MobileNetV2_TF2.md

TensorFlow MobileNetV2 with TensorFlow 2.4

Setup AI Model Efficiency Toolkit (AIMET)

Please install and setup AIMET before proceeding further. This evaluation was run using AIMET 1.26 for TensorFlow 2.4 i.e. please set release_tag="1.26" and AIMET_VARIANT="tf_gpu" in the above instructions.

Additional Dependencies

pip install numpy==1.19.5

Model checkpoint and dataset

The TF2 pretrained mobilenetv2 model is directly imported from package tensorflow.keras.applications

Dataset

• ImageNet can be downloaded from here:
  • http://www.image-net.org/
• The directory where the data is located should contains subdirectories, each containing images for a class
• The ImageNet validation dataset should be organized in the following way

< path to ImageNet validation dataset >
├── n01440764
├── n01443537
├── ...

Usage

• To run evaluation with QuantSim in AIMET, use the following:

python aimet_zoo_tensorflow/mobilenet_v2_tf2/evaluators/mobilenet_v2_tf2_quanteval.py \
    --dataset-path <path to imagenet dataset> \
    --batch-size <batch size for loading the dataset> \
    --model-config <model configuration to test>

Available model configurations are:

• mobilenetv2_w8a8

• Example: python aimet_zoo_tensorflow/mobilenet_v2_tf2/evaluators/mobilenet_v2_tf2_quanteval.py --dataset-path <imagenet_dataset_path> --batch_size 4 --model-config mobilenetv2_w8a8

Quantization configuration

In the evaluation script included, we have used the default config file, which configures the quantizer ops with the following assumptions:

• Weight quantization: 8 bits, symmetric quantization
• Bias parameters are not quantized
• Activation quantization: 8 bits, asymmetric quantization
• Model inputs are quantized

Results

Below are the top1 accuracy results of the TensorFlow 2.4 mobilenetv2 model for the imagenet dataset:

Model Configuration	Top1 (%)
Mobilenet_V2_FP32	71.6
Mobilenet_V2 + simple PTQ(w8a8)	71.0