OpenCompass/configs/multimodal/otter/otter_9b_mmbench.py
Li Bo a4d6840739
[Feat] Add Otter to OpenCompass MMBench Evaluation (#232)
* add otter model for opencompass mmbench

* add docs

* add readme docs

* debug for otter opencomass eval

* delete unused folders

* change to default data path

* remove unused files

* remove unused files

* update

* update config file

* flake8 lint formated and add prompt generator

* add prompt generator to config

* add a specific postproecss

* add post processor

* add post processor

* add post processor

* update according to suggestions

* remove unused redefinition
2023-08-31 12:55:53 +08:00

44 lines
1.4 KiB
Python

# dataloader settings
from opencompass.multimodal.models.otter import (
OTTERMMBenchPromptConstructor, OTTERMMBenchPostProcessor)
val_pipeline = [
dict(type="mmpretrain.torchvision/Resize", size=(224, 224), interpolation=3),
dict(type="mmpretrain.torchvision/ToTensor"),
dict(
type="mmpretrain.torchvision/Normalize",
mean=(0.48145466, 0.4578275, 0.40821073),
std=(0.26862954, 0.26130258, 0.27577711),
),
dict(
type="mmpretrain.PackInputs",
algorithm_keys=["question", "answer", "options", "category", "l2-category", "context", "index", "options_dict"],
),
]
dataset = dict(
type="opencompass.MMBenchDataset", data_file="/path/to/mmbench/mmbench_test_20230712.tsv", pipeline=val_pipeline
)
otter_9b_mmbench_dataloader = dict(
batch_size=1,
num_workers=4,
dataset=dataset,
collate_fn=dict(type="pseudo_collate"),
sampler=dict(type="DefaultSampler", shuffle=False),
)
# model settings
otter_9b_mmbench_model = dict(
type="otter-9b",
model_path="/path/to/OTTER-Image-MPT7B/", # noqa
load_bit="bf16",
prompt_constructor=dict(type=OTTERMMBenchPromptConstructor,
model_label='GPT',
user_label='User'),
post_processor=dict(type=OTTERMMBenchPostProcessor)
)
# evaluation settings
otter_9b_mmbench_evaluator = [dict(type="opencompass.DumpResults", save_path="work_dirs/otter-9b-mmbench.xlsx")]