README_segmentation.md

Training a network from a database.

We suggest the following organisation of folder for training and database

.  
├── weights #empty, to save model weights 
├── board #empty, for tensorboard visualisation  
└── data  
    └── dataset  
            ├── train  
            │       ├── images #training images (png or jpg) RGB
            │       └── labels #training labels (specific format, read the doc) 
            ├── val  
            │       ├── images  
            │       └── labels  
            └── test  
                    ├── images  
                    └── labels  
                    

The parent directory should be specified in the config file

[TrainingDirectory]
path = "/home/tim/shared/"
directory_weights = "weights/"
tsboard = "board/"
directory_dataset = "dataset_stem_green"

[Segmentation2D]
upstream_task = "Scan"
query = "{\"channel\":\"rgb\"}"
labels = "background,flower,peduncle,stem,bud,leaf,fruit"
model_name = "Resnet101"
model_segmentation_name = "Resnet101_896_896_epoch51.pt"
Sx = 896
Sy = 896
epochs = 5
batch = 1
learning_rate = 0.0001

Generate a dataset with blender

To be completed

Parameters

the training parameters are to be filled in parameters_train.toml.
-size of the training images (center crop from images in train/images)
-type of segmentation network (for now they are sourced in segmentation_models_pytorch
-batch size -number of epochs
-learning rate

Launch training

simply launch

train_cnn.py --config [path/to/config/file]  

you can also import the train_cnn function for multiple tests