SOT_CONFIG.md

Configuration yaml

Example experiments/siamfcpp/train/got10k/siamfcpp_alexnet-trn.yaml

Training

How to modify the number of epochs

• Modify the value in YAML
  • train.track.data.num_epochs
  • L88: num_epochs: &NUM_EPOCHS 20
• Modify the value in JSON (lr_policy)
  • train.track.optim.optimizer.SGD.lr_policy
  • L172: "max_epoch": 19

How to modify the minibatch size

• Modify the value in YAML
  • train.track.data.minibatch
  • L89: minibatch: &MINIBATCH 32

Note that the number of iterations is determined as follows: #iterations = _nr_image_per_epoch // minibatch

How to modify the number of image pairs of each epoch

• Modify the value in YAML
  • train.track.data.nr_image_per_epoch
  • L91: nr_image_per_epoch: &NR_IMAGE_PER_EPOCH 400000

How to modify the learning rate

• Modify the value in JSON (lr_policy)
  • train.track.optim.optimizer.SGD.lr_policy
  • L164: "end_lr": 0.08,
  • L170: "start_lr": 0.08,

How to modify the learning rate in backbone

• Modify the value in JSON (lr_policy)
  • train.track.optim.optimizer.SGD.lr_multiplier
  • L179: "ratio": 0.1

How to change the dynamic freezing schedule

• Modify the value in JSON (lr_policy)
  • train.track.optim.grad_modifier

How to change the number of GPUs

• Modify the value in YAML
  • train.track.num_processes
  • L51: num_processes: 2