mirror of
https://github.com/open-mmlab/mmclassification.git
synced 2025-06-03 21:53:55 +08:00
231 lines
7.9 KiB
Markdown
231 lines
7.9 KiB
Markdown
|
# 教程 4: 使用自定义数据集进行预训练
|
||
|
|
|
||
|
|
- [教程 4: 使用自定义数据集进行预训练](#教程-4-使用自定义数据集进行预训练)
|
||
|
|
- [在自定义数据集上使用 MAE 算法进行预训练](#在自定义数据集上使用-mae-算法进行预训练)
|
||
|
|
- [第一步：获取自定义数据路径](#第一步获取自定义数据路径)
|
||
|
|
- [第二步：选择一个配置文件作为模板](#第二步选择一个配置文件作为模板)
|
||
|
|
- [第三步：修改数据集相关的配置](#第三步修改数据集相关的配置)
|
||
|
|
- [在 COCO 数据集上使用 MAE 算法进行预训练](#在-coco-数据集上使用-mae-算法进行预训练)
|
||
|
|
|
||
|
|
在本教程中，我们将介绍如何使用自定义数据集(无需标注)进行自监督预训练。
|
||
|
|
|
||
|
|
## 在自定义数据集上使用 MAE 算法进行预训练
|
||
|
|
|
||
|
|
在 MMPretrain 中, 我们支持用户直接调用 MMPretrain 的 `CustomDataset` (类似于 `torchvision` 的 `ImageFolder`), 该数据集能自动的读取给的路径下的图片。你只需要准备你的数据集路径，并修改配置文件，即可轻松使用 MMPretrain 进行预训练。
|
||
|
|
|
||
|
|
### 第一步：获取自定义数据路径
|
||
|
|
|
||
|
|
路径应类似这种形式： `data/custom_dataset/`
|
||
|
|
|
||
|
|
### 第二步：选择一个配置文件作为模板
|
||
|
|
|
||
|
|
在本教程中，我们使用 `configs/selfsup/mae/mae_vit-base-p16_8xb512-coslr-400e_in1k.py`作为一个示例进行讲解。我们首先复制这个配置文件，将新复制的文件命名为`mae_vit-base-p16_8xb512-coslr-400e_${custom_dataset}.py`.
|
||
|
|
|
||
|
|
- `custom_dataset`: 表明你用的那个数据集。例如，用 `in1k` 代表ImageNet 数据集，`coco` 代表COCO数据集。
|
||
|
|
|
||
|
|
这个配置文件的内容如下：
|
||
|
|
|
||
|
|
```python
|
||
|
|
_base_ = [
|
||
|
|
'../_base_/models/mae_vit-base-p16.py',
|
||
|
|
'../_base_/datasets/imagenet_bs512_mae.py',
|
||
|
|
'../_base_/default_runtime.py',
|
||
|
|
]
|
||
|
|
# optimizer wrapper
|
||
|
|
optim_wrapper = dict(
|
||
|
|
type='AmpOptimWrapper',
|
||
|
|
loss_scale='dynamic',
|
||
|
|
optimizer=dict(
|
||
|
|
type='AdamW',
|
||
|
|
lr=1.5e-4 * 4096 / 256,
|
||
|
|
betas=(0.9, 0.95),
|
||
|
|
weight_decay=0.05),
|
||
|
|
paramwise_cfg=dict(
|
||
|
|
custom_keys={
|
||
|
|
'ln': dict(decay_mult=0.0),
|
||
|
|
'bias': dict(decay_mult=0.0),
|
||
|
|
'pos_embed': dict(decay_mult=0.),
|
||
|
|
'mask_token': dict(decay_mult=0.),
|
||
|
|
'cls_token': dict(decay_mult=0.)
|
||
|
|
}))
|
||
|
|
# learning rate scheduler
|
||
|
|
param_scheduler = [
|
||
|
|
dict(
|
||
|
|
type='LinearLR',
|
||
|
|
start_factor=0.0001,
|
||
|
|
by_epoch=True,
|
||
|
|
begin=0,
|
||
|
|
end=40,
|
||
|
|
convert_to_iter_based=True),
|
||
|
|
dict(
|
||
|
|
type='CosineAnnealingLR',
|
||
|
|
T_max=260,
|
||
|
|
by_epoch=True,
|
||
|
|
begin=40,
|
||
|
|
end=300,
|
||
|
|
convert_to_iter_based=True)
|
||
|
|
]
|
||
|
|
# runtime settings
|
||
|
|
train_cfg = dict(type='EpochBasedTrainLoop', max_epochs=300)
|
||
|
|
default_hooks = dict(
|
||
|
|
# only keeps the latest 3 checkpoints
|
||
|
|
checkpoint=dict(type='CheckpointHook', interval=1, max_keep_ckpts=3))
|
||
|
|
randomness = dict(seed=0, diff_rank_seed=True)
|
||
|
|
# auto resume
|
||
|
|
resume = True
|
||
|
|
# NOTE: `auto_scale_lr` is for automatically scaling LR
|
||
|
|
# based on the actual training batch size.
|
||
|
|
auto_scale_lr = dict(base_batch_size=4096)
|
||
|
|
```
|
||
|
|
|
||
|
|
### 第三步：修改数据集相关的配置
|
||
|
|
|
||
|
|
数据集相关的配置是定义在 `_base_`的`'../_base_/datasets/imagenet_mae.py'` 文件内。我们直接将其内容复制到刚刚创建的新的配置文件 `mae_vit-base-p16_8xb512-coslr-400e_${custom_dataset}.py` 中.
|
||
|
|
|
||
|
|
- 修改`dataset_type = 'CustomDataset'`和` data_root = /dataset/my_custom_dataset`.
|
||
|
|
- 删除 `train_dataloader`中的 `ann_file` ，同时根据自己的实际情况决定是否需要设定 `data_prefix`。
|
||
|
|
|
||
|
|
```{note}
|
||
|
|
`CustomDataset` 是在 MMPretrain 实现的, 因此我们使用这种方式 `dataset_type=CustomDataset` 来使用这个类。
|
||
|
|
```
|
||
|
|
|
||
|
|
此时，修改后的文件应如下：
|
||
|
|
|
||
|
|
```python
|
||
|
|
# >>>>>>>>>>>>>>>>>>>>> Start of Changed >>>>>>>>>>>>>>>>>>>>>>>>>
|
||
|
|
_base_ = [
|
||
|
|
'../_base_/models/mae_vit-base-p16.py',
|
||
|
|
'../_base_/datasets/imagenet_mae.py',
|
||
|
|
'../_base_/default_runtime.py',
|
||
|
|
]
|
||
|
|
# custom dataset
|
||
|
|
dataset_type = 'CustomDataset'
|
||
|
|
data_root = 'data/custom_dataset/'
|
||
|
|
train_dataloader = dict(
|
||
|
|
dataset=dict(
|
||
|
|
type=dataset_type,
|
||
|
|
data_root=data_root,
|
||
|
|
# ann_file='meta/train.txt', # removed if you don't have the annotation file
|
||
|
|
data_prefix=dict(img_path='./'))
|
||
|
|
# <<<<<<<<<<<<<<<<<<<<<< End of Changed <<<<<<<<<<<<<<<<<<<<<<<<<<<
|
||
|
|
# optimizer wrapper
|
||
|
|
optim_wrapper = dict(
|
||
|
|
type='AmpOptimWrapper',
|
||
|
|
loss_scale='dynamic',
|
||
|
|
optimizer=dict(
|
||
|
|
type='AdamW',
|
||
|
|
lr=1.5e-4 * 4096 / 256,
|
||
|
|
betas=(0.9, 0.95),
|
||
|
|
weight_decay=0.05),
|
||
|
|
paramwise_cfg=dict(
|
||
|
|
custom_keys={
|
||
|
|
'ln': dict(decay_mult=0.0),
|
||
|
|
'bias': dict(decay_mult=0.0),
|
||
|
|
'pos_embed': dict(decay_mult=0.),
|
||
|
|
'mask_token': dict(decay_mult=0.),
|
||
|
|
'cls_token': dict(decay_mult=0.)
|
||
|
|
}))
|
||
|
|
# learning rate scheduler
|
||
|
|
param_scheduler = [
|
||
|
|
dict(
|
||
|
|
type='LinearLR',
|
||
|
|
start_factor=0.0001,
|
||
|
|
by_epoch=True,
|
||
|
|
begin=0,
|
||
|
|
end=40,
|
||
|
|
convert_to_iter_based=True),
|
||
|
|
dict(
|
||
|
|
type='CosineAnnealingLR',
|
||
|
|
T_max=260,
|
||
|
|
by_epoch=True,
|
||
|
|
begin=40,
|
||
|
|
end=300,
|
||
|
|
convert_to_iter_based=True)
|
||
|
|
]
|
||
|
|
# runtime settings
|
||
|
|
train_cfg = dict(type='EpochBasedTrainLoop', max_epochs=300)
|
||
|
|
default_hooks = dict(
|
||
|
|
# only keeps the latest 3 checkpoints
|
||
|
|
checkpoint=dict(type='CheckpointHook', interval=1, max_keep_ckpts=3))
|
||
|
|
randomness = dict(seed=0, diff_rank_seed=True)
|
||
|
|
# auto resume
|
||
|
|
resume = True
|
||
|
|
# NOTE: `auto_scale_lr` is for automatically scaling LR
|
||
|
|
# based on the actual training batch size.
|
||
|
|
auto_scale_lr = dict(base_batch_size=4096)
|
||
|
|
```
|
||
|
|
|
||
|
|
使用上述配置文件，你就能够轻松的在自定义数据集上使用 `MAE` 算法来进行预训练了。
|
||
|
|
|
||
|
|
## 在 COCO 数据集上使用 MAE 算法进行预训练
|
||
|
|
|
||
|
|
```{note}
|
||
|
|
你可能需要参考[文档](https://github.com/open-mmlab/mmdetection/blob/3.x/docs/en/get_started.md)安装 MMDetection 来使用 `mmdet.CocoDataset`。
|
||
|
|
```
|
||
|
|
|
||
|
|
与在自定义数据集上进行预训练类似，我们在本教程中也提供了一个使用 COCO 数据集进行预训练的示例。修改后的文件如下：
|
||
|
|
|
||
|
|
```python
|
||
|
|
# >>>>>>>>>>>>>>>>>>>>> Start of Changed >>>>>>>>>>>>>>>>>>>>>>>>>
|
||
|
|
_base_ = [
|
||
|
|
'../_base_/models/mae_vit-base-p16.py',
|
||
|
|
'../_base_/datasets/imagenet_mae.py',
|
||
|
|
'../_base_/default_runtime.py',
|
||
|
|
]
|
||
|
|
# custom dataset
|
||
|
|
dataset_type = 'mmdet.CocoDataset'
|
||
|
|
data_root = 'data/coco/'
|
||
|
|
train_dataloader = dict(
|
||
|
|
dataset=dict(
|
||
|
|
type=dataset_type,
|
||
|
|
data_root=data_root,
|
||
|
|
ann_file='annotations/instances_train2017.json',
|
||
|
|
data_prefix=dict(img='train2017/')))
|
||
|
|
# <<<<<<<<<<<<<<<<<<<<<< End of Changed <<<<<<<<<<<<<<<<<<<<<<<<<<<
|
||
|
|
# optimizer wrapper
|
||
|
|
optim_wrapper = dict(
|
||
|
|
type='AmpOptimWrapper',
|
||
|
|
loss_scale='dynamic',
|
||
|
|
optimizer=dict(
|
||
|
|
type='AdamW',
|
||
|
|
lr=1.5e-4 * 4096 / 256,
|
||
|
|
betas=(0.9, 0.95),
|
||
|
|
weight_decay=0.05),
|
||
|
|
paramwise_cfg=dict(
|
||
|
|
custom_keys={
|
||
|
|
'ln': dict(decay_mult=0.0),
|
||
|
|
'bias': dict(decay_mult=0.0),
|
||
|
|
'pos_embed': dict(decay_mult=0.),
|
||
|
|
'mask_token': dict(decay_mult=0.),
|
||
|
|
'cls_token': dict(decay_mult=0.)
|
||
|
|
}))
|
||
|
|
# learning rate scheduler
|
||
|
|
param_scheduler = [
|
||
|
|
dict(
|
||
|
|
type='LinearLR',
|
||
|
|
start_factor=0.0001,
|
||
|
|
by_epoch=True,
|
||
|
|
begin=0,
|
||
|
|
end=40,
|
||
|
|
convert_to_iter_based=True),
|
||
|
|
dict(
|
||
|
|
type='CosineAnnealingLR',
|
||
|
|
T_max=260,
|
||
|
|
by_epoch=True,
|
||
|
|
begin=40,
|
||
|
|
end=300,
|
||
|
|
convert_to_iter_based=True)
|
||
|
|
]
|
||
|
|
# runtime settings
|
||
|
|
train_cfg = dict(type='EpochBasedTrainLoop', max_epochs=300)
|
||
|
|
default_hooks = dict(
|
||
|
|
# only keeps the latest 3 checkpoints
|
||
|
|
checkpoint=dict(type='CheckpointHook', interval=1, max_keep_ckpts=3))
|
||
|
|
randomness = dict(seed=0, diff_rank_seed=True)
|
||
|
|
# auto resume
|
||
|
|
resume = True
|
||
|
|
# NOTE: `auto_scale_lr` is for automatically scaling LR
|
||
|
|
# based on the actual training batch size.
|
||
|
|
auto_scale_lr = dict(base_batch_size=4096)
|
||
|
|
```
|