Awesome

Meta Gradient Fusion: Training Dynamic-Depth Neural Networks Harmoniously

Important Modification:

if you want to apply meta-gf to your multi-task learning program, please see the guideline in QuicklyApplication_MetaGF_ForMultitask.zip

Install:

Pytorch>=1.9.0

Dataset:

CIFAR10 CIFAR100 ImageNet

Usage:

Run the script in the folder named "Script"(select the corresponding networks)

After training finished, run the script in the "Script/TestScript"(select the corresponding networks)

Please downloading the Cifar datasets and put it to the "./data/cifar"

data └── cifar ├── cifar-100-python │ ├── meta │ ├── test │ └── train ├── cifar-100-python.tar.gz ├── cifar-10-batches-py │ ├── batches.meta │ ├── data_batch_1 │ ├── data_batch_2 │ ├── data_batch_3 │ ├── data_batch_4 │ ├── data_batch_5 │ ├── readme.html │ └── test_batch └── cifar-10-python.tar.gz

Notes:

• We adopt channel-wise weighting for vgg because the layer of vgg is too less
• We adopt layer-wise weighting policy for resnet and msdnet
• We adopt EMA updating policy for the meta-weights training
• We train the ImageNet in distributed mode

This version of code may have bugs. We will continue updating it.

New results:

The accuracy of Cagrad on ImageNet: 58.37/64.21/66.88/68.22/69.42

Acknowledgements

We thanks for the public codes provided by the following works:

Li, H., Zhang, H., Qi, X., Yang, R., Huang, G.: Improved techniques for training adaptive deep networks. In: Proceedings of the IEEE/CVF International Confer- ence on Computer Vision. pp. 1891–1900 (2019)

Kaya, Y., Hong, S., Dumitras, T.: Shallow-deep networks: Understanding and mit- igating network overthinking. In: International Conference on Machine Learning. pp. 3301–3310. PMLR