|
|
|
|
|
""" |
|
================================================ |
|
@author: Jaron |
|
@time: 2024/08/21 17:51:45 |
|
@email: [email protected] |
|
@description: |
|
================================================ |
|
""" |
|
from transformers import PretrainedConfig |
|
|
|
|
|
class VideoCCAMConfig(PretrainedConfig): |
|
model_type = 'videoccam' |
|
_auto_class = 'AutoConfig' |
|
|
|
def __init__( |
|
self, |
|
llm_name_or_path: str = None, |
|
projector_name_or_path: str = None, |
|
vision_encoder_name_or_path: str = None, |
|
image_token: str = '<image>', |
|
video_token: str = '<video>', |
|
vision_select_layer: int = -2, |
|
vision_max_chunk_size: int = 0, |
|
_attn_implementation: str = 'flash_attention_2', |
|
**kwargs |
|
): |
|
super().__init__(**kwargs) |
|
self.llm_name_or_path = llm_name_or_path |
|
self.projector_name_or_path = projector_name_or_path |
|
self.vision_encoder_name_or_path = vision_encoder_name_or_path |
|
self.image_token = image_token |
|
self.video_token = video_token |
|
self.vision_select_layer = vision_select_layer |
|
self.vision_max_chunk_size = vision_max_chunk_size |
|
self._attn_implementation = _attn_implementation |
|
|
