Image Classification STM32 Model Quantization

All the sections of the YAML file must be set as described in this readme with the operation_mode set to quantization. Alternatively, you can specify the operation_mode later when running the experience by using python stm32ai_main.py operation_mode=quantization.

1.1 Prepare the dataset

Information about the dataset you want to use for activations calibration is provided in the dataset section of the configuration file, as shown in the YAML code below.

dataset:
  dataset_name: tf_flowers 
  class_names: [daisy, dandelion, roses, sunflowers, tulips] # Optional
  test_path: # Optional
  quantization_path: ../datasets/flower_photos
  quantization_split: 0.4 # Optional
  seed: 0 # Optional

In this example, the only provided path is the quantization_path. It could be the full training set or a specific set dedicated to activations calibration. If you only want to quantize the model on a random part of your quantization set, simply set the percentage value in the quantization_split parameter.

State machine below describes the rules to follow when handling dataset path for the quantization.



Most of the time, it is not necessary to take thousands of images to calibrate your activations. It would take much more time and in the end does not necessarily bring an accuracy improvement. Although it is impossible to have a general rule, we have found that tens to hundreds of representative input samples are enough in most cases.

1.2 Apply preprocessing

The images from the dataset need to be preprocessed before they are presented to the network for quantization. This includes rescaling and resizing. In particular, they need to be rescaled exactly as they were at the training step. This is illustrated in the YAML code below:

preprocessing:
   rescaling: { scale: 1/127.5, offset: -1 }
   resizing: 
     aspect_ratio: "fit"
     interpolation: nearest
   color_mode: rgb

In this example, the pixels of the input images read in the dataset are in the interval [0, 255], that is UINT8. If you set scale to 1./255 and offset to 0, they will be rescaled to the interval [0.0, 1.0]. If you set scale to 1/127.5 and offset to -1, they will be rescaled to the interval [-1.0, 1.0].

The resizing attribute specifies the image resizing methods you want to use:

• The value of interpolation must be one of {"bilinear", "nearest", "bicubic", "area", "lanczos3", "lanczos5", "gaussian", "mitchellcubic"}.
• The value of aspect_ratio must be either "fit" or "crop". If you set it to "fit", the resized images will be distorted if their original aspect ratio is not the same as the resizing size. If you set it to "crop", images will be cropped as necessary to preserve the aspect ratio.

The color_mode attribute must be one of "grayscale", "rgb" or "rgba".

1.3 Set the model and quantization parameters

model:
   model_path: ../../stm32ai-modelzoo/image_classification/mobilenetv2/ST_pretrainedmodel_public_dataset/tf_flowers/mobilenetv2_a035_128_fft/mobilenetv2_a035_128_fft.keras
   
quantization:
  operating_mode: # Optional, if not set then operating_mode is 'default'
  quantizer: TFlite_converter
  quantization_type: PTQ
  quantization_input_type: uint8
  quantization_output_type: float
  granularity: per_channel  # Optional, defaults to "per_channel".
  optimize: False           # Optional, defaults to False.
  target_opset: 17          # Optional, defaults to 17 if not provided. Only used for when using Onnx_quantizer
  export_dir: quantized_models
  onnx_quant_parameters: # Optional, only applies to Onnx_quantizer
  onnx_extra_options:  # Optional, only applies to Onnx_quantizer
  iterative_quant_parameters: # Optional, only applies to Onnx_quantizer

where:

• operating_mode - String, can be "default", "inspection" or "full_auto". In case onnx quantizer is selected several operating modes are possible either for debug or for mixed precision quantization. They are all further detailed in a dedicated README.
• model_path - String, specifies the path of the model to be quantized. Can be '.onnx' or '.keras'
• quantizer - String, only options are "TFlite_converter" or "Onnx_quantizer" which will convert model trained weights from float to integer values. The quantized model will be saved in TensorFlow Lite or QDQ Onnx formats respectively.
• quantization_type - String, only option is "PTQ", i.e., "Post-Training Quantization".
• quantization_input_type - String, can be "int8", "uint8" or "float", represents the quantization type for the model input.
• quantization_output_type - String, can be "int8", "uint8" or "float", represents the quantization type for the model output.
• granularity - String, can be "per_tensor" or "per_channel", defines the quantization granularity.
• optimize - Boolean, can be either True or False, controls whether the user wants to optimize the model before attempting to quantize it. Only used when "TFlite_converter" and granularity is per_tensor.
• target_opset - Integer, the target opset for the ONNX models, only used when "Onnx_quantizer" is used.
• export_dir - String, refers to the directory name to save the quantized model.
• onnx_quant_parameters - Field, specific parameters for onnx quantizer only. Optional. For more details, please refer to the dedicated README.
• onnx_extra_options - Field, options supported for onnx quantizer. This field is optional. For more details, please refer to the dedicated README.
• iterative_quant_parameters - Field, for onnx quantizer only. Mandatory for mixed precision full_auto operating mode. For more details, please refer to the dedicated README.

Please note that whatever the values set in quantization_input_type and quantization_output_type, the ONNX QDQ model generated at this stage has always a floating-point input and output.

1.4 Random quantization

When no path is specified in quantization_path or training_path, the model is quantized after calibration on random data. There is no interest in evaluating the accuracy in this case. However, this random quantization can be useful to quickly estimate the model footprints on a target after quantization. We will see how to proceed in the next section.

1.5 Hydra and MLflow settings

The mlflow and hydra sections must always be present in the YAML configuration file. The hydra section can be used to specify the name of the directory where experiment directories are saved and/or the pattern used to name experiment directories. With the YAML code below, every time you run the Model Zoo, an experiment directory is created that contains all the directories and files created during the run. The names of experiment directories are all unique as they are based on the date and time of the run.

hydra:
   run:
      dir: ./tf/src/experiments_outputs/${now:%Y_%m_%d_%H_%M_%S}

The mlflow section is used to specify the location and name of the directory where MLflow files are saved, as shown below:

mlflow:
   uri: ./tf/src/experiments_outputs/mlruns

To launch your model quantization using a real dataset, run the following command from the UC folder:

python stm32ai_main.py --config-path ./config_file_examples/ --config-name quantization_config.yaml

The quantized TensorFlow Lite model can be found in the corresponding experiments_outputs/ folder.

In case you want to evaluate the accuracy of the quantized model, you can either launch the evaluation operation mode on the generated quantized model (please refer to the evaluation readme that describes in detail how to proceed) or you can use chained services like launching the chain_eqe example with the command below:

python stm32ai_main.py --config-path ./config_file_examples/ --config-name chain_eqe_config.yaml

In case you want to evaluate your quantized model footprints, you can either launch the benchmark operation mode on the generated quantized model (please refer to the benchmarking readme that describes in detail how to proceed) or you can use chained services like launching the chain_qb example with the command below:

python stm32ai_main.py --config-path ./config_file_examples/ --config-name chain_qb_config.yaml

Chained services work whether you specify a quantization dataset or not (random quantization).