How to train custom data?

[FrankFang0813]

Hi! @WalBouss
I use 'swin_small_patch4_window7_224.pth' pretrained model to train my data.
It can run ,but I got acc = 0 and loss = 0, like this picture.

I want to ask how to train custom data.
where I should modify?

thank you!

[WalBouss]

Hi,

To train on a custom dataset, first, you must organize and register your dataset by following MMsegmentation guidelines. But since the guidelines are not that clear, here are quick steps to follow:

1. arrange your custom dataset in the following manner

│   ├── my_dataset
│   │   ├── img_dir
│   │   │   ├── train
│   │   │   │   ├── xxx{img_suffix}
│   │   │   │   ├── yyy{img_suffix}
│   │   │   │   ├── zzz{img_suffix}
│   │   │   ├── val
│   │   ├── ann_dir
│   │   │   ├── train
│   │   │   │   ├── xxx{seg_map_suffix}
│   │   │   │   ├── yyy{seg_map_suffix}
│   │   │   │   ├── zzz{seg_map_suffix}
│   │   │   ├── val

2. Create a file under mmseg/datasets/my_dataset.py, and specify the classes and the palette used for the annotations (get inspiration from other datasets already implemented).

from .builder import DATASETS
from .custom import CustomDataset

@DATASETS.register_module()
class MyDataset(CustomDataset):
   """My Custom dataset """
   CLASSES = (
       'class0', 'class1', 'class2')

   PALETTE = [[0, 192, 64], [0, 192, 64], [0, 64, 96]] # these are just random numbers between 0 and 255

   def __init__(self, **kwargs):
       super(MyDataset, self).__init__(
           img_suffix='.jpg', seg_map_suffix='_labelTrainIds.png', **kwargs)

3. Register your dataset in mmseg/dataset/init.py
4. Create a config for your dataset in senformer_config/base/datasets ( I would advise you to copy and adapt the config file of ade20k).
5. Create a config file for your model that inherits from senformer_config/base/datasets/my_dataset.py. It should look to something like this:

_base_ = [
    '../../_base_/models/senformer_swin.py', '../../_base_/datasets/my_dataset.py', # <--- change name here
    '../../_base_/default_runtime.py', '../../_base_/schedules/schedule_160k.py'
]
model = dict(
    pretrained='pretrain/swin_tiny_patch4_window7_224.pth',
    backbone=dict(
        embed_dims=96,
        depths=[2, 2, 6, 2],
        num_heads=[3, 6, 12, 24],
        window_size=7,
        use_abs_pos_embed=False,
        drop_path_rate=0.3,
        patch_norm=True),
    neck=dict(
        type='FPNT',
        in_channels=[96, 192, 384, 768],
        out_channels=512,
        num_outs=4,
        depth_swin=1,
        num_heads=8),
    decode_head=dict(
        type='SenFormer',
        num_heads=8,
        branch_depth=6,
        in_channels=[512, 512, 512, 512],
        in_index=[0, 1, 2, 3],
        feature_strides=[4, 8, 16, 32],
        channels=128,  # not used
        dropout_ratio=0.1,
        num_classes=150, # <--- change number of classes
        align_corners=False),
    auxiliary_head=dict(in_channels=384, num_classes=150)) # <--- change number of classes

# AdamW optimizer, no weight decay for position embedding & layer norm in backbone
optimizer = dict(_delete_=True, type='AdamW', lr=0.00006, betas=(0.9, 0.999), weight_decay=0.01,
                 paramwise_cfg=dict(custom_keys={'absolute_pos_embed': dict(decay_mult=0.),
                                                 'relative_position_bias_table': dict(decay_mult=0.),
                                                 'norm': dict(decay_mult=0.),
                                                 'queries': dict(decay_mult=0.),
                                                 }))
optimizer_config = dict(_delete_=True, grad_clip=dict(max_norm=0.1, norm_type=2))

lr_config = dict(_delete_=True, policy='poly',
                 warmup='linear',
                 warmup_iters=1500,
                 warmup_ratio=1e-6,
                 power=1.0, min_lr=0.0, by_epoch=False)

# By default, models are trained on 8 GPUs with 2 images per GPU or 4 GPUs with 4 images per GPU
data=dict(samples_per_gpu=4)

Also, if your dataset does not contain many images, I recommend you start from a pretrained model (not just a pretrained backbone). You can download SwinTiny-SenFormer pretrained on ADE20K here. And note that the default learning rate and training schedule is for a batch size of 16 so you may have to adapt them if you change the batch size.

Hope that helps.

[FrankFang0813]

@WalBouss Thank you!
But I got a issue when I change num_classes of decode_head.
I use pretrained model,because I don't have many pictures.

_base_ = [
    '../../_base_/models/senformer_swin.py', '../../_base_/datasets/mydataset.py',
    '../../_base_/default_runtime.py', '../../_base_/schedules/schedule_160k.py'
]

model = dict(
    pretrained='pretrain/swin_tiny_patch4_window7_224.pth',
    backbone=dict(
        embed_dims=96,
        depths=[2, 2, 6, 2],
        num_heads=[3, 6, 12, 24],
        window_size=7,
        use_abs_pos_embed=False,
        drop_path_rate=0.3,
        patch_norm=True),
    neck=dict(
        type='FPNT',
        in_channels=[96, 192, 384, 768],
        out_channels=512,
        num_outs=4,
        depth_swin=1,
        num_heads=8),
    decode_head=dict(
        type='SenFormer',
        num_heads=8,
        branch_depth=6,
        in_channels=[512, 512, 512, 512],
        in_index=[0, 1, 2, 3],
        feature_strides=[4, 8, 16, 32],
        channels=128,  # not used
        dropout_ratio=0.1,
        num_classes=1, # I change here
        align_corners=False),
    auxiliary_head=dict(in_channels=384, num_classes=1)) # I change here

# AdamW optimizer, no weight decay for position embedding & layer norm in backbone
optimizer = dict(_delete_=True, type='AdamW', lr=0.00006, betas=(0.9, 0.999), weight_decay=0.01,
                 paramwise_cfg=dict(custom_keys={'absolute_pos_embed': dict(decay_mult=0.),
                                                 'relative_position_bias_table': dict(decay_mult=0.),
                                                 'norm': dict(decay_mult=0.),
                                                 'queries': dict(decay_mult=0.),
                                                 }))
optimizer_config = dict(_delete_=True, grad_clip=dict(max_norm=0.1, norm_type=2))

lr_config = dict(_delete_=True, policy='poly',
                 warmup='linear',
                 warmup_iters=1500,
                 warmup_ratio=1e-6,
                 power=1.0, min_lr=0.0, by_epoch=False)

# By default, models are trained on 8 GPUs with 2 images per GPU or 4 GPUs with 4 images per GPU
data=dict(samples_per_gpu=1)

Is this env problem? my torch == 1.10.1 and cuda == 10.2
I train ade20k datasets is no problem,but I change num_classes of decode_head got the issue.

[WalBouss]

Hi Frank,

I suspect that your dataset is not well registered, could you please share the files related to your own dataset implementation, namely mmseg/datasets/mydataset.py and senformer_config/base/datasets/mydataset.py.

Furthermore, you should have at least 2 classes (background + class_1).

[FrankFang0813]

Hi! ! @WalBouss
Thank you! I already trained my custom data!
#2 (comment)
This issue is my classes just have background and class1,but my class1 in picture is [128,128,128] > 2, so I can't train before.then I use opencv fillpoly with color [1,1,1],it can train!

[WalBouss]

Glad it helped you!
I am closing this issue; don't hesitate to reopen it if you need.