wanghao9610
commented
3 months ago 如果你是参考custom_ovd.py定义的数据集： 首先，你得看这个文件(custom_ovd.py)里面你定义的测试数据集是哪个，看报错似乎你并没有修改？ 如果你是参考coco_ovd.py定义的数据集： 你需要在register_custom_ovd.py修改相应的参数（是一个字典，改相应的value就可以）。
Closed menghan1124 closed 3 months ago
wanghao9610
commented
3 months ago 如果你是参考custom_ovd.py定义的数据集： 首先，你得看这个文件(custom_ovd.py)里面你定义的测试数据集是哪个，看报错似乎你并没有修改？ 如果你是参考coco_ovd.py定义的数据集： 你需要在register_custom_ovd.py修改相应的参数（是一个字典，改相应的value就可以）。
menghan1124
commented
3 months ago 写错了，是参考的参考custom_ovd.py定义的数据集，这里面已经修改改过了，下面是custom_ovd.py的代码，dataloader.evaluator我也试过了自带的代码，和自己定义的代码
import itertools
import detectron2.data.transforms as T
from detectron2.config import LazyCall as L
from detectron2.data import (
build_detection_test_loader,
build_detection_train_loader,
build_detection_val_loader,
get_detection_dataset_dicts,
)
from detectron2.evaluation import COCOEvaluator
from detrex.data import DetrDatasetMapper
from detrex.data.datasets import register_coco_ovd_instances
from omegaconf import OmegaConf
dataloader = OmegaConf.create()
# if you follow the coco format, you can use the following code.
# if you want to define it by yourself, you can change it on ovdino/detrex/data/datasets/custom_ovd.py.
register_coco_ovd_instances(
"custom_train1", # dataset_name
{}, # custom_data_info
"/workspace/OV-DINO/datas/custom/annotations/train.json", # annotations_jsonfile
"/workspace/OV-DINO/datas/custom/train", # image_root
27, # number_of_classes, default: 80
"full", # template, default: full
)
register_coco_ovd_instances(
"custom_val1",
{},
"/workspace/OV-DINO/datas/custom/annotations/val.json",
"/workspace/OV-DINO/datas/custom/val",
27,
"full",
)
register_coco_ovd_instances(
"custom_test1",
{},
"/workspace/OV-DINO/datas/custom/annotations/test.json",
"/workspace/OV-DINO/datas/custom/test",
27,
"identity",
)
dataloader.train = L(build_detection_train_loader)(
dataset=L(get_detection_dataset_dicts)(names="custom_train1"),
mapper=L(DetrDatasetMapper)(
augmentation=[
L(T.RandomFlip)(),
L(T.ResizeShortestEdge)(
short_edge_length=(
480,
512,
544,
576,
608,
640,
672,
704,
736,
768,
800,
),
max_size=1333,
sample_style="choice",
),
],
augmentation_with_crop=[
L(T.RandomFlip)(),
L(T.ResizeShortestEdge)(
short_edge_length=(400, 500, 600),
sample_style="choice",
),
L(T.RandomCrop)(
crop_type="absolute_range",
crop_size=(384, 600),
),
L(T.ResizeShortestEdge)(
short_edge_length=(
480,
512,
544,
576,
608,
640,
672,
704,
736,
768,
800,
),
max_size=1333,
sample_style="choice",
),
],
is_train=True,
mask_on=False,
img_format="RGB",
),
total_batch_size=16,
num_workers=4,
)
dataloader.test = L(build_detection_test_loader)(
dataset=L(get_detection_dataset_dicts)(
names="custom_test1", filter_empty=False
),
mapper=L(DetrDatasetMapper)(
augmentation=[
L(T.ResizeShortestEdge)(
short_edge_length=800,
max_size=1333,
),
],
augmentation_with_crop=None,
is_train=False,
mask_on=False,
img_format="RGB",
),
num_workers=4,
)
# dataloader.evaluator = L(COCOEvaluator)(
# dataset_name="${..test.dataset.names}",
# )
dataloader.evaluator = L(build_detection_val_loader)(
dataset=L(get_detection_dataset_dicts)(
names="custom_val1", filter_empty=False
),
mapper=L(DetrDatasetMapper)(
augmentation=[
L(T.ResizeShortestEdge)(
short_edge_length=800,
max_size=1333,
),
],
augmentation_with_crop=None,
is_train=False,
mask_on=False,
img_format="RGB",
),
num_workers=4,
)看上去似乎没有什么问题，你可以debug看看/workspace/OV-DINO/ovdino/detrex/data/datasets/coco_ovd.py", line 321 这个位置传递的参数是否是正确的。
menghan1124
commented
3 months ago 看上去似乎没有什么问题，你可以debug看看/workspace/OV-DINO/ovdino/detrex/data/datasets/coco_ovd.py", line 321 这个位置传递的参数是否是正确的。
目前是用这份数据微调没有问题，微调过程中Evaluate也没问题，用微调后的模型跑scripts/eval.sh脚本出现的报错
wanghao9610
commented
3 months ago 配置文件错了，应该用这个：
sh scripts/eval.sh projects/ovdino/configs/ovdino_swin_tiny224_bert_base_ft_custom_24ep.py ../wkdrs/ovdino_swin_tiny224_bert_base_ft_custom_24ep/model_0034999.pth evaldirssh scripts/eval.sh projects/ovdino/configs/ovdino_swin_tiny224_bert_base_ft_coco_24ep.py ../wkdrs/ovdino_swin_tiny224_bert_base_ft_custom_24ep/model_0034999.pth evaldirs/
配置文件错了，应该用这个：
sh scripts/eval.sh projects/ovdino/configs/ovdino_swin_tiny224_bert_base_ft_custom_24ep.py ../wkdrs/ovdino_swin_tiny224_bert_base_ft_custom_24ep/model_0034999.pth evaldirs
好的感谢，已经跑通
otakue
commented
3 months ago 请问微调后会出现遗忘性问题吗?
wanghao9610
commented
3 months ago @otakue 微调能提升在相应数据集上的性能，但相应的泛化（zero-shot）性能会降低。
命令：
sh scripts/eval.sh projects/ovdino/configs/ovdino_swin_tiny224_bert_base_ft_coco_24ep.py ../wkdrs/ovdino_swin_tiny224_bert_base_ft_custom_24ep/model_0034999.pth evaldirs/报错信息：
用自定义数据训练时正常。用训练的最优模型参数评估样本数据时报错，请问需要修改数据配置的那一部分？
custom_ovd.py中已经加上coco格式自定义数据集信息