MobileNetV2.md

Tensorflow Mobilenetv2 1.4

Setup AI Model Efficiency Toolkit (AIMET)

Please install and setup AIMET before proceeding further. This evaluation was run using AIMET 1.22.2 for TensorFlow 1.15 i.e. please set release_tag="1.22.2" and AIMET_VARIANT="tf_gpu_tf115" in the above instructions.

NOTE: This model is expected not to work with GPUs at or after NVIDIA 30-series (e.g. RTX 3050), as those bring a new architecture not fully compatible with TF 1.X.

Additional Dependencies

Setup TensorFlow Models repo

• Clone the TensorFlow Models repo
  git clone https://github.com/tensorflow/models.git
cd models
• Checkout this commit id:
  git checkout 104488e40bc2e60114ec0212e4e763b08015ef97
• Append the repo location to your PYTHONPATH with the following:
  export PYTHONPATH=$PYTHONPATH:<path to tensorflow/models repo>/research/slim

Model checkpoint and dataset

• Downloading model checkpoint and config file are handled by evaluation script.
• The optimized Mobilenet v2 1.4 checkpoint can be downloaded from Releases.
• The Quantization Simulation (Quantsim) Configuration file can be downloaded from here: default_config.json (Please see this page for more information on this file).

Dataset

• ImageNet can be downloaded from here:
  • http://www.image-net.org/
• For this evaluation, Tf-records of ImageNet validation dataset are required. (See https://github.com/tensorflow/models/tree/master/research/slim#Data for details)
• The Tf-records of ImageNet validation dataset should be organized in the following way

< path to ImageNet validation dataset Tf-records>
├── validation-00000-of-00128
├── validation-00001-of-00128
├── ...

Usage

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

python mobilenet_v2_140_quanteval.py \
    --dataset-path <path to imagenet validation TFRecords> \
    --batch-size <batch size for loading the dataset> \
    --model-to-eval <which model to evaluate. Two options are available: 'fp32' for evaluating original fp32 model, 'int8' for evaluating quantized int8 model.>

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, asymmetric quantization
• Bias parameters are not quantized
• Activation quantization: 8 bits, asymmetric quantization
• Model inputs are quantized
• Operations which shuffle data such as reshape or transpose do not require additional quantizers