The role of dataset classes in Transfer Learning

Original article can be found here (source): Deep Learning on Medium

Here we have 2 datasets

  • Cats-Dogs dataset having 2 classes.
  • Logo classification having 16 classes.

Creating and managing experiments

  • Provide project name
  • Provide experiment name

This creates files and directories as per the following structure

workspace 
|--Project
|--study-num-classes | |--experiment-state.json | |--output | |--logs (All training logs and graphs saved here) | |--models (all trained models saved here)

Setup Default Params with Cats-Dogs dataset

gtf.Default(dataset_path="study_classes/dogs_vs_cats", 
model_name="resnet18",
num_epochs=5)

Visualize network

gtf.Visualize_With_Netron(data_shape=(3, 224, 224), port=8081)

The final layer

Reset Default Params with new dataset — Logo classification

gtf.Default(dataset_path="study_classes/logos", 
model_name="resnet18",
num_epochs=5)

Visualize network

gtf.Visualize_With_Netron(data_shape=(3, 224, 224), port=8082)

The final layer