MobileNetV3-Large-075 is a width-reduced version of MobileNetV3-Large, using 0.75x width multiplier to reduce channels while maintaining the Large series network depth. Trained on ImageNet-1K, input size 224x224.
Original model: timm/tf_mobilenetv3_large_075.in1k
| Component | Version |
|---|---|
| NPU | Ascend910 |
| CANN | 25.5.2 |
| PyTorch | 2.9.0 |
| torch_npu | 2.9.0 |
| timm | 1.0.27 |
| Python | 3.11.14 |
| OS | Linux (aarch64) |
pip install torch torchvision timm Pillow numpy modelscope safetensors
export MODELSCOPE_CACHE=/tmp/modelscope# CPU inference
python3 inference.py --model_name tf_mobilenetv3_large_075.in1k --device cpu --output_dir ./tf_mobilenetv3_large_075.in1k
# NPU inference
python3 inference.py --model_name tf_mobilenetv3_large_075.in1k --device npu --output_dir ./tf_mobilenetv3_large_075.in1k| Metric | CPU | NPU (Ascend910) |
|---|---|---|
| Input size | 224x224 | 224x224 |
| Avg inference | 24.18 ms | 5.83 ms |
| Top-1 class | class 549 | class 549 |
| Top-1 prob | 7.4729% | 7.4055% |
NPU speedup: 4.1x vs CPU.
| Metric | Value |
|---|---|
| Max logits diff | 0.01664400 |
| Mean logits diff | 0.00292952 |
| Cosine Similarity | 0.999994 |
| Max prob diff | 0.067369% |
| Mean prob diff | 0.000392% |
| Relative L2 error | 0.335626% |
| Top-1 match | Yes |
| Top-5 overlap | 5/5 |
Top-5 predictions:
| Rank | CPU class | CPU prob | NPU class | NPU prob |
|---|---|---|---|---|
| 1 | class 549 | 7.4729% | class 549 | 7.4055% |
| 2 | class 902 | 4.7883% | class 902 | 4.7955% |
| 3 | class 916 | 3.7425% | class 916 | 3.7318% |
| 4 | class 446 | 2.7991% | class 446 | 2.7828% |
| 5 | class 409 | 2.7624% | class 409 | 2.7593% |
NPU vs CPU error < 1%. All metrics show NPU results match CPU:

import torch
import timm
from PIL import Image
from timm.data import create_transform, resolve_data_config
model = timm.create_model('tf_mobilenetv3_large_075.in1k', pretrained=True)
model.eval()
if torch.npu.is_available():
model = model.npu()
transform = create_transform(**resolve_data_config(model.pretrained_cfg, model=model))
img = Image.open('image.jpg').convert('RGB')
input_tensor = transform(img).unsqueeze(0)
if torch.npu.is_available():
input_tensor = input_tensor.npu()
with torch.no_grad():
logits = model(input_tensor)
probs = torch.nn.functional.softmax(logits, dim=-1)
top5 = torch.topk(probs, k=5)
for i in range(5):
print(f"class {top5.indices[0][i].item()}: {top5.values[0][i].item()*100:.2f}%")