Download a public baseline model

Choose a pretrained baseline

Download an existing quantized export rather than retraining or calibrating a model. Use mobilenet_v1_0.25_224_quant.tflite from the TensorFlow MobileNet V1 model catalog . It’s a public pretrained image classifier with a width multiplier of 0.25 and a fixed input shape of [1, 224, 224, 3].

The model has quantized eight-bit weights and activations, with unsigned eight-bit input and output tensors. It isn’t a neural upscaling model. However, it provides a real convolutional workload for learning the Arm ML Inference Advisor (MLIA) analysis workflow.

Download and verify the model

From mlia-demo, download the versioned public archive, and extract the LiteRT model. Name your local copy baseline.tflite:

    

        
        
mkdir -p models runs
curl -fL https://storage.googleapis.com/download.tensorflow.org/models/mobilenet_v1_2018_08_02/mobilenet_v1_0.25_224_quant.tgz -o models/baseline.tgz
tar -xzf models/baseline.tgz -C models ./mobilenet_v1_0.25_224_quant.tflite
mv models/mobilenet_v1_0.25_224_quant.tflite models/baseline.tflite
echo "ce8f61fe3e4e29560c4ea8038e4b91eb354be330fc428348b9645bd1e7de03bf  models/baseline.tflite" | sha256sum --check

    

The expected output is:

    

        
        models/baseline.tflite: OK

        
    

MLIA runs the installed converter and saves the intermediate artifacts with its reports. Keep the downloaded model in your working directory.

What you’ve accomplished and what’s next

You’ve downloaded and verified the quantized MobileNet baseline.

Next, you’ll check whether its operators are supported by passing models/baseline.tflite directly to MLIA.