Computer Vision

ResNet10 ImageNet Classification on the MLA-100 NPU

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

ModelPrecisionProduct NameTrainVal
ResNet10-baseFP32NVIDIA A10084.00%69.50%
ResNet10-2FP32NVIDIA A10081.00%72.00%
ResNet10-3FP32NVIDIA A10083.27%71.00%
ResNet10-4FP32NVIDIA L492.11%82.9%
ResNet10-5FP32NVIDIA L484.27%86.30%
ResNet10-2INT8 (PTQ)Mobilint MLA100xx
ResNet10-5-calibratedv3INT8 (PTQ)Mobilint MLA100x85.30%

Processing Performance

ModelProduct NamePerformance (ms/image)Images Per Second (img/sec)Power (W)
ResNet10-5Intel(R) AI Boost (Ultra 5 255)2.86349.865
ResNet10-5MLA1001.07938.122

ResNet10

IDNotebookHuggingfaceReport
ResNet10-baseNotebookbdanko/imagenetsub20/imagenetsub20resnet10.pthReport
ResNet10-2Notebookbdanko/imagenetsub20/imagenetsub20resnet10-2.pthReport
ResNet10-3NotebookN/AReport
ResNet10-4Notebookbdanko/imagenetsub20/imagenetsub20resnet10-4.pthReport
ResNet10-5Notebookbdanko/imagenetsub20/imagenetsub20resnet10-5.pthReport

Training Configs

IDCheckpointEpochsOptimLRBatch
ResNet10-baseNone52Adam0.0011024
ResNet10-2ResNet10-base14Adam0.0011024
ResNet10-3ResNet10-base14Adam0.0011024
ResNet10-4None93Adam/Cosine Annealing0.00164
ResNet10-5None163Adam/Cosine Annealing/EMA0.00164

Quantization

Models were quantized using the 100 samples in calib_npy as calibration.