cnn-segmentation

CNN based Deep Learning project for image segmentation. Heavily relies on transfer learning.

Training

The available model has been trained with Kaggle, to take adavantage of the GPU. Kaggle notebook

Running the scripts

You can run the main.py script in order to :

• train for finetuning
• train for feature extraction
• run inferences

...depending on the values of the booleans train, and ft

See PyTorch's definitions of these terms. Pytorch's documentation

A DeepLab MobileNet example is already there, you can load it by setting model_choice to 'dlab_large'

The expected file structure if you use the MS COCO Dataset locally is :

In the data file :

• images
  • train
  • val
  • test
• annotations

Running the GUI

You can run the gui.py script You can use it without the MS COCO dataset if you set the coco parameter to False

Kaggle notebook used for training : link