Vision Outlooker (VOLO), a novel outlook attention, presents a simple and general architecture. Unlike self-attention that focuses on global dependency modeling at a coarse level, the outlook attention efficiently encodes finer-level features and contexts into tokens, which is shown to be critically beneficial to recognition performance but largely ignored by the self-attention. Five versions different from model scaling are introduced based on the proposed VOLO: VOLO-D1 with 27M parameters to VOLO-D5 with 296M. Experiments show that the best one, VOLO-D5, achieves 87.1% top-1 accuracy on ImageNet-1K classification, which is the first model exceeding 87% accuracy on this competitive benchmark, without using any extra training data.
Figure 1. Illustration of outlook attention. [1]
Our reproduced model performance on ImageNet-1K is reported as follows.
| Model | Context | Top-1 (%) | Top-5 (%) | Params (M) | Recipe | Weight |
|---|---|---|---|---|---|---|
| volo_d1 | D910x8-G | 82.59 | 95.99 | 27 | yaml | weights |
| volo_d2 | D910x8-G | 82.95 | 96.13 | 59 | yaml | weights |
| volo_d3 | D910x8-G | 83.38 | 96.28 | 87 | yaml | weights |
| volo_d4 | D910x8-G | 82.5 | 95.86 | 193 | yaml | weights |
- Context: Training context denoted as {device}x{pieces}-{MS mode}, where mindspore mode can be G - graph mode or F - pynative mode with ms function. For example, D910x8-G is for training on 8 pieces of Ascend 910 NPU using graph mode.
- Top-1 and Top-5: Accuracy reported on the validation set of ImageNet-1K.
Please refer to the installation instruction in MindCV.
Please download the ImageNet-1K dataset for model training and validation.
- Distributed Training
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/volo/volo_d1_ascend.yaml --data_dir /path/to/imagenetIf the script is executed by the root user, the
--allow-run-as-rootparameter 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.
- Standalone Training
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/volo/volo_d1_ascend.yaml --data_dir /path/to/dataset --distribute FalseTo 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/volo/volo_d1_ascend.yaml --data_dir /path/to/imagenet --ckpt_path /path/to/ckptTo deploy online inference services with the trained model efficiently, please refer to the deployment tutorial.
[1] Yuan L , Hou Q , Jiang Z , et al. VOLO: Vision Outlooker for Visual Recognition[J]. . arXiv preprint arXiv:2106.13112, 2021.