Swin Transformer is a pioneering neural network architecture that introduces a novel approach to handling local and global information in computer vision tasks. Departing from traditional self-attention mechanisms, Swin Transformer adopts a hierarchical design, organizing its attention windows in a shifted manner. This innovation enables more efficient modeling of contextual information across different scales, enhancing the model's capability to capture intricate patterns.
| GPU | IXUCA SDK | Release | Branch |
|---|---|---|---|
| MR-V100 | 4.4.0 | 26.03 | release/26.03 |
| MR-V100 | 4.3.0 | 25.12 | release/25.12 |
Note: 请切换到与您的 SDK 版本对应的 Release 分支进行测试。请勿直接在 master 分支上运行测试，因为 master 分支可能包含与您的本地 SDK 版本不兼容的最新更改。
切换分支命令示例：
git checkout release/26.03
Pretrained model: http://files.deepspark.org.cn:880/deepspark/data/checkpoints/swin_s_model_sim.onnx
Dataset: https://www.image-net.org/download.php to download the validation dataset.
mkdir -p checkpoints
# download swin_s_model_sim.onnx into checkpointsexport PROJ_DIR=./
export DATASETS_DIR=./imagenet-val/
export CHECKPOINTS_DIR=./checkpoints
export RUN_DIR=../../ixrt_common/# Accuracy
bash scripts/infer_swin_transformer_fp16_accuracy.sh
# Performance
bash scripts/infer_swin_transformer_fp16_performance.sh| Model | BatchSize | Precision | FPS | Top-1(%) | Top-5(%) |
|---|---|---|---|---|---|
| Swin Transformer | 32 | FP16 | 231.428 | 82.782 | 96.296 |