Here is a semantic segmentation project, training on the capsicum dataset, using deeplabv3_resnet defined in Pytorch.

The project structure:

Capsicum Segmentation
--data (for dataset)
--dataloaders
--trained_models (for saving trained models)
--images    (for the images you wanna predict)
--output (for the decoded predicted masks)
--train_log_dir (for logging information)
--install.sh    (for the installation instruction)
--make_train_val_test.py    (for splitting dataset into 3 parts)
--predict.py    
--train.py
--README.MD

(1) Run make_train_val_test.py to make train, val, test dataset csv file,

(2) Run train.py to train the model,

(3) Use following command to check the metrics ( Loss, Acc, IoU ) and intermediate predicted images

tensorboard --logdir=train_log_dir --port=6006

(4) Run predict.py to predict an image, change the variable "image_path" to predict your image, change the variable "model_name" to use another trained model