twins_pcpvt_small

Converted TIMM image classification model for LiteRT.

• Source architecture: twins_pcpvt_small
• Source checkpoint: timm/twins_pcpvt_small.in1k
• File: model.tflite
• Input: float32 tensor in NCHW layout, shape [1, 3, 224, 224]
• Output: ImageNet-1K logits, shape [1, 1000]

Runtime Status

• CPU smoke test: passed with LiteRT CompiledModel.
• GPU delegation: currently blocked for this model by rank-5 tensor patterns in the GPU backend, mostly RESHAPE, TRANSPOSE, and related window/attention operations. The model is published as CPU-ready while GPU support is being improved.

Model Details

• Model Type: Image classification / feature backbone
• Model Stats:
  • Params (M): 24.1
  • GMACs: 3.8
  • Activations (M): 18.1
  • Image size: 224 x 224
• Papers:
  • Twins: Revisiting the Design of Spatial Attention in Vision Transformers: https://arxiv.org/abs/2104.13840
• Dataset: ImageNet-1k
• Original: https://github.com/Meituan-AutoML/Twins

Citation

@inproceedings{chu2021Twins,
    title={Twins: Revisiting the Design of Spatial Attention in Vision Transformers},
    author={Xiangxiang Chu and Zhi Tian and Yuqing Wang and Bo Zhang and Haibing Ren and Xiaolin Wei and Huaxia Xia and Chunhua Shen},
    booktitle={NeurIPS 2021},
    url={https://openreview.net/forum?id=5kTlVBkzSRx},
    year={2021}
}