xujiashuai/timm-swinv2_cr_small_ns_224.sw_in1k
Model Introduction
Files & versions
Pull Requests
Discussion
Analysis

timm/swinv2_cr_small_ns_224.sw_in1k - Ascend NPU Adaptation

1. Model Overview

SwinV2-CR Small is a Compact Restructured Swin Transformer V2 model trained with neighborhood attention on ImageNet-1k at 224x224 resolution.

  • Original: timm/swinv2_cr_small_ns_224.sw_in1k
  • Architecture: SwinV2-CR (Compact Restructured)
  • Parameters: 49,696,444
  • Input Resolution: 224x224
  • Output Classes: 1000

2. NPU Adaptation

  • Device: Ascend NPU (npu:0)
  • Framework: PyTorch + torch_npu

3. Accuracy

MetricValue
Cosine Similarity0.999986
Max Abs Error0.001687

4. Latency

DeviceAvg Latency
NPU (npu:0)22.31 ms

5. CPU vs NPU

MetricCPUNPU
Output Shape[1, 1000][1, 1000]
Prediction640640

6. Usage

source setup_env.sh
HF_ENDPOINT=https://hf-mirror.com python3 inference.py --device npu:0