英文

maxxvitv2_rmlp_base_rw_384.sw_in12k_ft_in1k 的模型卡

这是一个基于 timm 的特定MaxxViT-V2模型(使用MLP Log-CPB(受Swin-V2启发的连续对数坐标相对位置偏差)进行图像分类)。在ImageNet-12k(完整ImageNet-22k的11821类子集)上进行预训练,并由Ross Wightman在ImageNet-1k上进行微调。

在8x GPU Lambda Labs 云实例上进行了ImageNet-12k预训练和ImageNet-1k微调。

maxxvit.py 中的模型变体

MaxxViT包括一系列相关的模型架构,它们共享一个共同的结构,包括:

  • CoAtNet-在早期阶段结合MBConv(深度可分离)卷积块和后期阶段的自注意力变换块。
  • MaxViT-在所有阶段均统一使用的块,每个块包含一个MBConv(深度可分离)卷积块,后面是两个具有不同分区方案(窗口和网格)的自注意力块。
  • CoAtNeXt-一个特定于timm的架构,其在CoAtNet中使用ConvNeXt块代替MBConv块。所有的归一化层都是LayerNorm(没有BatchNorm)。
  • MaxxViT-一个特定于timm的架构,其在MaxViT中使用ConvNeXt块代替MBConv块。所有的归一化层都是LayerNorm(没有BatchNorm)。
  • MaxxViT-V2-对MaxxViT的变种,去除了窗口块的注意力,只留下ConvNeXt块和带有更大宽度的网格注意力以进行补偿。

除了上述主要变体之外,模型之间还存在更细微的变化。任何模型名称中包含字符串rw的都是timm特定配置,其中进行了建模调整,以适应PyTorch的即时使用需求。这些模型是在训练初始的模型复现过程中创建的,因此存在一些不同。所有包含字符串tf的模型都是与原始论文作者基于Tensorflow的模型完全匹配的模型,其权重已转换为PyTorch格式。这包括多个MaxViT模型。官方CoAtNet模型从未发布过。

模型细节

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('maxxvitv2_rmlp_base_rw_384.sw_in12k_ft_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)

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(
    'maxxvitv2_rmlp_base_rw_384.sw_in12k_ft_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, 128, 192, 192])
    #  torch.Size([1, 128, 96, 96])
    #  torch.Size([1, 256, 48, 48])
    #  torch.Size([1, 512, 24, 24])
    #  torch.Size([1, 1024, 12, 12])

    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(
    'maxxvitv2_rmlp_base_rw_384.sw_in12k_ft_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, 1024, 12, 12) 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)
12311321 88.53 98.64 21.76 475.77 534.14 1413.22
12312321 88.32 98.54 42.53 475.32 292.78 668.76
12313321 88.20 98.53 50.87 119.88 138.02 703.99
12314321 88.04 98.40 36.42 212.33 244.75 942.15
12315321 87.98 98.56 71.75 212.03 132.55 445.84
12316321 87.92 98.54 104.71 119.65 73.80 332.90
12317321 87.81 98.37 106.55 116.14 70.97 318.95
12318321 87.47 98.37 149.49 116.09 72.98 213.74
12319321 87.39 98.31 160.80 73.88 47.69 209.43
12320321 86.89 98.02 375.86 116.14 23.15 92.64
12321321 86.64 98.02 501.03 116.09 24.20 62.77
12322321 86.60 97.92 50.75 119.88 138.02 703.99
12323321 86.57 97.89 631.88 73.87 15.09 49.22
12324321 86.52 97.88 36.04 212.33 244.75 942.15
12325321 86.49 97.90 620.58 73.88 15.18 54.78
12326321 86.29 97.80 101.09 119.65 73.80 332.90
12327321 86.23 97.69 70.56 212.03 132.55 445.84
12328321 86.10 97.76 88.63 69.13 67.26 383.77
12329321 85.67 97.58 144.25 31.05 33.49 257.59
12330321 85.54 97.46 188.35 69.02 35.87 183.65
12331321 85.11 97.38 293.46 30.98 17.53 123.42
12332321 84.93 96.97 247.71 211.79 43.68 127.35
12333321 84.90 96.96 1025.45 41.72 8.11 40.13
12334321 84.85 96.99 358.25 119.47 24.04 95.01
12335321 84.63 97.06 575.53 66.01 14.67 58.38
12336321 84.61 96.74 625.81 73.88 15.18 54.78
12337321 84.49 96.76 693.82 64.90 10.75 49.30
12338321 84.43 96.83 647.96 68.93 11.66 53.17
12339321 84.23 96.78 807.21 29.15 6.77 46.92
12340321 83.62 96.38 989.59 41.72 8.04 34.60
12341321 83.50 96.50 1100.53 29.06 5.11 33.11
12342321 83.41 96.59 1004.94 30.92 5.60 35.78
12343321 83.36 96.45 1093.03 41.69 7.85 35.47
12344321 83.11 96.33 1276.88 23.70 6.26 23.05
12345321 83.03 96.34 1341.24 16.78 4.37 26.05
12346321 82.96 96.26 1283.24 15.50 4.47 31.92
12347321 82.93 96.23 1218.17 15.45 4.46 30.28
12348321 82.39 96.19 1600.14 27.44 4.67 22.04
12349321 82.39 95.84 1831.21 27.44 4.43 18.73
12350321 82.05 95.87 2109.09 15.15 2.62 20.34
12351321 81.95 95.92 2525.52 14.70 2.47 12.80
12352321 81.70 95.64 2344.52 15.14 2.41 15.41
12353321 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)
12351321 81.95 95.92 2525.52 14.70 2.47 12.80
12352321 81.70 95.64 2344.52 15.14 2.41 15.41
12350321 82.05 95.87 2109.09 15.15 2.62 20.34
12349321 82.39 95.84 1831.21 27.44 4.43 18.73
12348321 82.39 96.19 1600.14 27.44 4.67 22.04
12353321 80.53 95.21 1594.71 7.52 1.85 24.86
12345321 83.03 96.34 1341.24 16.78 4.37 26.05
12346321 82.96 96.26 1283.24 15.50 4.47 31.92
12344321 83.11 96.33 1276.88 23.70 6.26 23.05
12347321 82.93 96.23 1218.17 15.45 4.46 30.28
12341321 83.50 96.50 1100.53 29.06 5.11 33.11
12343321 83.36 96.45 1093.03 41.69 7.85 35.47
12333321 84.90 96.96 1025.45 41.72 8.11 40.13
12342321 83.41 96.59 1004.94 30.92 5.60 35.78
12340321 83.62 96.38 989.59 41.72 8.04 34.60
12339321 84.23 96.78 807.21 29.15 6.77 46.92
12337321 84.49 96.76 693.82 64.90 10.75 49.30
12338321 84.43 96.83 647.96 68.93 11.66 53.17
12323321 86.57 97.89 631.88 73.87 15.09 49.22
12336321 84.61 96.74 625.81 73.88 15.18 54.78
12325321 86.49 97.90 620.58 73.88 15.18 54.78
12335321 84.63 97.06 575.53 66.01 14.67 58.38
12321321 86.64 98.02 501.03 116.09 24.20 62.77
12320321 86.89 98.02 375.86 116.14 23.15 92.64
12334321 84.85 96.99 358.25 119.47 24.04 95.01
12331321 85.11 97.38 293.46 30.98 17.53 123.42
12332321 84.93 96.97 247.71 211.79 43.68 127.35
12330321 85.54 97.46 188.35 69.02 35.87 183.65
12319321 87.39 98.31 160.80 73.88 47.69 209.43
12318321 87.47 98.37 149.49 116.09 72.98 213.74
12329321 85.67 97.58 144.25 31.05 33.49 257.59
12317321 87.81 98.37 106.55 116.14 70.97 318.95
12316321 87.92 98.54 104.71 119.65 73.80 332.90
12326321 86.29 97.80 101.09 119.65 73.80 332.90
12328321 86.10 97.76 88.63 69.13 67.26 383.77
12315321 87.98 98.56 71.75 212.03 132.55 445.84
12327321 86.23 97.69 70.56 212.03 132.55 445.84
12313321 88.20 98.53 50.87 119.88 138.02 703.99
12322321 86.60 97.92 50.75 119.88 138.02 703.99
12312321 88.32 98.54 42.53 475.32 292.78 668.76
12314321 88.04 98.40 36.42 212.33 244.75 942.15
12324321 86.52 97.88 36.04 212.33 244.75 942.15
12311321 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}
}