timm/maxvit_tiny_tf_224.in1k - Ascend NPU Adaptation
1. Model Overview
MaxViT-Tiny TF 224 (TensorFlow-style) vision transformer
- Original: timm/maxvit_tiny_tf_224.in1k
- Framework: PyTorch (timm)
- Task: Image Classification (ImageNet-1K)
- Parameters: 30,916,528
- Status: SUCCESS
2. Environment
| Component | Version |
|---|
| torch | 2.9.0 |
| torch-npu | 2.9.0.post1 |
| CANN | 8.5.1 |
| NPU | Ascend 910B4 |
3. Accuracy
| Metric | Value |
|---|
| Cosine Similarity | 1.000000 |
| Max Abs Error | 0.003341 |
| CPU/NPU Prediction | Match (class 111) |
| Metric | Value |
|---|
| Avg Latency | 22.88 ms |
| Output Shape | [1, 1000] |
5. CPU vs NPU
| Metric | CPU | NPU | Diff |
|---|
| Cosine Similarity | 1.0 | 1.000000 | - |
| Output Shape | [1,1000] | [1,1000] | Same |
| Prediction | 111 | 111 | Same |