input_size- input image resolution. If the image size differs from the defined value, it will be resized automatically.
batch_size- batch size per single GPU, e.g. if you train NN with 4 GPUs the number of processed images on each iteration equals to
4 * batch_size
dataset_tags- tags that are used to split input data to train and validation datasets. Default tag for training set is
train, default tag for validation set is
val. If you are going to use your custom tags, you shoud define this field otherwise training process will be crash with error. Here is an example of how you can use custom tags for splitting:
data_workers- defines the number of threads for batch preparation
special_classes- objects with specified classes will be interpreted in a specific way. The default class name for
bg, the default class name for
neutral. All pixels from
neutralobjects will be ignored in the loss function. Here is an example of how you can use custom tags for
epochs- number of training epochs, i.e. how many times NN will look at each image from the training dataset.
val_every- defines how frequently validation will be performed.
0.5means that model wil be validated 2 times per epoch.
lr- learning rate
gpu_devices- list of GPU devices used in training, e.g. for 4 GPU training you should set
"gpu_devices": [0, 1, 2, 3]
weights_init_type- how NN weights will be inited. In
"transfer_learning"mode all possible weights will be transfered except the last layer. In
"continue_training"mode all weights will be transfered and validation for classes number and classes names order will be performed.