IRED Autoencoder — roy-W/ired-reasoning

Frozen BART autoencoder with compression/reconstruction split.

• Pool type: conv
• Latent slots (K): 384
• d_ae: 768
• Decoder fine-tuned: yes
• Base model: facebook/bart-base
• Training steps: 5

Usage

from ired.model.autoencoder import FrozenBartAutoencoder
import torch

ae = FrozenBartAutoencoder(
    model_name='facebook/bart-base',
    k=384,
    pool_type='conv',
    d_ae=768,
    train_decoder=True,
)
ckpt = torch.load('ae_checkpoint.pt', map_location='cpu')
ae.load_ae(ckpt['ae'])