Are you sure you want to delete this task? Once this task is deleted, it cannot be recovered.
EmmaHAN 062fe1a7c9 | 3 months ago | |
---|---|---|
.. | ||
README.md | 8 months ago | |
halonet_50t_ascend.yaml | 3 months ago |
Scaling Local Self-Attention for Parameter Efficient Visual Backbones
Researchers from Google Research and UC Berkeley have developed a new model of self-attention that can outperform standard baseline models and even high-performance convolutional models.[1]
Blocked Self-Attention:The whole input image is divided into multiple blocks and self-attention is applied to each block.However, if only the information inside the block is considered each time, it will inevitably lead to the loss of information.Therefore, before calculating the SA, a haloing operation is performed on each block, i.e., outside of each block, the information of the original image is used to padding a circle, so that the sensory field of each block can be appropriately larger and focus on more information.
Figure 1. Architecture of Blocked Self-Attention [1]
Down Sampling:In order to reduce the amount of computation, each block is sampled separately, and then attentions are performed on this sampled information to reach the effect of down sampling.
Figure 2. Architecture of Down Sampling [1]
Our reproduced model performance on ImageNet-1K is reported as follows.
Model | Context | Top-1 (%) | Top-5 (%) | Params (M) | Recipe | Download |
---|---|---|---|---|---|---|
halonet_50t | D910X8-G | 79.53 | 94.79 | 22.79 | yaml | weights |
Please refer to the installation instruction in MindCV.
Please download the ImageNet-1K dataset for model training and validation.
It is easy to reproduce the reported results with the pre-defined training recipe. For distributed training on multiple Ascend 910 devices, please run
# distributed training on multiple GPU/Ascend devices
mpirun -n 8 python train.py --config configs/halonet/halonet_50t_ascend.yaml --data_dir /path/to/imagenet
If the script is executed by the root user, the
--allow-run-as-root
parameter must be added tompirun
.
Similarly, you can train the model on multiple GPU devices with the above mpirun
command.
For detailed illustration of all hyper-parameters, please refer to config.py.
Note: As the global batch size (batch_size x num_devices) is an important hyper-parameter, it is recommended to keep the global batch size unchanged for reproduction or adjust the learning rate linearly to a new global batch size.
If you want to train or finetune the model on a smaller dataset without distributed training, please run:
# standalone training on a CPU/GPU/Ascend device
python train.py --config configs/halonet/halonet_50t_ascend.yaml --data_dir /path/to/dataset --distribute False
To validate the accuracy of the trained model, you can use validate.py
and parse the checkpoint path with --ckpt_path
.
python validate.py -c configs/halonet/halonet_50t_ascend.yaml --data_dir /path/to/imagenet --ckpt_path /path/to/ckpt
Please refer to the deployment tutorial in MindCV.
[1] Vaswani A, Ramachandran P, Srinivas A, et al. Scaling local self-attention for parameter efficient visual backbones[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2021: 12894-12904.
A toolbox of vision models and algorithms based on MindSpore
https://github.com/mindspore-lab/mindcv
Python Markdown other
Dear OpenI User
Thank you for your continuous support to the Openl Qizhi Community AI Collaboration Platform. In order to protect your usage rights and ensure network security, we updated the Openl Qizhi Community AI Collaboration Platform Usage Agreement in January 2024. The updated agreement specifies that users are prohibited from using intranet penetration tools. After you click "Agree and continue", you can continue to use our services. Thank you for your cooperation and understanding.
For more agreement content, please refer to the《Openl Qizhi Community AI Collaboration Platform Usage Agreement》