模型:

timm/maxvit_xlarge_tf_512.in21k_ft_in1k

英文

maxvit_xlarge_tf_512.in21k_ft_in1k的模型卡片

官方的MaxViT图像分类模型。在ImageNet-21k(21843个谷歌特定实例的ImageNet-22k)上使用tensorflow进行预训练,并由论文作者在ImageNet-1k上进行微调。

由Ross Wightman将官方的Tensorflow实现( https://github.com/google-research/maxvit )转换为PyTorch。

maxxvit.py 中的模型变种

MaxxViT涵盖了许多相关的模型架构,它们共享一个共同的结构,包括:

  • CoAtNet - 在早期阶段将MBConv(深度可分离)卷积块与后期的自注意力变换块结合起来。
  • MaxViT - 在所有阶段中均使用统一的块,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划分方案(窗口后跟网格)的自注意力块。
  • CoAtNeXt - 这是一个timm特定的架构,它在CoAtNet中使用ConvNeXt块而不是MBConv块。所有的归一化层都是LayerNorm(没有BatchNorm)。
  • MaxxViT - 这是一个timm特定的架构,它在MaxViT中使用ConvNeXt块而不是MBConv块。所有的归一化层都是LayerNorm(没有BatchNorm)。
  • MaxxViT-V2 - MaxxViT的一个变种,删除了窗口块注意力,只保留了ConvNeXt块和网格注意力,通过增加宽度来进行补偿。

除了上面列出的主要变体外,从一个模型到另一个模型还有一些细微的变化。带有字符串rw的模型名称是timm特定的配置,其中进行了模型调整以支持PyTorch的即时使用。这些模型是在训练模型的初始重现期间创建的,因此变种较多。所有带有字符串tf的模型均与原始论文作者基于Tensorflow的模型完全匹配,将权重移植到PyTorch上。这涵盖了许多MaxViT模型。官方的CoAtNet模型从未发布过。

模型详细信息

  • 模型类型:图像分类/特征骨干
  • 模型统计信息:
    • 参数数目(M):475.8
    • GMACs:534.1
    • 激活数目(M):1413.2
    • 图像尺寸:512 x 512
  • 论文:
  • 数据集:ImageNet-1k
  • 预训练数据集:ImageNet-21k

模型用法

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('maxvit_xlarge_tf_512.in21k_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(
    'maxvit_xlarge_tf_512.in21k_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, 192, 256, 256])
    #  torch.Size([1, 192, 128, 128])
    #  torch.Size([1, 384, 64, 64])
    #  torch.Size([1, 768, 32, 32])
    #  torch.Size([1, 1536, 16, 16])

    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(
    'maxvit_xlarge_tf_512.in21k_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, 1536, 16, 16) 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

model top1 top5 samples / sec Params (M) GMAC Act (M)
1239321 88.53 98.64 21.76 475.77 534.14 1413.22
12310321 88.32 98.54 42.53 475.32 292.78 668.76
12311321 88.20 98.53 50.87 119.88 138.02 703.99
12312321 88.04 98.40 36.42 212.33 244.75 942.15
12313321 87.98 98.56 71.75 212.03 132.55 445.84
12314321 87.92 98.54 104.71 119.65 73.80 332.90
12315321 87.81 98.37 106.55 116.14 70.97 318.95
12316321 87.47 98.37 149.49 116.09 72.98 213.74
12317321 87.39 98.31 160.80 73.88 47.69 209.43
12318321 86.89 98.02 375.86 116.14 23.15 92.64
12319321 86.64 98.02 501.03 116.09 24.20 62.77
12320321 86.60 97.92 50.75 119.88 138.02 703.99
12321321 86.57 97.89 631.88 73.87 15.09 49.22
12322321 86.52 97.88 36.04 212.33 244.75 942.15
12323321 86.49 97.90 620.58 73.88 15.18 54.78
12324321 86.29 97.80 101.09 119.65 73.80 332.90
12325321 86.23 97.69 70.56 212.03 132.55 445.84
12326321 86.10 97.76 88.63 69.13 67.26 383.77
12327321 85.67 97.58 144.25 31.05 33.49 257.59
12328321 85.54 97.46 188.35 69.02 35.87 183.65
12329321 85.11 97.38 293.46 30.98 17.53 123.42
12330321 84.93 96.97 247.71 211.79 43.68 127.35
12331321 84.90 96.96 1025.45 41.72 8.11 40.13
12332321 84.85 96.99 358.25 119.47 24.04 95.01
12333321 84.63 97.06 575.53 66.01 14.67 58.38
12334321 84.61 96.74 625.81 73.88 15.18 54.78
12335321 84.49 96.76 693.82 64.90 10.75 49.30
12336321 84.43 96.83 647.96 68.93 11.66 53.17
12337321 84.23 96.78 807.21 29.15 6.77 46.92
12338321 83.62 96.38 989.59 41.72 8.04 34.60
12339321 83.50 96.50 1100.53 29.06 5.11 33.11
12340321 83.41 96.59 1004.94 30.92 5.60 35.78
12341321 83.36 96.45 1093.03 41.69 7.85 35.47
12342321 83.11 96.33 1276.88 23.70 6.26 23.05
12343321 83.03 96.34 1341.24 16.78 4.37 26.05
12344321 82.96 96.26 1283.24 15.50 4.47 31.92
12345321 82.93 96.23 1218.17 15.45 4.46 30.28
12346321 82.39 96.19 1600.14 27.44 4.67 22.04
12347321 82.39 95.84 1831.21 27.44 4.43 18.73
12348321 82.05 95.87 2109.09 15.15 2.62 20.34
12349321 81.95 95.92 2525.52 14.70 2.47 12.80
12350321 81.70 95.64 2344.52 15.14 2.41 15.41
12351321 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)
12349321 81.95 95.92 2525.52 14.70 2.47 12.80
12350321 81.70 95.64 2344.52 15.14 2.41 15.41
12348321 82.05 95.87 2109.09 15.15 2.62 20.34
12347321 82.39 95.84 1831.21 27.44 4.43 18.73
12346321 82.39 96.19 1600.14 27.44 4.67 22.04
12351321 80.53 95.21 1594.71 7.52 1.85 24.86
12343321 83.03 96.34 1341.24 16.78 4.37 26.05
12344321 82.96 96.26 1283.24 15.50 4.47 31.92
12342321 83.11 96.33 1276.88 23.70 6.26 23.05
12345321 82.93 96.23 1218.17 15.45 4.46 30.28
12339321 83.50 96.50 1100.53 29.06 5.11 33.11
12341321 83.36 96.45 1093.03 41.69 7.85 35.47
12331321 84.90 96.96 1025.45 41.72 8.11 40.13
12340321 83.41 96.59 1004.94 30.92 5.60 35.78
12338321 83.62 96.38 989.59 41.72 8.04 34.60
12337321 84.23 96.78 807.21 29.15 6.77 46.92
12335321 84.49 96.76 693.82 64.90 10.75 49.30
12336321 84.43 96.83 647.96 68.93 11.66 53.17
12321321 86.57 97.89 631.88 73.87 15.09 49.22
12334321 84.61 96.74 625.81 73.88 15.18 54.78
12323321 86.49 97.90 620.58 73.88 15.18 54.78
12333321 84.63 97.06 575.53 66.01 14.67 58.38
12319321 86.64 98.02 501.03 116.09 24.20 62.77
12318321 86.89 98.02 375.86 116.14 23.15 92.64
12332321 84.85 96.99 358.25 119.47 24.04 95.01
12329321 85.11 97.38 293.46 30.98 17.53 123.42
12330321 84.93 96.97 247.71 211.79 43.68 127.35
12328321 85.54 97.46 188.35 69.02 35.87 183.65
12317321 87.39 98.31 160.80 73.88 47.69 209.43
12316321 87.47 98.37 149.49 116.09 72.98 213.74
12327321 85.67 97.58 144.25 31.05 33.49 257.59
12315321 87.81 98.37 106.55 116.14 70.97 318.95
12314321 87.92 98.54 104.71 119.65 73.80 332.90
12324321 86.29 97.80 101.09 119.65 73.80 332.90
12326321 86.10 97.76 88.63 69.13 67.26 383.77
12313321 87.98 98.56 71.75 212.03 132.55 445.84
12325321 86.23 97.69 70.56 212.03 132.55 445.84
12311321 88.20 98.53 50.87 119.88 138.02 703.99
12320321 86.60 97.92 50.75 119.88 138.02 703.99
12310321 88.32 98.54 42.53 475.32 292.78 668.76
12312321 88.04 98.40 36.42 212.33 244.75 942.15
12322321 86.52 97.88 36.04 212.33 244.75 942.15
1239321 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}
}