config_loss_and_miners

loss_funcs

The loss functions are given embeddings and labels, and output a value on which back propagation can be performed. This config option is a mapping from strings to loss classes. The strings should match the loss names used by your trainer.

Default yaml:

loss_funcs:
  metric_loss: 
    ContrastiveLoss:

Example command line modification:

# Use a different loss function
--loss_funcs {metric_loss~OVERRIDE~: {MultiSimilarityLoss: {alpha: 0.1, beta: 40, base: 0.5}}}

sampler

The sampler is passed to the PyTorch dataloader, and determines how batches are formed. Use {} if you want random sampling.

Default yaml:

sampler:
  MPerClassSampler:
    m: 4

Example command line modification:

# Use random sampling
--sampler~OVERRIDE~ {}

mining_funcs

Mining functions determine the best tuples to train on, within an arbitrarily formed batch. This config option is a mapping from strings to miner classes. The strings should match the miner names used by your trainer.

Default yaml:

mining_funcs: {}

Example command line modification:

# Use a miner
--mining_funcs {tuple_miner: {MultiSimilarityMiner: {epsilon: 0.1}}}