CNN_via_Tensorflow2_low-level

This project's purpose is building convolutional neural network implementation with Tensorflow2.0 low-level API only. It looks somewhat useless things because Keras layers are so easy to handle. However, I have many layers that were built in TF1 low-level API, and surely, there is no Keras layer for them. So, I have to do this, and perhaps some guys need it too.

Require

• Tensorflow > 2.0

Contents

• nets/tcl.py

  • Build a custom layer and add trainable or untrainable parameters.
  • Add regularization for each trainable parameters.
  • Define update function for the moving mean and the moving standard deviation.
  • Conditioning for training and inference phase.
  • Prototype of arg_scope. (will be updated)

• op_util.py

  • Define a loss function with regularization losses.
  • Build optimizer with computing and applying gradients.
  • Define steps for training and inference.
  • Learning rate scheduler

• train_and_validate.py

  • Load dataset, pre_processing algorithn, model, and optimizer.
  • Do train and validate.
  • Visualize the log via Tensorboard.

• dataloader.py

  • Load dataset
  • Define pre-processing algorithm.

• nets/ResNet.py and WResNet.py

  • Build a custom model via custom layers.
  • How to use implemented arg_scope

To do

• Write Readme and milestones.
• Codes to save and load models without a checkpoint.
• Improve readability of a custom model.
• Find more things to do...