Request for trainable DialogueSidon Lightning checkpoint
Thank you for uploading ssl_vae_l32/epoch=0-step=510000.ckpt in response to discussion #1.
We would like to fine-tune DialogueSidon on additional paired 24 kHz full-duplex dialogue data. The current release contains the three torch.export inference artifacts, but not the original DialogueSidonDiffusionLightningModule checkpoint. Would you be willing to share the .ckpt corresponding to the published DialogueSidon exports, or an official conversion/loading script that initializes the trainable module from those exports? Optimizer state is not necessary.
Could you also confirm that the newly published latent-32 checkpoint (run qmrvcukw) is the exact SSL-VAE used by the current DialogueSidon export? The published metadata says latent_dim: 32, while the default training YAML currently references a different private run path (jhqmq4u3).
Thank you!