Skip to content

Quantization Traning

Jump to bottom
webstermaster777 edited this page Sep 21, 2022 · 9 revisions

Quantization

The project makes it possible to get a trained statically or dynamically quantized model.

Dynamic quantization

To obtain a dynamically quantized model, it is necessary to transfer the already trained model to the script dynamic_quantization.py with config file (see how get trained model). After executing the program, the dynamically compressed model will be saved as qmodel.pt.

python dynamic_quantization.py

Static quantization

To obtain a statically quantized model, you need to repeat the same steps as with the training of a conventional model, but the word "quat" is attributed to the names of the scripts being executed.

Clone Prject from git

git clone https://github.com/WEBSTERMASTER777/TripletRecognitionSystem.git

Load a data set, any data set with the following structure is suitable:

-- DIR WITH DATASET NAME:

    --- Dir_class_1

    --- Dir_class_2
    ...

    --- Dir class_n

The architecture of our network is great for training on small datasets, examples of datasets on kaggle:

Install requirements:

pip install -r requirements.txt -f https://download.pytorch.org/whl/torch_stable.html

Train

Starting network training, all necessary configurations need to be configured in config/train_parameters.yaml

# fix random 
random_seed: 7
cudnn_deterministic: True 

# Number of classes to classify
num_classes: 5

# Path to the archive with data
data_zip_path: 'archive_full.zip'

# Path to the directory to save files
data_save_path: 'data/'

# Fraction of the test dataset
test_size: 0.2

# Number of sample per draw ( to increase probability to
#               contain valid triplets in a batch )
n_sample: 3

batch_size: 12

# Count of epochs:
epochs: 30

# Path to save the plot with embeddings
plot_embeddings_img: './PDD.png'

# Path to save model parameters 
model_save_path: 'triplet_model_param.pt'

# Path to save optimizer parameters
optim_save_path: 'triplet_optim_param.pt'

# Colors of points on plot 
point_colors: ['#00ffff', '#000000', '#0000ff', '#ff00ff',
          '#808080', '#008000', '#00ff00', '#800000',
          '#000080', '#808000', '#800080', '#ff0000',
          '#c0c0c0', '#008080', '#ffff00']

knn_metric: 'cosine'

#optimazer params
lr: 0.0001

weight_decay: 0.00005

#scheduler params

step_size: 7
 
gamma: 0.1

Train:

python quant_train.py

Train classifier:

python clf_train.py
python classifier_train.py

As a result of executing the script, three models appear in the root folder of the repository.

  • classifier.pt - MLP classifier model
  • model.pt - Static quantization CNN + MLP model
  • mobilemodel.pt - Model for Android application

To get the predictions of the model, you need to use the script perceptron_script.py with config on config/script_percep_param.yaml

# Path to model 
model: model.pt

# How many of the nearest points are required for the extraction
topn: 3

# Path to file with class names
class_names: classes2.txt

#Path to save prediction 
prediction_savefile: data_file.json # or False if not save

Run script:

python perceptron_script.py

Example output:

        1 rice_leaf_diseases__Brown spot
        2 rice_leaf_diseases__Bacterial leaf blight
        3 rice_leaf_diseases__Leaf smut

You can check tutorial for quantization training on notebook

Wiki

Clone this wiki locally