_in1k

A HGNet (High Performance GPU Net) image classification model. Trained by model authors on mined ImageNet-22k and ImageNet-1k using SSLD distillation.

Please see details at https://github.com/PaddlePaddle/PaddleClas/blob/develop/docs/zh_CN/models/ImageNet1k/PP-HGNetV2.md

Model Details

• Model Type: Image classification / feature backbone
• Model Stats:
  • Params (M): 71.6
  • GMACs: 25.1
  • Activations (M): 15.5
  • Image size: train = 224 x 224, test = 288 x 288
• Pretrain Dataset: ImageNet-22k
• Dataset: ImageNet-1k
• Papers:
  • Model paper unknown: TBD
  • Beyond Self-Supervision: A Simple Yet Effective Network Distillation Alternative to Improve Backbones: https://arxiv.org/abs/2103.05959
• Original: https://github.com/PaddlePaddle/PaddleClas

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('hgnet_base.ssld_in1k', 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(
    'hgnet_base.ssld_in1k',
    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, 320, 56, 56])
    #  torch.Size([1, 640, 28, 28])
    #  torch.Size([1, 960, 14, 14])
    #  torch.Size([1, 1280, 7, 7])

    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(
    'hgnet_base.ssld_in1k',
    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, 1280, 7, 7) shaped tensor

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

Model Comparison

By Top-1

model	top1	top1_err	top5	top5_err	param_count	img_size
hgnetv2_b6.ssld_stage2_ft_in1k	86.36	13.64	97.934	2.066	75.26	288
hgnetv2_b6.ssld_stage1_in22k_in1k	86.294	13.706	97.948	2.052	75.26	288
hgnetv2_b6.ssld_stage2_ft_in1k	86.204	13.796	97.81	2.19	75.26	224
hgnetv2_b6.ssld_stage1_in22k_in1k	86.028	13.972	97.804	2.196	75.26	224
hgnet_base.ssld_in1k	85.474	14.526	97.632	2.368	71.58	288
hgnetv2_b5.ssld_stage2_ft_in1k	85.146	14.854	97.612	2.388	39.57	288
hgnetv2_b5.ssld_stage1_in22k_in1k	84.928	15.072	97.514	2.486	39.57	288
hgnet_base.ssld_in1k	84.912	15.088	97.342	2.658	71.58	224
hgnetv2_b5.ssld_stage2_ft_in1k	84.808	15.192	97.3	2.7	39.57	224
hgnetv2_b5.ssld_stage1_in22k_in1k	84.458	15.542	97.22	2.78	39.57	224
hgnet_small.ssld_in1k	84.376	15.624	97.128	2.872	24.36	288
hgnetv2_b4.ssld_stage2_ft_in1k	83.912	16.088	97.06	2.94	19.8	288
hgnet_small.ssld_in1k	83.808	16.192	96.848	3.152	24.36	224
hgnetv2_b4.ssld_stage2_ft_in1k	83.694	16.306	96.786	3.214	19.8	224
hgnetv2_b3.ssld_stage2_ft_in1k	83.58	16.42	96.81	3.19	16.29	288
hgnetv2_b4.ssld_stage1_in22k_in1k	83.45	16.55	96.92	3.08	19.8	288
hgnetv2_b3.ssld_stage1_in22k_in1k	83.116	16.884	96.712	3.288	16.29	288
hgnetv2_b3.ssld_stage2_ft_in1k	82.916	17.084	96.364	3.636	16.29	224
hgnetv2_b4.ssld_stage1_in22k_in1k	82.892	17.108	96.632	3.368	19.8	224
hgnetv2_b3.ssld_stage1_in22k_in1k	82.588	17.412	96.38	3.62	16.29	224
hgnet_tiny.ssld_in1k	82.524	17.476	96.514	3.486	14.74	288
hgnetv2_b2.ssld_stage2_ft_in1k	82.346	17.654	96.394	3.606	11.22	288
hgnet_small.paddle_in1k	82.222	17.778	96.22	3.78	24.36	288
hgnet_tiny.ssld_in1k	81.938	18.062	96.114	3.886	14.74	224
hgnetv2_b2.ssld_stage2_ft_in1k	81.578	18.422	95.896	4.104	11.22	224
hgnetv2_b2.ssld_stage1_in22k_in1k	81.46	18.54	96.01	3.99	11.22	288
hgnet_small.paddle_in1k	81.358	18.642	95.832	4.168	24.36	224
hgnetv2_b2.ssld_stage1_in22k_in1k	80.75	19.25	95.498	4.502	11.22	224
hgnet_tiny.paddle_in1k	80.64	19.36	95.54	4.46	14.74	288
hgnetv2_b1.ssld_stage2_ft_in1k	79.904	20.096	95.148	4.852	6.34	288
hgnet_tiny.paddle_in1k	79.894	20.106	95.052	4.948	14.74	224
hgnetv2_b1.ssld_stage1_in22k_in1k	79.048	20.952	94.882	5.118	6.34	288
hgnetv2_b1.ssld_stage2_ft_in1k	78.872	21.128	94.492	5.508	6.34	224
hgnetv2_b0.ssld_stage2_ft_in1k	78.586	21.414	94.388	5.612	6.0	288
hgnetv2_b1.ssld_stage1_in22k_in1k	78.05	21.95	94.182	5.818	6.34	224
hgnetv2_b0.ssld_stage1_in22k_in1k	78.026	21.974	94.242	5.758	6.0	288
hgnetv2_b0.ssld_stage2_ft_in1k	77.342	22.658	93.786	6.214	6.0	224
hgnetv2_b0.ssld_stage1_in22k_in1k	76.844	23.156	93.612	6.388	6.0	224

Citation

@article{cui2021beyond,
  title={Beyond Self-Supervision: A Simple Yet Effective Network Distillation Alternative to Improve Backbones},
  author={Cui, Cheng and Guo, Ruoyu and Du, Yuning and He, Dongliang and Li, Fu and Wu, Zewu and Liu, Qiwen and Wen, Shilei and Huang, Jizhou and Hu, Xiaoguang and others},
  journal={arXiv preprint arXiv:2103.05959},
  year={2021}
}