Microsoft's ResNet-18 Hits 840K Monthly Downloads Despite Being a Decade Old
A decade-old 11.7M parameter image classifier still pulls close to a million monthly downloads, and the reasons tell you a lot about modern vision pipelines.
Takeaways
- microsoft/resnet-18 keeps trending with 840K+ downloads despite being nearly a decade old.
- 11.7M parameters, 1.8 GFLOPs, roughly 69.8% top-1 and 89.1% top-5 on ImageNet-1k.
- Apache 2.0 licensed, loads in two lines via the Transformers
image-classificationpipeline. - Residual connections introduced in the original paper now power transformers and diffusion models.
- INT8 quantization drops accuracy by about 0.1%, making it ideal for edge and on-device inference.
- Serves as a go-to backbone, teaching tool, and robotics vision encoder, including in lerobot tutorials.
A recent Hugging Face snapshot placed microsoft/resnet-18 among trending vision models, with roughly 840,000 downloads during the preceding month and references across dozens of Spaces. ResNet-18 also appears in robot-learning examples such as LeRobot. Hub downloads count file requests rather than unique deployments, but the activity shows continued demand for a checkpoint first associated with the 2015 ImageNet competition.
The appeal is practical: ResNet-18 is compact, stable, widely supported, and easy to fine-tune. It gives developers a credible baseline without the memory, latency, and integration costs of larger convolutional networks or vision transformers.
Small model, explicit tradeoffs
The ResNet-18 name counts 17 main-path convolutional layers and one classifier layer. Projection shortcuts add 1x1 convolutions where feature dimensions change, so implementations contain more convolution modules than the name suggests.
| Metric | Reference value | Practical meaning |
|---|---|---|
| Parameters | 11,689,512 | About 45 MiB of raw FP32 weights |
| Input | 224 × 224 RGB | Modest memory use and predictable preprocessing |
| Compute | About 1.81 GFLOPs | Suitable for CPU, edge, and real-time evaluation |
| ImageNet-1k top-1 | 69.758% | A solid baseline rather than current best accuracy |
| ImageNet-1k top-5 | 89.078% | The correct label appears among five predictions in most validation examples |
An INT8 export can reduce weight storage by roughly four times, although the exact size, accuracy, and latency depend on the runtime. Convolution-heavy networks usually benefit more from static post-training quantization or quantization-aware training than from dynamic quantization, which primarily targets linear and recurrent layers. Teams should benchmark the exported artifact on the intended processor because FLOPs alone do not predict latency.