Author:Ziyan Jiang([email protected]) Dalian University of Technology

MNIST(Mixed National Institute of Standards and Technology database), is a classic written number dataset. To complete this task, you need to do as follows:

• First, use GitBash to clone the repositories:

  git clone https://github.com/jiangziyan-693/MNIST.git
  

• Second, you need to create a new virtual environment，please input the code below in terminal:

  conda create -n your_environment_name python=3.9
  conda activate your_environment_name
  

• Third, you need to install the dependency package, please input the code below in terminal:

  pip install -r requirements.txt
  

Researching the influence of Hyperparameter to LLM is a key task of this work, we have created a easy way for you to change the Hyperparameters, just open config.yaml, and change the parameters as you like:

BATCH_SIZE: 256
EPOCHS: 10
LEARNING_RATE: 0.0001
ACTIVATION_FUNCTION: 'ReLU' # 'ReLU' or 'Sigmoid' or 'Tanh'

• BATCH_SIZE: control the number of samples processed before the model's internal parameters are updated.
• EPOCHS: controls the number of complete passes through the entire training dataset.
• LEARNING_RATE: controls the speed at which a model's weights are updated during training.
• ACTIVATION_FUCTION: controls the output of a neural network node by introducing non-linearity.

To run the program, please input the code below in terminal:

sh run_source.sh