TexasInstruments-EdgeAI/ConvNeXt-Classification

ConvNeXt for TI EdgeAI

A ConvNet for the 2020s — Pure ConvNet Matching Transformer Accuracy

---

Overview

ConvNeXt is a pure convolutional network modernized by incorporating design principles from Vision Transformers (Swin Transformer). Introduced in A ConvNet for the 2020s (Liu et al., CVPR 2022), ConvNeXt matches or surpasses Swin Transformers in accuracy while retaining the simplicity, efficiency, and hardware-friendliness of standard CNNs — no attention mechanisms, no positional encodings.

Pretrained weights are sourced from torchvision (BSD-3-Clause), trained on ImageNet-1K using a modernized training recipe.

All variants take a 224×224 input with ImageNet normalization (mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]). The ideal pre-crop resize is variant-specific (236px for Tiny, 230px for Small, 232px for Base/Large) and is already set correctly in each variant's config YAML.

---

Model Variants

Model	Architecture	Params	GFLOPs	Top-1 Acc	Top-5 Acc	Validated Devices	Config
convnext_tiny	ConvNeXt-Tiny	28.6M	4.46	82.52%	96.15%	TDA4VH	convnext_tiny_config.yaml
convnext_small	ConvNeXt-Small	50.2M	8.68	83.62%	96.65%	TDA4VH	convnext_small_config.yaml
convnext_base	ConvNeXt-Base	88.6M	15.36	84.06%	96.87%	TDA4VH	convnext_base_config.yaml
convnext_large	ConvNeXt-Large	197.8M	34.36	84.41%	96.98%	TDA4VH	convnext_large_config.yaml

Recommended for edge deployment: convnext_tiny delivers competitive accuracy (82.5%) at the lowest compute (4.46 GFLOPs, 28.6M params), making it the most practical choice for edge deployment. Larger variants offer incremental accuracy gains at significantly higher compute cost.

---

Quick Start

Prerequisites

pip install torch torchvision onnx>=1.22.0 onnxruntime>=1.23.2
# Optional but recommended for model optimization:
pip install onnx-simplifier

Export the Model

# Export the default model (convnext_tiny)
python prepare_model.py

# Export a specific model variant
python prepare_model.py --model convnext_base

# Export all supported models
python prepare_model.py --model all

# List all available variants
python prepare_model.py --list-models

# Use a custom checkpoint
python prepare_model.py --model convnext_tiny --weights /path/to/checkpoint.pth

The script automatically:

• Downloads pretrained ImageNet-1K weights from torchvision (first run only)
• Exports the model to ONNX (opset 17) with static input shape [1, 3, 224, 224]
• Runs ONNX shape inference
• Optionally simplifies the graph with onnx-simplifier

Compile and Infer uing edgeai-tidlrunner

Note: Run the commands below from inside the tidlrunner directory (the cloned edgeai-tidlrunner repository), with --config_path pointing to this model's config file.

Compile using edgeai-tidlrunner - on PC

cd /path/to/edgeai-tidlrunner
tidlrunner-cli compile --target_device J784S4 \
  --config_path /path/to/convnext_tiny_config.yaml

Run Inference Benchmark - on device

cd /path/to/edgeai-tidlrunner
tidlrunner-cli infer --target_device J784S4 \
  --config_path /path/to/convnext_tiny_config.yaml

Compile and Infer using edgeai-tidl-tools (Advanced):

Follow the instructions at https://github.com/TexasInstruments/edgeai-tidl-tools

Deploy using edgeai-tidl-tools:

Deplyment can be done using edgeai-tidl-tools. For ONNX models, onnxruntime-tidl with TIDL acceleration can be used. Consult the documentation of edgeai-tidl-tools for more details.

---

Citation

@inproceedings{liu2022convnet,
  title     = {A ConvNet for the 2020s},
  author    = {Zhuang Liu and Hanzi Mao and Chao-Yuan Wu and
               Christoph Feichtenhofer and Trevor Darrell and Saining Xie},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision
               and Pattern Recognition (CVPR)},
  year      = {2022},
  url       = {https://arxiv.org/abs/2201.03545}
}

---

🔗 Resources

Resource	Link
Paper	arXiv:2201.03545
Source Repo	facebookresearch/ConvNeXt
Torchvision Docs	ConvNeXt
HuggingFace	facebook/convnext-tiny-224
edgeai-tidl-tools	GitHub
edgeai-tidlrunner	GitHub
EdgeAI SDK	Documentation

---

Related Models

ViT Pure Transformer Attention-based

DINOv2 Self-supervised ViT Higher accuracy

DINO Self-supervised ViT Linear head

ResNet CNN baseline Lower compute

---

Maintained by: Texas Instruments EdgeAI Team
Last Updated: August 2026