Question about the SemanticTransformerTrainer

[djqualia]

I have a naive/silly question, and based on the answer I get I'll submit a PR to address my confusion (either with a comment that explains, or fixing the example)...

In the main readme instructions, when giving an example for how to train the SemanticTransformerTrainer [1], I noticed that num_train_steps is set to 1.

Is this intentional? (i.e. is there just some basic conversion going on from the underlying wave2vec/hubert model?)

Or is this just a poor example? (i.e. you'd never actually train with just 1 step, but the reader/me should have gotten the idea?)

The similar example below it that includes text conditioning does have a much higher value set. So I wondered whether training more than 1 steps was only necessary with text conditioning?

Thanks in advance for any insight you can provide!

[1]
trainer = SemanticTransformerTrainer(
transformer = semantic_transformer,
wav2vec = wav2vec,
folder = '/home/phil/dl/data/LibriSpeech',
batch_size = 1,
data_max_length = 320 * 32,
num_train_steps = 1
)

[StevenSchrembeck]

I think it's just a little code hack to reuse the same class instance again for generation. (real experts: correct me if i'm wrong)

at inference time you can also just create the underlying SemanticTransformerWrapper yourself using yourNewSemanticTransformer.load_state_dict(YOUR_SAVED_TRAINED_MODEL_HERE) as the transformer param. then you can call that thing's generate method, shedding the much larger trainer class. that's what i did

hope that helps!