Computer Vision
ResNet10 ImageNet Classification on the MLA-100 NPU
Training, compiling and quantization notebooks
Bence Danko
Last updated March 1, 2026
Prediction of 20 classes from ImageNet, using 6,000 images for training, and 1,000 for evaluation. Each class equally has 300 images for training, and 50 to validate. We must omit annotation data, working only off of this limited subset, and achieve at least 60% on validation accuracy.
Dataset and Models.
The dataset used can be located at https://huggingface.co/datasets/bdanko/imagenetsubset20. All models are published at https://huggingface.co/bdanko/imagenetsub20.
Supported Operations
The final model can only support operations from the MLA-100 qb compiler.
Summary
| Model | Precision | Product Name | Train | Val |
|---|---|---|---|---|
| ResNet10-base | FP32 | NVIDIA A100 | 84.00% | 69.50% |
| ResNet10-2 | FP32 | NVIDIA A100 | 81.00% | 72.00% |
| ResNet10-3 | FP32 | NVIDIA A100 | 83.27% | 71.00% |
| ResNet10-4 | FP32 | NVIDIA L4 | 92.11% | 82.9% |
| ResNet10-5 | FP32 | NVIDIA L4 | 84.27% | 86.30% |
| ResNet10-2 | INT8 (PTQ) | Mobilint MLA100 | x | x |
| ResNet10-5-calibratedv3 | INT8 (PTQ) | Mobilint MLA100 | x | 85.30% |
Processing Performance
| Model | Product Name | Performance (ms/image) | Images Per Second (img/sec) | Power (W) |
|---|---|---|---|---|
| ResNet10-5 | Intel(R) AI Boost (Ultra 5 255) | 2.86 | 349.8 | 65 |
| ResNet10-5 | MLA100 | 1.07 | 938.1 | 22 |
ResNet10
| ID | Notebook | Huggingface | Report |
|---|---|---|---|
| ResNet10-base | Notebook | bdanko/imagenetsub20/imagenetsub20resnet10.pth | Report |
| ResNet10-2 | Notebook | bdanko/imagenetsub20/imagenetsub20resnet10-2.pth | Report |
| ResNet10-3 | Notebook | N/A | Report |
| ResNet10-4 | Notebook | bdanko/imagenetsub20/imagenetsub20resnet10-4.pth | Report |
| ResNet10-5 | Notebook | bdanko/imagenetsub20/imagenetsub20resnet10-5.pth | Report |
Training Configs
| ID | Checkpoint | Epochs | Optim | LR | Batch |
|---|---|---|---|---|---|
| ResNet10-base | None | 52 | Adam | 0.001 | 1024 |
| ResNet10-2 | ResNet10-base | 14 | Adam | 0.001 | 1024 |
| ResNet10-3 | ResNet10-base | 14 | Adam | 0.001 | 1024 |
| ResNet10-4 | None | 93 | Adam/Cosine Annealing | 0.001 | 64 |
| ResNet10-5 | None | 163 | Adam/Cosine Annealing/EMA | 0.001 | 64 |
Quantization
Models were quantized using the 100 samples in calib_npy as calibration.