模型:

timm/coatnet_3_rw_224.sw_in12k

英文

coatnet_3_rw_224.sw_in12k模型卡片

一种特定于timm的CoAtNet图像分类模型。由Ross Wightman在ImageNet-12k上使用timm进行训练(ImageNet-12k是ImageNet-22k的一个11821类子集)。

模型变体 maxxvit.py

MaxxViT涵盖了一系列相关的模型架构,这些模型架构共享相同的结构,包括:

  • CoAtNet - 在早期阶段将MBConv(深度可分离)卷积块与后期的自注意力Transformer块相结合。
  • MaxViT - 在所有阶段中使用统一的块,每个块包含一个MBConv(深度可分离)卷积块,后面跟随两个具有不同分区方案(窗口和网格)的自注意力块。
  • CoAtNeXt - timm特定的架构,将MBConv块替换为CoAtNet中的ConvNeXt块。所有归一化层都是LayerNorm(没有BatchNorm)。
  • MaxxViT - timm特定的架构,将MBConv块替换为MaxViT中的ConvNeXt块。所有归一化层都是LayerNorm(没有BatchNorm)。
  • MaxxViT-V2 - MaxxViT的一个变体,去除了窗口块关注,只保留ConvNeXt块和网格注意力,并增加了更多的宽度以进行补偿。

除了上述主要变体之外,模型之间还存在细微的差异。带有字符串rw的模型名称是timm特定的配置,具有有利于PyTorch eager使用的建模调整。这些模型是在训练初始复现模型时创建的,因此存在一些变化。所有包含字符串tf的模型与原始论文作者基于Tensorflow的模型完全相匹配,并将权重转移到PyTorch中。这涵盖了许多MaxViT模型。官方的CoAtNet模型从未发布过。

模型详情

  • 模型类型:图像分类/特征主干
  • 模型统计信息:
    • 参数数量(M):181.8
    • GMACs:33.4
    • 激活函数数量(M):73.8
    • 图像尺寸:224 x 224
  • 论文:
  • 数据集: ImageNet-12k

模型用途

图像分类

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('coatnet_3_rw_224.sw_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)

特征图提取

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(
    'coatnet_3_rw_224.sw_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, 192, 112, 112])
    #  torch.Size([1, 192, 56, 56])
    #  torch.Size([1, 384, 28, 28])
    #  torch.Size([1, 768, 14, 14])
    #  torch.Size([1, 1536, 7, 7])

    print(o.shape)

图像嵌入

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(
    'coatnet_3_rw_224.sw_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, 1536, 7, 7) shaped tensor

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

模型比较

按Top-1指标

model top1 top5 samples / sec Params (M) GMAC Act (M)
1238321 88.53 98.64 21.76 475.77 534.14 1413.22
1239321 88.32 98.54 42.53 475.32 292.78 668.76
12310321 88.20 98.53 50.87 119.88 138.02 703.99
12311321 88.04 98.40 36.42 212.33 244.75 942.15
12312321 87.98 98.56 71.75 212.03 132.55 445.84
12313321 87.92 98.54 104.71 119.65 73.80 332.90
12314321 87.81 98.37 106.55 116.14 70.97 318.95
12315321 87.47 98.37 149.49 116.09 72.98 213.74
12316321 87.39 98.31 160.80 73.88 47.69 209.43
12317321 86.89 98.02 375.86 116.14 23.15 92.64
12318321 86.64 98.02 501.03 116.09 24.20 62.77
12319321 86.60 97.92 50.75 119.88 138.02 703.99
12320321 86.57 97.89 631.88 73.87 15.09 49.22
12321321 86.52 97.88 36.04 212.33 244.75 942.15
12322321 86.49 97.90 620.58 73.88 15.18 54.78
12323321 86.29 97.80 101.09 119.65 73.80 332.90
12324321 86.23 97.69 70.56 212.03 132.55 445.84
12325321 86.10 97.76 88.63 69.13 67.26 383.77
12326321 85.67 97.58 144.25 31.05 33.49 257.59
12327321 85.54 97.46 188.35 69.02 35.87 183.65
12328321 85.11 97.38 293.46 30.98 17.53 123.42
12329321 84.93 96.97 247.71 211.79 43.68 127.35
12330321 84.90 96.96 1025.45 41.72 8.11 40.13
12331321 84.85 96.99 358.25 119.47 24.04 95.01
12332321 84.63 97.06 575.53 66.01 14.67 58.38
12333321 84.61 96.74 625.81 73.88 15.18 54.78
12334321 84.49 96.76 693.82 64.90 10.75 49.30
12335321 84.43 96.83 647.96 68.93 11.66 53.17
12336321 84.23 96.78 807.21 29.15 6.77 46.92
12337321 83.62 96.38 989.59 41.72 8.04 34.60
12338321 83.50 96.50 1100.53 29.06 5.11 33.11
12339321 83.41 96.59 1004.94 30.92 5.60 35.78
12340321 83.36 96.45 1093.03 41.69 7.85 35.47
12341321 83.11 96.33 1276.88 23.70 6.26 23.05
12342321 83.03 96.34 1341.24 16.78 4.37 26.05
12343321 82.96 96.26 1283.24 15.50 4.47 31.92
12344321 82.93 96.23 1218.17 15.45 4.46 30.28
12345321 82.39 96.19 1600.14 27.44 4.67 22.04
12346321 82.39 95.84 1831.21 27.44 4.43 18.73
12347321 82.05 95.87 2109.09 15.15 2.62 20.34
12348321 81.95 95.92 2525.52 14.70 2.47 12.80
12349321 81.70 95.64 2344.52 15.14 2.41 15.41
12350321 80.53 95.21 1594.71 7.52 1.85 24.86

按吞吐量(样本/秒)

model top1 top5 samples / sec Params (M) GMAC Act (M)
12348321 81.95 95.92 2525.52 14.70 2.47 12.80
12349321 81.70 95.64 2344.52 15.14 2.41 15.41
12347321 82.05 95.87 2109.09 15.15 2.62 20.34
12346321 82.39 95.84 1831.21 27.44 4.43 18.73
12345321 82.39 96.19 1600.14 27.44 4.67 22.04
12350321 80.53 95.21 1594.71 7.52 1.85 24.86
12342321 83.03 96.34 1341.24 16.78 4.37 26.05
12343321 82.96 96.26 1283.24 15.50 4.47 31.92
12341321 83.11 96.33 1276.88 23.70 6.26 23.05
12344321 82.93 96.23 1218.17 15.45 4.46 30.28
12338321 83.50 96.50 1100.53 29.06 5.11 33.11
12340321 83.36 96.45 1093.03 41.69 7.85 35.47
12330321 84.90 96.96 1025.45 41.72 8.11 40.13
12339321 83.41 96.59 1004.94 30.92 5.60 35.78
12337321 83.62 96.38 989.59 41.72 8.04 34.60
12336321 84.23 96.78 807.21 29.15 6.77 46.92
12334321 84.49 96.76 693.82 64.90 10.75 49.30
12335321 84.43 96.83 647.96 68.93 11.66 53.17
12320321 86.57 97.89 631.88 73.87 15.09 49.22
12333321 84.61 96.74 625.81 73.88 15.18 54.78
12322321 86.49 97.90 620.58 73.88 15.18 54.78
12332321 84.63 97.06 575.53 66.01 14.67 58.38
12318321 86.64 98.02 501.03 116.09 24.20 62.77
12317321 86.89 98.02 375.86 116.14 23.15 92.64
12331321 84.85 96.99 358.25 119.47 24.04 95.01
12328321 85.11 97.38 293.46 30.98 17.53 123.42
12329321 84.93 96.97 247.71 211.79 43.68 127.35
12327321 85.54 97.46 188.35 69.02 35.87 183.65
12316321 87.39 98.31 160.80 73.88 47.69 209.43
12315321 87.47 98.37 149.49 116.09 72.98 213.74
12326321 85.67 97.58 144.25 31.05 33.49 257.59
12314321 87.81 98.37 106.55 116.14 70.97 318.95
12313321 87.92 98.54 104.71 119.65 73.80 332.90
12323321 86.29 97.80 101.09 119.65 73.80 332.90
12325321 86.10 97.76 88.63 69.13 67.26 383.77
12312321 87.98 98.56 71.75 212.03 132.55 445.84
12324321 86.23 97.69 70.56 212.03 132.55 445.84
12310321 88.20 98.53 50.87 119.88 138.02 703.99
12319321 86.60 97.92 50.75 119.88 138.02 703.99
1239321 88.32 98.54 42.53 475.32 292.78 668.76
12311321 88.04 98.40 36.42 212.33 244.75 942.15
12321321 86.52 97.88 36.04 212.33 244.75 942.15
1238321 88.53 98.64 21.76 475.77 534.14 1413.22

引用

@misc{rw2019timm,
  author = {Ross Wightman},
  title = {PyTorch Image Models},
  year = {2019},
  publisher = {GitHub},
  journal = {GitHub repository},
  doi = {10.5281/zenodo.4414861},
  howpublished = {\url{https://github.com/huggingface/pytorch-image-models}}
}
@article{tu2022maxvit,
  title={MaxViT: Multi-Axis Vision Transformer},
  author={Tu, Zhengzhong and Talebi, Hossein and Zhang, Han and Yang, Feng and Milanfar, Peyman and Bovik, Alan and Li, Yinxiao},
  journal={ECCV},
  year={2022},
}        
@article{dai2021coatnet,
  title={CoAtNet: Marrying Convolution and Attention for All Data Sizes},
  author={Dai, Zihang and Liu, Hanxiao and Le, Quoc V and Tan, Mingxing},
  journal={arXiv preprint arXiv:2106.04803},
  year={2021}
}