KeyError: 'Domain' when retraining model about voice-type-classifier HOT 4 CLOSED

The error probably comes from the Task specification.
In the config.yml, could you try replacing :

task:
  params:
    alpha: 0
    duration: 2.0
    domain: domain
    batch_size: 64
    attachment: 0
    per_epoch: 1
    domain_loss: MSELoss
    rnn:
      unit: LSTM
      hidden_size: 128
      num_layers: 1
      bidirectional: False
      pool: max
    labels_spec:
      regular: ['KCHI', 'CHI', 'MAL', 'FEM']
      union:
        SPEECH: ['KCHI', 'CHI', 'FEM', 'MAL', 'SPEECH']
  name: DomainAdversarialMultilabelDetection

by

task:
  params:
    duration: 2.0
    batch_size: 64
    per_epoch: 1
    labels_spec:
      regular: ['KCHI', 'CHI', 'MAL', 'FEM']
      union:
        SPEECH: ['KCHI', 'CHI', 'FEM', 'MAL', 'SPEECH']
  name: MultilabelDetection

It should solve the issue !
Let me know if it works please :)

PS : For more information, pyannote is expecting a domain key to train the model adversarially w.r.t the domain information such as in here. The weight associated to the domain loss being 0 in this case, it is strictly equivalent than not considering at all this domain adversarial loss.

from voice-type-classifier.

I've decided to update the config.yml in the repo directly to make the model retrainable more easily.
I'm waiting for your feedback before closing the issue :)

from voice-type-classifier.

This worked, thank you! Actually it resolved the first issue and then I experienced RuntimeError: CUDA out of memory errors. By changing the Batch Size=32, this resolve and it seems to be retaining now.

from voice-type-classifier.

Training completed successfully (we think). However, I cannot run the validate step - will open a new issue.

from voice-type-classifier.