weixin_72661020/convnextv2-tiny-22k-224
Model Introduction
Files & versions
Pull Requests
Discussion
Analysis

ConvNeXtV2-Tiny-22k-224

This model is a tiny variant of ConvNeXtV2 architecture, pretrained on ImageNet-22K and fine-tuned on ImageNet-1K (224x224 resolution).

Model parameters: 28.6M Input size: 3x224x224

Model Description

ConvNeXtV2 is an improved convolutional neural network architecture that incorporates techniques from vision transformers (e.g., LayerNorm, GELU, large kernels) while maintaining the efficiency of pure ConvNets. The V2 version introduces ConvNeXtV2Block with GRN (Global Response Normalization) for better performance.

Usage

from PIL import Image
import requests
from transformers import AutoImageProcessor, ConvNextV2ForImageClassification

model_path = "weixin_72661020/convnextv2-tiny-22k-224"
processor = AutoImageProcessor.from_pretrained(model_path)
model = ConvNextV2ForImageClassification.from_pretrained(model_path)

# Example inference
url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = Image.open(requests.get(url, stream=True).raw)
inputs = processor(image, return_tensors="pt")

with torch.no_grad():
    outputs = model(**inputs)
    logits = outputs.logits
    predicted_class = logits.argmax(-1).item()
print(f"Predicted class: {predicted_class}")

NPU Performance (Ascend910)

Batch SizeMean Latency (ms)Throughput (fps)
17.78128.6
49.55418.7
816.57482.7

Accuracy

Top-1 consistency between CPU and NPU: 100% KL Divergence (max): < 0.00001

Files

  • pytorch_model.bin / model.safetensors: Model weights
  • config.json: Model configuration
  • preprocessor_config.json: Image processor configuration