mmpretrain/configs/eva/eva-l-p14_headless.py

27 lines
668 B
Python

model = dict(
type='ImageClassifier',
backbone=dict(
type='BEiTViT',
arch='l',
img_size=224,
patch_size=14,
avg_token=True,
layer_scale_init_value=0.0,
output_cls_token=False,
use_abs_pos_emb=True,
use_rel_pos_bias=False,
use_shared_rel_pos_bias=False,
layer_cfgs=dict(bias=True),
),
neck=None,
head=None,
)
data_preprocessor = dict(
# RGB format normalization parameters
mean=[0.48145466 * 255, 0.4578275 * 255, 0.40821073 * 255],
std=[0.26862954 * 255, 0.26130258 * 255, 0.27577711 * 255],
# convert image from BGR to RGB
to_rgb=True,
)