metadata
library_name: pytorch
EfficientNet-Lite is a family of mobile and IoT-friendly image classification models derived from EfficientNet and optimized for efficient inference on mobile CPUs, GPUs, and Edge TPUs, with modifications that improve quantization and hardware compatibility.
Original paper: EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks
EfficientNet-Lite0
This model uses the EfficientNet-Lite0 architecture, which removes squeeze-and-excitation modules, replaces Swish with ReLU6, and keeps the stem and head fixed across scaled variants to reduce latency and improve post-training quantization.
Model Configuration:
- Reference implementation: timm.models.efficientnet_lite0
- Original Weight: efficientnet_lite0.ra_in1k
- Dataset: ImageNet
- Resolution: 3x224x224
- Support Cooper version:
- Cooper SDK: [2.5.4]
- Cooper Foundry: [2.3]
| Model | Device | compression | Model Link |
|---|---|---|---|
| EfficientNet-Lite0 | N1-655 | Activation_fp16 | Model_Link |
| EfficientNet-Lite0 | N1-655 | Amba_optimized | Model_Link |
| EfficientNet-Lite0 | X7 | Activation_fp16 | Model_Link |
| EfficientNet-Lite0 | X7 | Amba_optimized | Model_Link |
| EfficientNet-Lite0 | CV7 | Activation_fp16 | Model_Link |
| EfficientNet-Lite0 | CV7 | Amba_optimized | Model_Link |
| EfficientNet-Lite0 | CV72 | Activation_fp16 | Model_Link |
| EfficientNet-Lite0 | CV72 | Amba_optimized | Model_Link |
| EfficientNet-Lite0 | CV75 | Activation_fp16 | Model_Link |
| EfficientNet-Lite0 | CV75 | Amba_optimized | Model_Link |