clip_vit_large_14_336

→CLIPModel

clip_vit_large_14_336(pretrained: bool = False, overrides: object = {})

source edit

Construct CLIP with a ViT-L/14 image tower at 336 pixels.

Model Size

Parameters

pretrainedbool= False

Load the OpenAI WIT-400M checkpoint from the Lucid hub.

**overridesobject= {}

Optional CLIPConfig field overrides.

Returns

CLIPModel

Both towers and the learned temperature.

Notes

Reference: Radford et al., ICML 2021 (arXiv:2103.00020), §2.5 — "for the ViT-L/14 we also pre-train at a higher 336 pixel resolution for one additional epoch", denoted ViT-L/14@336px.

Architecturally this differs from clip_vit_large_14 in one number, but that number changes the positional table from 257 rows to 577, so the two are not weight-compatible.

Examples

>>> from lucid.models import clip_vit_large_14_336
>>> model = clip_vit_large_14_336()
>>> model.config.image_size
336

Used by 1

lucid.models.multimodal.clip