EfficientNet-B0 RA Hits 800K Downloads as Developers Chase Tiny 5M Models
A tiny 5.3M parameter EfficientNet-B0 in timm keeps racking up nearly a million monthly downloads thanks to its modern RandAugment training recipe.
Takeaways
- timm's efficientnet_b0.ra_in1k pulls ~800K monthly downloads as a go-to lightweight CNN backbone.
- 5.3M params, 0.4 GMACs, 224x224 input, Apache 2.0 licensed, safetensors format.
- Retrained with RandAugment RA recipe, RMSProp, EMA weights, step LR schedule with warmup.
- Recipe published as B in ResNet Strikes Back, applied to EfficientNet-B0.
- Supports classification, multi-scale feature extraction, and 1280-dim embeddings via timm API.
- Best for edge, mobile, and transfer learning where budget beats raw accuracy.
EfficientNet-B0 RA recorded more than 800,000 Hugging Face downloads in a recent month. The checkpoint pairs the compact EfficientNet-B0 convolutional architecture with a newer timm training recipe, giving developers a 5.3 million-parameter ImageNet classifier and feature backbone for constrained inference workloads.
Ross Wightman’s timm repository provides the weights. The model name describes the package: efficientnet_b0 identifies the architecture, ra identifies the RandAugment-based recipe, and in1k indicates training on the 1,000-class ImageNet-1k dataset.
Inside the .ra_in1k recipe
The underlying network follows the EfficientNet paper by Mingxing Tan and Quoc V. Le. EfficientNet scales network depth, width, and input resolution together, producing mobile-friendly convolutional models with relatively low parameter and compute requirements.
This checkpoint uses a timm recipe derived from RandAugment experiments and the “B” recipe described in ResNet Strikes Back. Its training stack combines RandAugment, RMSProp, learning-rate warmup, a step-based learning-rate schedule, and an exponential moving average of model weights.
RandAugment samples transformations such as rotation, shear, and color adjustment during training. Exponential moving averaging maintains a smoothed copy of the learned weights. Together, these techniques improve generalization without adding layers or increasing inference cost.