[Doc] Add explanation and usage instructions for data configuration (#…

…1548)

* [WIP] Data configuration

* [Doc] Add data configuration

* version info

* grammar

* typo

* typo

* format

* fix based on comments

* grammar

* comments

docs/en/tutorials/customize_datasets.md

@@ -1,5 +1,84 @@
 # Tutorial 2: Customize Datasets
 
+## Data configuration
+
+`data` in config file is the variable for data configuration, to define the arguments that are used in datasets and dataloaders.
+
+Here is an example of data configuration:
+
+```python
+data = dict(
+    samples_per_gpu=4,
+    workers_per_gpu=4,
+    train=dict(
+        type='ADE20KDataset',
+        data_root='data/ade/ADEChallengeData2016',
+        img_dir='images/training',
+        ann_dir='annotations/training',
+        pipeline=train_pipeline),
+    val=dict(
+        type='ADE20KDataset',
+        data_root='data/ade/ADEChallengeData2016',
+        img_dir='images/validation',
+        ann_dir='annotations/validation',
+        pipeline=test_pipeline),
+    test=dict(
+        type='ADE20KDataset',
+        data_root='data/ade/ADEChallengeData2016',
+        img_dir='images/validation',
+        ann_dir='annotations/validation',
+        pipeline=test_pipeline))
+```
+
+- `train`, `val` and `test`: The [`config`](https://github.com/open-mmlab/mmcv/blob/master/docs/en/understand_mmcv/config.md)s to build dataset instances for model training, validation and testing by
+using [`build and registry`](https://github.com/open-mmlab/mmcv/blob/master/docs/en/understand_mmcv/registry.md) mechanism.
+
+- `samples_per_gpu`: How many samples per batch and per gpu to load during model training, and the `batch_size` of training is equal to `samples_per_gpu` times gpu number, e.g. when using 8 gpus for distributed data parallel trainig and `samples_per_gpu=4`, the `batch_size` is `8*4=16`.
+If you would like to define `batch_size` for testing and validation, please use `test_dataloaser` and
+`val_dataloader` with mmseg >=0.24.1.
+
+- `workers_per_gpu`: How many subprocesses per gpu to use for data loading. `0` means that the data will be loaded in the main process.
+
+**Note:** `samples_per_gpu` only works for model training, and the default setting of `samples_per_gpu` is 1 in mmseg when model testing and validation (DO NOT support batch inference yet).
+
+**Note:** before v0.24.1, except `train`, `val` `test`, `samples_per_gpu` and `workers_per_gpu`, the other keys in `data` must be the
+input keyword arguments for `dataloader` in pytorch, and the dataloaders used for model training, validation and testing have the same input arguments.
+In v0.24.1, mmseg supports to use `train_dataloader`, `test_dataloaser` and `val_dataloader` to specify different keyword arguments, and still supports the overall arguments definition but the specific dataloader setting has a higher priority.
+
+Here is an example for specific dataloader:
+
+```python
+data = dict(
+    samples_per_gpu=4,
+    workers_per_gpu=4,
+    shuffle=True,
+    train=dict(type='xxx', ...),
+    val=dict(type='xxx', ...),
+    test=dict(type='xxx', ...),
+    # Use different batch size during validation and testing.
+    val_dataloader=dict(samples_per_gpu=1, workers_per_gpu=4, shuffle=False),
+    test_dataloader=dict(samples_per_gpu=1, workers_per_gpu=4, shuffle=False))
+```
+
+Assume only one gpu used for model training and testing, as the priority of the overall arguments definition is low, the batch_size
+for training is `4` and dataset will be shuffled, and batch_size for testing and validation is `1`, and dataset will not be shuffled.
+
+To make data configuration much clearer, we recommend use specific dataloader setting instead of overall dataloader setting after v0.24.1, just like:
+
+```python
+data = dict(
+    train=dict(type='xxx', ...),
+    val=dict(type='xxx', ...),
+    test=dict(type='xxx', ...),
+    # Use specific dataloader setting
+    train_dataloader=dict(samples_per_gpu=4, workers_per_gpu=4, shuffle=True),
+    val_dataloader=dict(samples_per_gpu=1, workers_per_gpu=4, shuffle=False),
+    test_dataloader=dict(samples_per_gpu=1, workers_per_gpu=4, shuffle=False))
+```
+
+**Note:** in model training, default values in the script of mmseg for dataloader are `shuffle=True, and drop_last=True`,
+in model validation and testing, default values are `shuffle=False, and drop_last=False`
+
 ## Customize datasets by reorganizing data
 
 The simplest way is to convert your dataset to organize your data into folders.