47 lines
1.3 KiB
Python
47 lines
1.3 KiB
Python
_base_ = [
|
|
'../_base_/datasets/nocaps.py',
|
|
'../_base_/default_runtime.py',
|
|
]
|
|
|
|
# model settings
|
|
model = dict(
|
|
type='BlipCaption',
|
|
vision_encoder=dict(
|
|
type='VisionTransformer',
|
|
arch='b',
|
|
img_size=384,
|
|
patch_size=16,
|
|
out_type='raw',
|
|
),
|
|
tokenizer=dict(type='BlipTokenizer', name_or_path='bert-base-uncased'),
|
|
decoder_head=dict(
|
|
type='SeqGenerationHead',
|
|
decoder=dict(
|
|
type='XBertLMHeadDecoder',
|
|
med_config=dict(
|
|
architectures=['BertModel'],
|
|
attention_probs_dropout_prob=0.1,
|
|
hidden_act='gelu',
|
|
hidden_dropout_prob=0.1,
|
|
hidden_size=768,
|
|
initializer_range=0.02,
|
|
intermediate_size=3072,
|
|
layer_norm_eps=1e-12,
|
|
max_position_embeddings=512,
|
|
model_type='bert',
|
|
num_attention_heads=12,
|
|
num_hidden_layers=12,
|
|
pad_token_id=0,
|
|
add_type_embeddings=False,
|
|
vocab_size=30524,
|
|
encoder_width=768,
|
|
add_cross_attention=True),
|
|
),
|
|
),
|
|
prompt='a picture of ',
|
|
max_txt_len=20,
|
|
)
|
|
|
|
val_cfg = dict()
|
|
test_cfg = dict()
|