Identity Mappings in Deep Residual Networks
Paper
•
1603.05027
•
Published
•
4
PreResNet
Use case : Image classification
Model description
Pre-activation ResNet (PreResNet) is a variant of ResNet that places batch normalization and activation before convolutions. This simple change improves both training dynamics and final accuracy.
PreResNet employs a pre-activation design with BN-ReLU-Conv order instead of Conv-BN-ReLU, enabling cleaner identity mappings for improved information flow through residual connections. The improved gradient flow during training results from full pre-activation applied to both main path and shortcut connections.
The architecture is well-suited for deep learning research, transfer learning with pre-activation benefits, and applications where training dynamics matter.
(source: https://arxiv.org/abs/1603.05027)
The model is quantized to int8 using ONNX Runtime and exported for efficient deployment.
Network information
| Network Information | Value |
|---|---|
| Framework | Torch |
| MParams | ~3.75 M |
| Quantization | Int8 |
| Provenance | https://github.com/KaimingHe/resnet-1k-layers |
| Paper | https://arxiv.org/abs/1603.05027 |
Network inputs / outputs
For an image resolution of NxM and P classes
| Input Shape | Description |
|---|---|
| (1, N, M, 3) | Single NxM RGB image with UINT8 values between 0 and 255 |
| Output Shape | Description |
|---|---|
| (1, P) | Per-class confidence for P classes in FLOAT32 |
Recommended platforms
| Platform | Supported | Recommended |
|---|---|---|
| STM32L0 | [] | [] |
| STM32L4 | [] | [] |
| STM32U5 | [] | [] |
| STM32H7 | [] | [] |
| STM32MP1 | [] | [] |
| STM32MP2 | [] | [] |
| STM32N6 | [x] | [x] |
Performances
Metrics
- Measures are done with default STEdgeAI Core configuration with enabled input / output allocated option.
- All the models are trained from scratch on Imagenet dataset
Reference NPU memory footprint on Imagenet dataset (see Accuracy for details on dataset)
| Model | Dataset | Format | Resolution | Series | Internal RAM (KiB) | External RAM (KiB) | Weights Flash (KiB) | STEdgeAI Core version |
|---|---|---|---|---|---|---|---|---|
| preresnet18_a025_pt_224 | Imagenet | Int8 | 224×224×3 | STM32N6 | 1323 | 0 | 3843.64 | 3.0.0 |
Reference NPU inference time on Imagenet dataset (see Accuracy for details on dataset)
| Model | Dataset | Format | Resolution | Board | Execution Engine | Inference time (ms) | Inf / sec | STEdgeAI Core version |
|---|---|---|---|---|---|---|---|---|
| preresnet18_a025_pt_224 | Imagenet | Int8 | 224×224×3 | STM32N6570-DK | NPU/MCU | 14.35 | 69.69 | 3.0.0 |
Accuracy with Imagenet dataset
| Model | Format | Resolution | Top 1 Accuracy |
|---|---|---|---|
| preresnet18_a025_pt | Float | 224x224x3 | 60.99 % |
| preresnet18_a025_pt | Int8 | 224x224x3 | 59.79 % |
Dataset details: link Number of classes: 1000. To perform the quantization, we calibrated the activations with a random subset of the training set. For the sake of simplicity, the accuracy reported here was estimated on the 50000 labelled images of the validation set.
| Model | Format | Resolution | Top 1 Accuracy |
|---|---|---|---|
| preresnet18_a025_pt | Float | 224x224x3 | 60.99 % |
| preresnet18_a025_pt | Int8 | 224x224x3 | 59.79 % |
Retraining and Integration in a simple example:
Please refer to the stm32ai-modelzoo-services GitHub here
References
[1] - Dataset: Imagenet (ILSVRC 2012) — https://www.image-net.org/
[2] - Model: PreResNet — https://github.com/KaimingHe/resnet-1k-layers