Skip to content

Latest commit

 

History

History
62 lines (50 loc) · 10.9 KB

classification.md

File metadata and controls

62 lines (50 loc) · 10.9 KB
Raw

ImageNet-1k Classification

Remarks: All the models are directly converted from the previous implementation, the AOGNet-V2 repo. There are very minor performance difference. This refactored code can be used to retrain those or train new models from scratch.

Top-1 and Top-5 Accuracy (w/ a normal training setup)

Model Params (M) imagenet-1k imagenet-1k-reassessed imagenetv2-matched-frequency imagenetv2-threshold0.7 imagenetv2-topimages imagenet-sketch imagenet-a imagenet-o
aognet_12m_bn_imagenet 12.259 (0.775, 0.939) (0.84, 0.961) (0.648, 0.861) (0.738, 0.929) (0.79, 0.953) (0.257, 0.447) (0.054, 0.318) (0.017, 0.054)
aognet_12m_an_imagenet 12.373 (0.787, 0.942) (0.849, 0.964) (0.666, 0.873) (0.757, 0.936) (0.802, 0.959) (0.269, 0.454) (0.095, 0.385) (0.019, 0.054)
aognet_40m_bn_imagenet 40.153 (0.802, 0.951) (0.859, 0.969) (0.682, 0.882) (0.769, 0.942) (0.812, 0.964) (0.277, 0.459) (0.118, 0.42) (0.016, 0.057)
aognet_40m_an_imagenet 40.389 (0.807, 0.953) (0.861, 0.97) (0.692, 0.886) (0.779, 0.944) (0.823, 0.963) (0.287, 0.474) (0.157, 0.462) (0.017, 0.055)
resnet_34_bn_imagenet 21.798 (0.745, 0.919) (0.812, 0.948) (0.62, 0.837) (0.708, 0.91) (0.765, 0.943) (0.233, 0.407) (0.022, 0.215) (0.018, 0.052)
resnet_34_an_bn1_imagenet 21.996 (0.755, 0.926) (0.822, 0.952) (0.64, 0.851) (0.724, 0.917) (0.781, 0.947) (0.24, 0.415) (0.038, 0.259) (0.018, 0.058)
resnet_34_an_bn12_imagenet 22.194 (0.751, 0.924) (0.814, 0.95) (0.629, 0.845) (0.713, 0.912) (0.771, 0.944) (0.227, 0.398) (0.044, 0.271) (0.016, 0.052)
resnet_50_bn_imagenet 25.557 (0.769, 0.934) (0.835, 0.958) (0.65, 0.857) (0.742, 0.928) (0.79, 0.952) (0.247, 0.422) (0.033, 0.267) (0.02, 0.06)
resnet_50_an_bn2_imagenet 25.755 (0.784, 0.941) (0.844, 0.961) (0.666, 0.865) (0.748, 0.929) (0.798, 0.955) (0.247, 0.422) (0.082, 0.349) (0.017, 0.053)
resnet_50_an_bn3_imagenet 26.349 (0.782, 0.94) (0.839, 0.961) (0.659, 0.863) (0.746, 0.93) (0.796, 0.955) (0.246, 0.416) (0.086, 0.36) (0.014, 0.049)
resnet_50_an_all_imagenet 26.947 (0.778, 0.938) (0.837, 0.96) (0.657, 0.859) (0.743, 0.925) (0.794, 0.952) (0.236, 0.407) (0.084, 0.356) (0.016, 0.048)
resnet_101_bn_imagenet 44.549 (0.788, 0.942) (0.849, 0.963) (0.667, 0.866) (0.757, 0.935) (0.805, 0.958) (0.276, 0.456) (0.064, 0.323) (0.021, 0.059)
resnet_101_an_imagenet 45.001 (0.794, 0.946) (0.852, 0.965) (0.675, 0.877) (0.759, 0.936) (0.806, 0.959) (0.277, 0.456) (0.118, 0.404) (0.017, 0.051)
mobilenetv2_bn_imagenet 3.505 (0.717, 0.904) (0.79, 0.937) (0.585, 0.81) (0.678, 0.889) (0.738, 0.925) (0.185, 0.343) (0.016, 0.186) (0.016, 0.049)
mobilenetv2_an_imagenet 3.673 (0.732, 0.912) (0.803, 0.943) (0.605, 0.823) (0.695, 0.899) (0.756, 0.933) (0.204, 0.372) (0.028, 0.225) (0.018, 0.05)
densenet_121_an_imagenet 8.342 (0.774, 0.936) (0.837, 0.961) (0.651, 0.86) (0.742, 0.926) (0.792, 0.954) (0.249, 0.427) (0.085, 0.37) (0.014, 0.053)
  • Models trained with a normal training setup: 120 epochs, cosine learning rate scheduler, SGD+Momentum, etc. (MobileNet-v2 trained with 150 epochs). All models are trained with 8 Nvidia V100 GPUs. See Attentive Normalization(ECCV2020) for details.
  • 'bn': BatchNorm (BN) is used.
  • 'an': Attentive Normalization (AN) with BN backbone is used. 'an_bn1' means that 'an' is used to replace the first 'bn' in a building block (basic block or bottleneck block), and similarly for 'an_bn2', 'an_bn3' and 'an_bn12'. 'an_all' means that all 'bn' layers are replaced by 'an'.

Top-1 and Top-5 Accuracy (w/ an advance training setup)

Model Params (M) imagenet-1k imagenet-1k-reassessed imagenetv2-matched-frequency imagenetv2-threshold0.7 imagenetv2-topimages imagenet-sketch imagenet-a imagenet-o
aognet_12m_bn_imagenet_200e_ls_mixup 12.259 (0.782, 0.943) (0.85, 0.967) (0.666, 0.869) (0.752, 0.935) (0.8, 0.958) (0.286, 0.471) (0.086, 0.367) (0.017, 0.053)
aognet_12m_an_imagenet_200e_ls_mixup 12.373 (0.795, 0.947) (0.857, 0.969) (0.681, 0.88) (0.767, 0.943) (0.815, 0.963) (0.285, 0.471) (0.133, 0.426) (0.019, 0.055)
aognet_40m_bn_imagenet_200e_ls_mixup 40.153 (0.812, 0.956) (0.868, 0.972) (0.699, 0.891) (0.781, 0.948) (0.825, 0.967) (0.306, 0.489) (0.184, 0.496) (0.017, 0.052)
aognet_40m_an_imagenet_200e_ls_mixup 40.389 (0.818, 0.957) (0.872, 0.973) (0.71, 0.898) (0.792, 0.951) (0.83, 0.968) (0.306, 0.49) (0.233, 0.549) (0.017, 0.052)
resnetv1d_50_bn_imagenet_200e_ls_mixup 25.576 (0.792, 0.945) (0.851, 0.965) (0.669, 0.868) (0.758, 0.935) (0.805, 0.959) (0.271, 0.45) (0.088, 0.359) (0.019, 0.055)
resnetv1d_50_an_imagenet_200e_ls_mixup 25.775 (0.8, 0.95) (0.857, 0.969) (0.686, 0.882) (0.768, 0.937) (0.817, 0.961) (0.265, 0.441) (0.146, 0.432) (0.016, 0.049)
resnetv1d_101_bn_imagenet_200e_ls_mixup 44.568 (0.802, 0.951) (0.86, 0.969) (0.685, 0.882) (0.772, 0.945) (0.817, 0.967) (0.295, 0.477) (0.141, 0.43) (0.018, 0.051)
resnetv1d_101_an_imagenet_200e_ls_mixup 45.020 (0.81, 0.954) (0.863, 0.971) (0.696, 0.887) (0.778, 0.945) (0.82, 0.965) (0.3, 0.483) (0.199, 0.488) (0.017, 0.049)
  • Models trained with an advanced training setup: 200 epochs, 0.1 label smoothing, 0.2 mixup, cosine learning rate scheduler, SGD+Momentum, etc. All models are trained with 8 Nvidia V100 GPUs. See Attentive Normalization(ECCV2020) for details.

Top-1 and Top-5 Accuracy (AN as a strong alternative to the Squeeze-Excitation (SE) module)

Model Params (M) imagenet-1k imagenet-1k-reassessed imagenetv2-matched-frequency imagenetv2-threshold0.7 imagenetv2-topimages imagenet-sketch imagenet-a imagenet-o
resnet_50_se_bn2_imagenet 26.186 (0.779, 0.939) (0.842, 0.961) (0.658, 0.867) (0.748, 0.933) (0.796, 0.957) (0.236, 0.409) (0.062, 0.326) (0.018, 0.054)
resnet_50_an_bn2_imagenet 25.755 (0.784, 0.941) (0.844, 0.961) (0.666, 0.865) (0.748, 0.929) (0.798, 0.955) (0.247, 0.422) (0.082, 0.349) (0.017, 0.053)
resnet_50_se_bn3_imagenet 28.072 (0.777, 0.939) (0.84, 0.961) (0.654, 0.863) (0.74, 0.928) (0.794, 0.954) (0.239, 0.411) (0.06, 0.318) (0.017, 0.057)
resnet_50_an_bn3_imagenet 26.349 (0.782, 0.94) (0.839, 0.961) (0.659, 0.863) (0.746, 0.93) (0.796, 0.955) (0.246, 0.416) (0.086, 0.36) (0.014, 0.049)
resnet_50_se_bn123_imagenet 29.329 (0.779, 0.94) (0.842, 0.962) (0.658, 0.863) (0.745, 0.926) (0.796, 0.952) (0.235, 0.405) (0.061, 0.328) (0.018, 0.056)
resnet_50_an_all_imagenet 26.947 (0.778, 0.938) (0.837, 0.96) (0.657, 0.859) (0.743, 0.925) (0.794, 0.952) (0.236, 0.407) (0.084, 0.356) (0.016, 0.048)