[Feature] Support DSDL Dataset

open-mmlab · May 10, 2023 · 9d7876c · 9d7876c
1 parent 77591b9
commit 9d7876c
Show file tree

Hide file tree

Showing 15 changed files with 464 additions and 5 deletions.
diff --git a/configs/dsdl/README.md b/configs/dsdl/README.md
@@ -0,0 +1,105 @@
+# DSDL: Standard Description Language for DataSet
+
+<!-- [ALGORITHM] -->
+
+<!-- [DATASET] -->
+
+## Abstract
+
+<!-- [ABSTRACT] -->
+
+Data is the cornerstone of artificial intelligence. The efficiency of data acquisition, exchange, and application directly impacts the advances in technologies and applications. Over the long history of AI, a vast quantity of data sets have been developed and distributed. However, these datasets are defined in very different forms, which incurs significant overhead when it comes to exchange, integration, and utilization -- it is often the case that one needs to develop a new customized tool or script in order to incorporate a new dataset into a workflow.
+
+To overcome such difficulties, we develop **Data Set Description Language (DSDL)**. More details please visit our [official documents](https://opendatalab.github.io/dsdl-docs/getting_started/overview/), dsdl datasets can be downloaded from our platform [OpenDataLab](https://opendatalab.com/).
+
+<!-- [IMAGE] -->
+
+## Steps
+
+- install dsdl and opendatalab:
+
+  ```
+  pip install dsdl
+  pip install opendatalab
+  ```
+
+- install mmseg and pytorch:
+  please refer this [installation documents](https://mmsegmentation.readthedocs.io/en/latest/get_started.html).
+
+- prepare dsdl dataset (take voc2012 as an example)
+
+  - dowaload dsdl dataset (you will need an opendatalab account to do so. [register one now](https://opendatalab.com/))
+
+    ```
+    cd data
+
+    odl login
+    odl get PASCAL_VOC2012
+    ```
+
+    usually, dataset is compressed on opendatalab platform, the downloaded voc 2012 dataset should be like this:
+
+    ```
+    data/
+    ├── PASCAL_VOC2012
+    │   ├── dsdl
+    │   │   ├── dsdl_Det_full.zip
+    │   │   └── dsdl_SemSeg_full.zip
+    │   ├── raw
+    │   │   ├── VOC2012test.tar
+    │   │   ├── VOCdevkit_18-May-2011.tar
+    │   │   └── VOCtrainval_11-May-2012.tar
+    │   └── README.md
+    └── ...
+    ```
+
+  - decompress dataset
+
+    ```
+    cd dsdl
+    unzip dsdl_SemSeg_full.zip
+    ```
+
+    as we do not need detection dsdl files, we only decompress the semantic segmentation files here.
+
+    ```
+    cd ../raw
+    tar -xvf VOCtrainval_11-May-2012.tar
+    tar -xvf VOC2012test.tar
+
+    cd ../../
+    ```
+
+- change traning config
+
+  here , we open the [voc config file](voc.py), and set some file paths as below:
+
+  ```
+  data_root = 'data/PASCAL_VOC2012'
+  img_prefix = 'raw/VOCdevkit/VOC2012'
+  train_ann = 'dsdl/dsdl_SemSeg_full/set-train/train.yaml'
+  val_ann = 'dsdl/dsdl_SemSeg_full/set-val/val.yaml'
+  ```
+
+  as dsdl datasets with one task using one dataloader, we can simplly change these file paths to train a model on a different dataset.
+
+- train:
+
+  - using single gpu:
+
+  ```
+  python tools/train.py {config_file}
+  ```
+
+  - using slrum:
+
+  ```
+  ./tools/slurm_train.sh {partition} {job_name} {config_file} {work_dir} {gpu_nums}
+  ```
+
+## Test Results
+
+|  Datasets  |                                                                                        Model                                                                                         | mIoU(%) |          Config           |
+| :--------: | :----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------: | :-----: | :-----------------------: |
+|  voc2012   |    [model](https://download.openmmlab.com/mmsegmentation/v0.5/deeplabv3/deeplabv3_r50-d8_512x512_20k_voc12aug/deeplabv3_r50-d8_512x512_20k_voc12aug_20200617_010906-596905ef.pth)    |  76.73  |    [config](./voc.py)     |
+| cityscapes | [model](https://download.openmmlab.com/mmsegmentation/v0.5/deeplabv3/deeplabv3_r50-d8_512x1024_40k_cityscapes/deeplabv3_r50-d8_512x1024_40k_cityscapes_20200605_022449-acadc2f8.pth) |  79.01  | [config](./cityscapes.py) |
diff --git a/configs/dsdl/cityscapes.py b/configs/dsdl/cityscapes.py
@@ -0,0 +1,70 @@
+_base_ = [
+    '../_base_/models/deeplabv3_r50-d8.py', '../_base_/default_runtime.py',
+    '../_base_/schedules/schedule_40k.py'
+]
+
+crop_size = (512, 1024)
+data_preprocessor = dict(size=crop_size)
+model = dict(data_preprocessor=data_preprocessor)
+# dataset settings
+dataset_type = 'DSDLSegDataset'
+data_root = 'data/CityScapes'
+img_prefix = 'raw/CityScapes'
+train_ann = 'dsdl/dsdl_SemSeg_full/set-train/train.yaml'
+val_ann = 'dsdl/dsdl_SemSeg_full/set-val/val.yaml'
+
+used_labels = [
+    'road', 'sidewalk', 'building', 'wall', 'fence', 'pole', 'traffic_light',
+    'traffic_sign', 'vegetation', 'terrain', 'sky', 'person', 'rider', 'car',
+    'truck', 'bus', 'train', 'motorcycle', 'bicycle'
+]
+
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='LoadAnnotations'),
+    dict(
+        type='RandomResize',
+        scale=(2048, 1024),
+        ratio_range=(0.5, 2.0),
+        keep_ratio=True),
+    dict(type='RandomCrop', crop_size=crop_size, cat_max_ratio=0.75),
+    dict(type='RandomFlip', prob=0.5),
+    dict(type='PhotoMetricDistortion'),
+    dict(type='PackSegInputs')
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='Resize', scale=(2048, 1024), keep_ratio=True),
+    # add loading annotation after ``Resize`` because ground truth
+    # does not need to do resize data transform
+    dict(type='LoadAnnotations'),
+    dict(type='PackSegInputs')
+]
+train_dataloader = dict(
+    batch_size=2,
+    num_workers=2,
+    persistent_workers=True,
+    sampler=dict(type='InfiniteSampler', shuffle=True),
+    dataset=dict(
+        type=dataset_type,
+        data_root=data_root,
+        data_prefix=dict(img_path=img_prefix, seg_map_path=img_prefix),
+        ann_file=train_ann,
+        used_labels=used_labels,
+        pipeline=train_pipeline))
+val_dataloader = dict(
+    batch_size=1,
+    num_workers=4,
+    persistent_workers=True,
+    sampler=dict(type='DefaultSampler', shuffle=False),
+    dataset=dict(
+        type=dataset_type,
+        data_root=data_root,
+        data_prefix=dict(img_path=img_prefix, seg_map_path=img_prefix),
+        ann_file=val_ann,
+        used_labels=used_labels,
+        pipeline=test_pipeline))
+test_dataloader = val_dataloader
+
+val_evaluator = dict(type='IoUMetric', iou_metrics=['mIoU'])
+test_evaluator = val_evaluator
diff --git a/configs/dsdl/metafile.yaml b/configs/dsdl/metafile.yaml
@@ -0,0 +1,12 @@
+Collections:
+- Name: ''
+  License: Apache License 2.0
+  Metadata:
+    Training Data: []
+  Paper:
+    Title: ''
+    URL: ''
+  README: configs/dsdl/README.md
+  Frameworks:
+  - PyTorch
+Models: []
diff --git a/configs/dsdl/voc.py b/configs/dsdl/voc.py
@@ -0,0 +1,65 @@
+_base_ = [
+    '../_base_/models/deeplabv3_r50-d8.py', '../_base_/default_runtime.py',
+    '../_base_/schedules/schedule_20k.py'
+]
+
+# dataset settings
+dataset_type = 'DSDLSegDataset'
+data_root = 'data/PASCAL_VOC2012'
+img_prefix = 'raw/VOCdevkit/VOC2012'
+train_ann = 'dsdl/dsdl_SemSeg_full/set-train/train.yaml'
+val_ann = 'dsdl/dsdl_SemSeg_full/set-val/val.yaml'
+crop_size = (512, 512)
+train_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='LoadAnnotations'),
+    dict(
+        type='RandomResize',
+        scale=(2048, 512),
+        ratio_range=(0.5, 2.0),
+        keep_ratio=True),
+    dict(type='RandomCrop', crop_size=crop_size, cat_max_ratio=0.75),
+    dict(type='RandomFlip', prob=0.5),
+    dict(type='PhotoMetricDistortion'),
+    dict(type='PackSegInputs')
+]
+test_pipeline = [
+    dict(type='LoadImageFromFile'),
+    dict(type='Resize', scale=(2048, 512), keep_ratio=True),
+    # add loading annotation after ``Resize`` because ground truth
+    # does not need to do resize data transform
+    dict(type='LoadAnnotations'),
+    dict(type='PackSegInputs')
+]
+train_dataloader = dict(
+    batch_size=4,
+    num_workers=4,
+    persistent_workers=True,
+    sampler=dict(type='InfiniteSampler', shuffle=True),
+    dataset=dict(
+        type=dataset_type,
+        data_root=data_root,
+        data_prefix=dict(img_path=img_prefix, seg_map_path=img_prefix),
+        ann_file=train_ann,
+        pipeline=train_pipeline))
+val_dataloader = dict(
+    batch_size=1,
+    num_workers=4,
+    persistent_workers=True,
+    sampler=dict(type='DefaultSampler', shuffle=False),
+    dataset=dict(
+        type=dataset_type,
+        data_root=data_root,
+        data_prefix=dict(img_path=img_prefix, seg_map_path=img_prefix),
+        ann_file=val_ann,
+        pipeline=test_pipeline))
+test_dataloader = val_dataloader
+
+val_evaluator = dict(type='IoUMetric', iou_metrics=['mIoU'])
+test_evaluator = val_evaluator
+
+data_preprocessor = dict(size=crop_size)
+model = dict(
+    data_preprocessor=data_preprocessor,
+    decode_head=dict(num_classes=21),
+    auxiliary_head=dict(num_classes=21))
diff --git a/docs/en/user_guides/2_dataset_prepare.md b/docs/en/user_guides/2_dataset_prepare.md
@@ -2,6 +2,7 @@
 
 It is recommended to symlink the dataset root to `$MMSEGMENTATION/data`.
 If your folder structure is different, you may need to change the corresponding paths in config files.
+For users in China, we also recommend you get the dsdl dataset from our opensource platform [OpenDataLab](https://opendatalab.com/), for better download and use experience，here is an example: [DSDLReadme](../../../configs/dsdl/README.md)， welcome to try.
 
 ```none
 mmsegmentation

diff --git a/docs/zh_cn/user_guides/2_dataset_prepare.md b/docs/zh_cn/user_guides/2_dataset_prepare.md
@@ -2,6 +2,7 @@
 
 我们建议将数据集根目录符号链接到 `$MMSEGMENTATION/data`。
 如果您的目录结构不同，您可能需要更改配置文件中相应的路径。
+对于中国境内的用户，我们也推荐通过开源数据平台 [OpenDataLab](https://opendatalab.com/) 来下载dsdl标准数据，以获得更好的下载和使用体验，这里有一个下载dsdl数据集并进行训练的案例[DSDLReadme](../../../configs/dsdl/README.md)，欢迎尝试。
 
 ```none
 mmsegmentation

diff --git a/mmseg/datasets/__init__.py b/mmseg/datasets/__init__.py
@@ -9,6 +9,7 @@
 from .dataset_wrappers import MultiImageMixDataset
 from .decathlon import DecathlonDataset
 from .drive import DRIVEDataset
+from .dsdl import DSDLSegDataset
 from .hrf import HRFDataset
 from .isaid import iSAIDDataset
 from .isprs import ISPRSDataset
@@ -54,7 +55,11 @@
     'BioMedicalGaussianNoise', 'BioMedicalGaussianBlur',
     'BioMedicalRandomGamma', 'BioMedical3DPad', 'RandomRotFlip',
     'SynapseDataset', 'REFUGEDataset', 'MapillaryDataset_v1',
+<<<<<<< HEAD
     'MapillaryDataset_v2', 'Albu', 'LEVIRCDDataset',
     'LoadMultipleRSImageFromFile', 'LoadSingleRSImageFromFile',
     'ConcatCDInput', 'BaseCDDataset'
+=======
+    'MapillaryDataset_v2', 'DSDLSegDataset'
+>>>>>>> [Feature] Support DSDL Dataset
 ]