Stuck on an issue?

Lightrun Answers was designed to reduce the constant googling that comes with debugging 3rd party libraries. It collects links to all the places you might be looking at while hunting down a tough bug.

And, if you’re still stuck at the end, we’re happy to hop on a call to see how we can help out.

Load pre-trained models

See original GitHub issue

Hey @mravanelli @pplantinga @Gastron @TParcollet ,

With my transducer, I’m thinking about adding specific pre-trained option to handle the load of pre-trained models from CTC for example…

The first solution is to add an option in checkpointer like this:

checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
    checkpoints_dir: !ref <save_folder>
    recoverables:
        encoder_crdnn: !ref <encoder_crdnn>
        encoder_lin: !ref <encoder_lin>
        decoder_gru: !ref <decoder_gru>
        decoder_lin: !ref <decoder_lin>
        joint_lin: !ref <joint_lin>
        output: !ref <output>
        optimizer: !ref <optimizer>
        scheduler: !ref <lr_annealing>
        normalizer: !ref <normalize>
        counter: !ref <epoch_counter>

And consider that encoder_crdnn and encoder_lin have same dim as CTC recipe…I’m thinking about creating an option like:

checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
    checkpoints_dir: !ref <save_folder>
    pre-trained-dir: !ref <pretrained_folder>
    load-pre-trained:
        encoder_crdnn: !ref <encoder_crdnn>
        encoder_lin: !ref <encoder_lin>
    recoverables:
        encoder_crdnn: !ref <encoder_crdnn>
        encoder_lin: !ref <encoder_lin>
        decoder_gru: !ref <decoder_gru>
        decoder_lin: !ref <decoder_lin>
        joint_lin: !ref <joint_lin>
        output: !ref <output>
        optimizer: !ref <optimizer>
        scheduler: !ref <lr_annealing>
        normalizer: !ref <normalize>
        counter: !ref <epoch_counter>

Or adding a class as @TParcollet propose to load model at the begning…