MCPcopy Create free account
hub / github.com/allenai/molmo2 / TokenizedVisionData

Class TokenizedVisionData

olmo/preprocessing/preprocessor_utils.py:15–46  ·  view source on GitHub ↗

Data for visual input that has been converted into tokens

Source from the content-addressed store, hash-verified

13
14@dataclasses.dataclass
15class TokenizedVisionData:
16 """Data for visual input that has been converted into tokens"""
17
18 tokens: np.ndarray
19 """Token ids to use in the LLM"""
20
21 images: np.ndarray
22 """Images in [n_images, n_patches, patch_size] format"""
23
24 image_masks: Optional[np.ndarray] = None
25 """Images in [n_images, n_patches] format"""
26
27 token_pooling: Optional[np.ndarray] = None
28 """[n_visual_tokens, pooling_dim] array of of how to pool ViT patches for patch_id tokens"""
29
30 low_res_token_pooling: Optional[np.ndarray] = None
31 """[n_visual_tokens, pooling_dim] array of of how to pool ViT patches for low_res tokens"""
32
33 position_ids: Optional[np.ndarray] = None
34 """Position id for each token if not sequential"""
35
36 cum_token_pooling_bounds: Optional[Dict[str, np.ndarray]] = None
37 "cumulative token pooling boundaries used when training with context parallelism on images and videos"
38
39 cum_image_bounds: Optional[int] = None
40 """boundaries of crops in an input"""
41
42 other_data: Optional[Dict[str, np.ndarray]] = None
43 """Any extra data that needs to be passed to the model"""
44
45 token_mapping: Optional[np.ndarray] = None
46 """Map patch coordinates -> patch ids"""
47
48
49@dataclasses.dataclass

Callers 2

__call__Method · 0.90

Calls

no outgoing calls

Tested by

no test coverage detected