Hugging Face Adds RepVGG-A0 to timm With 9M Parameters and One-Line Deployment
The compact RepVGG-A0 checkpoint is live on the Hugging Face Hub via timm, offering a 9M-parameter plain-conv backbone trained on ImageNet-1k.
- RepVGG-A0 pretrained weights are now available on Hugging Face via timm.
- 9.1M params, 1.5 GMACs, 224x224 input, trained on ImageNet-1k, MIT licensed.
- Inference graph is pure 3x3 conv plus ReLU after structural re-parameterization.
- Multi-branch training blocks collapse into a single 3x3 conv via BN folding.
- Loadable in one line through
timm.create_modelor Transformers pipeline. - Underpins the backbone design used in YOLOv6 and YOLOv7.
RepVGG-A0 joins Hugging Face’s timm catalog
Hugging Face now hosts a RepVGG-A0 checkpoint that timm can download and instantiate with one create_model call. The 9.1-million-parameter weights came from the paper’s authors and were trained for 1,000-class ImageNet classification on 224 × 224 RGB images.
RepVGG addresses a practical hardware constraint. Multi-branch blocks optimize effectively during training, and homogeneous stacks of 3 × 3 convolutions map efficiently to common inference kernels. Structural reparameterization combines those properties by training with several branches and later folding each block into a single convolution.
Nine million parameters, one standard API
| Property | Value |
|---|---|
| Parameters | 9.1 million |
| Compute | 1.5 GMACs per 224 × 224 image |
| Activations | 3.6 million |
| Input | 3 × 224 × 224 |
| Classifier output | 1,000 ImageNet logits |
| Checkpoint source | Original RepVGG authors |
| License | MIT |
The timm implementation uses its configurable BYOBNet implementation. That scaffold supports custom stages, stems, normalization layers, output strides, gradient checkpointing, stochastic depth, classifier removal, and per-stage feature extraction without changing the checkpoint-loading interface.
Load the saved recipe
Installing current versions of timm and the Hub client provides the model loader and checkpoint download path:
python -m pip install -U timm huggingface_hubThe model’s saved data configuration supplies its resize, crop, interpolation, normalization, and input-size settings:
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
