ijepa_large_16(pretrained: bool = False, overrides: object = {})I-JEPA with a ViT-L/16 encoder.
Model Size
Parameters
pretrainedbool= FalseNo weights are published for this size;
True raises.**overridesobject= {}Optional
IJEPAConfig field overrides.Returns
IJEPAModelContext encoder, target encoder and predictor, untrained.
Notes
Reference: Assran et al., arXiv:2301.08243, Table 1 — 77.5% ImageNet-1k linear probe after 600 epochs, and 69.4% on 1% of ImageNet (Table 2). Table 13 reports 77.8% for the same model under the weight-decay schedule it prefers, which is the paper disagreeing with itself by 0.3.
Examples
>>> from lucid.models import AutoConfig
>>> config = AutoConfig.from_pretrained("ijepa_large_16")
>>> config.dim, config.depth, config.num_heads
(1024, 24, 16)