CLI train¶
The help documentation is available in the terminal from:
peerannot train --help
All computer vision models that can be used for training are available through the Torchvision
library and can be found running:
peerannot modelinfo --help
peerannot train¶
Train a classification neural network given a dataset path, an output name and the number of classes
peerannot train [OPTIONS] [DATAPATH]
Options
- -o, --output-name <output_name>¶
Name of the generated results file
- -K, --n-classes <n_classes>¶
Number of classes to separate
- --labels <labels>¶
Path to file of labels
- -optim, --optimizer <optimizer>¶
Optimizer for the neural network
- --model <model>¶
Name of neural network to use. The list is available at peerannot modelinfo
- --metadata_path <metadata_path>¶
Path to the metadata of the dataset if different than default
- --img-size <img_size>¶
Size of image (square)
- --data-augmentation¶
Perform data augmentation on training set with a random choice between RandomAffine(shear=15), RandomHorizontalFlip(0.5) and RandomResizedCrop
- Default:
False
- --path-remove <path_remove>¶
Path to file of index to prune from the training set
- --pretrained¶
Use torch available weights to initialize the network
- Default:
False
- --n-epochs <n_epochs>¶
Number of training epochs
- --lr <lr>¶
Learning rate
- --momentum <momentum>¶
Momentum for the optimizer
- --decay <decay>¶
Weight decay for the optimizer
- --scheduler <scheduler>¶
Use a multistepscheduler for the learning rate by default. To use the cosine annealing use the keyword ‘cosine’
- Default:
'multistep'
- -m, --milestones <milestones>¶
Milestones for the learning rate decay scheduler
- --n-params <n_params>¶
Number of parameters for the logistic regression only
- --lr-decay <lr_decay>¶
Learning rate decay for the scheduler
- --num-workers <num_workers>¶
Number of workers
- --batch-size <batch_size>¶
Batch size
- --freeze¶
Freeze all layers of the network except for the last one
- Default:
False
- --seed <seed>¶
random state
Arguments
- DATAPATH¶
Optional argument