Tutorials

In this chapter, we will provide detailed tutorials to help the users learn how to use SSSegmentation.

Learn about Config

We incorporate modular design into our config system, which is convenient to conduct various experiments.

Config File Structure

Now, there are 2 basic component types under ssseg/configs/_base_, i.e., datasets and dataloaders, which are responsible for defining configs of various datasets with different runtime settings (e.g., batch size, image size, data augmentation, to name a few).

For example, to train FCN segmentor on Pascal VOC dataset (assuming that we hope the total batch size is 16 and the training image size is 512x512), you can import the corresponding pre-defined config like this,

from .._base_ import DATASET_CFG_VOCAUG_512x512, DATALOADER_CFG_BS16

Then, modify SEGMENTOR_CFG in the corresponding method config file (e.g., ssseg/configs/fcn/fcn_resnet50os16_voc.py) as follows,

# modify dataset config
SEGMENTOR_CFG['dataset'] = DATASET_CFG_VOCAUG_512x512.copy()
# modify dataloader config
SEGMENTOR_CFG['dataloader'] = DATALOADER_CFG_BS16.copy()

From this, you are not required to define the same dataloader and dataset configs over and over again if you just want to follow the conventional training and testing settings of various datasets to train your segmentors.

Next, we talk about config files in specific segmentation algorithm directories (e.g., ssseg/configs/fcn directory). There is also one and only one base_cfg.py used to define some necessary configs for the corresponding algorithm (e.g., ssseg/configs/fcn/base_cfg.py is used to define the basic configs for FCN segmentation algorithm). Like loading configs from ssseg/configs/_base_, you can also import the base_cfg.py and simply modify some key values in SEGMENTOR_CFG to customize and train the corresponding segmentor.

For instance, to customize FCN with ResNet-50-D16 backbone and train it on Pascal VOC dataset, you can create a config file in ssseg/configs/fcn directory, named fcn_resnet50os16_voc.py and write in some contents like this,

import copy
from .base_cfg import SEGMENTOR_CFG
from .._base_ import DATASET_CFG_VOCAUG_512x512, DATALOADER_CFG_BS16

# deepcopy
SEGMENTOR_CFG = copy.deepcopy(SEGMENTOR_CFG)
# modify dataset config
SEGMENTOR_CFG['dataset'] = DATASET_CFG_VOCAUG_512x512.copy()
# modify dataloader config
SEGMENTOR_CFG['dataloader'] = DATALOADER_CFG_BS16.copy()
# modify scheduler config
SEGMENTOR_CFG['scheduler']['max_epochs'] = 60
# modify other segmentor configs
SEGMENTOR_CFG['num_classes'] = 21
SEGMENTOR_CFG['backbone'] = {
    'type': 'ResNet', 'depth': 50, 'structure_type': 'resnet50conv3x3stem',
    'pretrained': True, 'outstride': 16, 'use_conv3x3_stem': True, 'selected_indices': (2, 3),
}
SEGMENTOR_CFG['work_dir'] = 'fcn_resnet50os16_voc'
SEGMENTOR_CFG['logfilepath'] = 'fcn_resnet50os16_voc/fcn_resnet50os16_voc.log'
SEGMENTOR_CFG['resultsavepath'] = 'fcn_resnet50os16_voc/fcn_resnet50os16_voc_results.pkl'

After that, you can train this segmentor with the following command,

bash scripts/dist_train.sh 4 ssseg/configs/fcn/fcn_resnet50os16_voc.py

How relaxing and enjoyable! Maybe, you can read more config examples in ssseg/configs to help you learn about how to create a valid config file in SSSegmentation.

An Example of PSPNet

To help the users have a basic idea of a complete config and the modules in SSSegmentation, we make brief comments on the config of PSPNet using ResNet-101-D8 as the following,

import os

# dataset configs
DATASET_CFG_ADE20k_512x512 = {
    'type': 'ADE20kDataset', # the dataset type used to instance the specific dataset class in builder.py defined in "ssseg/modules/datasets/builder.py"
    'rootdir': os.path.join(os.getcwd(), 'ADE20k'), # the corresponding dataset path you download
    'train': {
        'set': 'train', # train the models with the train set defined in the dataset
        'data_pipelines': [
            ('Resize', {'output_size': (2048, 512), 'keep_ratio': True, 'scale_range': (0.5, 2.0)}),
            ('RandomCrop', {'crop_size': (512, 512), 'one_category_max_ratio': 0.75}),
            ('RandomFlip', {'flip_prob': 0.5}),
            ('PhotoMetricDistortion', {}),
            ('Normalize', {'mean': [123.675, 116.28, 103.53], 'std': [58.395, 57.12, 57.375]}),
            ('ToTensor', {}),
            ('Padding', {'output_size': (512, 512), 'data_type': 'tensor'}),
        ], # define some pre-processing operations for the loaded images and segmentation targets during training, you can refer to "ssseg/modules/datasets/pipelines" for more details
    },
    'test': {
        'set': 'val', # test the models with the validation set defined in the dataset
        'data_pipelines': [
            ('Resize', {'output_size': (2048, 512), 'keep_ratio': True, 'scale_range': None}),
            ('Normalize', {'mean': [123.675, 116.28, 103.53], 'std': [58.395, 57.12, 57.375]}),
            ('ToTensor', {}),
        ], # define some pre-processing operations for the loaded images and segmentation targets during testing, you can refer to "ssseg/modules/datasets/pipelines" for more details
    }
}
DATALOADER_CFG_BS16 = {
    'expected_total_train_bs_for_assert': 16, # it is defined for asserting whether the users adopt the correct batch size for training the models
	'auto_adapt_to_expected_train_bs': True, # if set Ture, the "expected_total_train_bs_for_assert" will be used to determine the value of "batch_size_per_gpu" rather than using the specified value in the config
    'train': {
        'batch_size_per_gpu': 2, # number of images in each gpu during training
        'num_workers_per_gpu': 2, # number of workers for dataloader in each gpu during training
        'shuffle': True, # whether to shuffle the image order during training
        'pin_memory': True, # whether to achieve higher bandwidth between the host and the device using pinned memory during training
        'drop_last': True, # whether to drop out the last images which cannot form a batch size of "expected_total_train_bs_for_assert" during training
    },
    'test': {
        'batch_size_per_gpu': 1, # number of images in each gpu during testing
        'num_workers_per_gpu': 2, # number of workers for dataloader in each gpu during testing
        'shuffle': False, # whether to shuffle the image order during testing
        'pin_memory': True, # whether to achieve higher bandwidth between the host and the device using pinned memory during testing
        'drop_last': False, # whether to drop out the last images which cannot form a batch size of "expected_total_train_bs_for_assert" during testing
    }
}
SEGMENTOR_CFG = {
    'type': 'PSPNet', # the segmentor type defined in "ssseg/modules/models/segmentors/builder.py"
    'num_classes': -1, # number of classes in the dataset
    'benchmark': True, # set True for speeding up training
    'align_corners': False, # align_corners in torch.nn.functional.interpolate
    'backend': 'nccl', # backend for DDP training and testing
    'work_dir': 'ckpts', # directory used to save checkpoints and training and testing logs
    'eval_interval_epochs': 10, # evaluate models after "eval_interval_epochs" epochs
    'save_interval_epochs': 1, # save the checkpoints of models after "save_interval_epochs" epochs
    'evaluate_results_filename': '', # path used to save the testing results used in `ssseg/test.py`
    'logger_handle_cfg': {'type': 'LocalLoggerHandle', 'logfilepath': ''}, # config used to instance logger_handle, which is used to print and save logs
    'training_logging_manager_cfg': {'log_interval_iters': 50}, # config used to instance training_logging_manager, which is used to manage training logs
    'norm_cfg': {'type': 'SyncBatchNorm'}, # config for normalization layer in the segmentor
    'act_cfg': {'type': 'ReLU', 'inplace': True}, # config for activation layer in the segmentor
    'backbone': {
        'type': 'ResNet', 'depth': 101, 'structure_type': 'resnet101conv3x3stem',
        'pretrained': True, 'outstride': 8, 'use_conv3x3_stem': True, 'selected_indices': (2, 3),
    }, # define backbones, refer to "ssseg/modules/models/backbones/builder.py" for more details
    'head': {
        'in_channels': 2048, 'feats_channels': 512, 'pool_scales': [1, 2, 3, 6], 'dropout': 0.1,
    }, # define decoder heads, refer to "ssseg/modules/models/segmentors" for more details
    'auxiliary': {
        'in_channels': 1024, 'out_channels': 512, 'dropout': 0.1,
    }, # define auxiliary decoder, refer to "ssseg/modules/models/segmentors/base/base.py" for more details
    'losses': {
        'loss_aux': {'type': 'CrossEntropyLoss', 'scale_factor': 0.4, 'ignore_index': 255, 'reduction': 'mean'},
        'loss_cls': {'type': 'CrossEntropyLoss', 'scale_factor': 1.0, 'ignore_index': 255, 'reduction': 'mean'},
    }, # define objective functions, refer to "ssseg/modules/models/losses/builder.py" for more details
    'inference': {
        'mode': 'whole',
        'opts': {}, 
        'tricks': {
            'multiscale': [1], 'flip': False, 'use_probs_before_resize': False,
        }
    }, # define the inference config, refer to "ssseg/test.py" for more details
    'scheduler': {
        'type': 'PolyScheduler', 'max_epochs': 0, 'power': 0.9,
        'optimizer': {
            'type': 'SGD', 'lr': 0.01, 'momentum': 0.9, 'weight_decay': 5e-4, 'params_rules': {},
        }
    }, # define the scheduler configs, refer to "ssseg/modules/models/schedulers/builder.py" for more details
    'dataset': DATASET_CFG_ADE20k_512x512, # define the dataset configs
    'dataloader': DATALOADER_CFG_BS16, # define the dataloader configs
}

In the following sections, we will give more explanations and examples about each module specified in the config above.

Customize Datasets

Dataset classes in SSSegmentation have two functions: (1) load data information after data preparation and (2) construct sample_meta for the subsequent segmentor training and testing.

The type of sample_meta is dict which includes several keys:

• image: The loaded image data,

• seg_target: The loaded ground truth segmentation mask data,

• width and height: The original size of the image (i.e., the image size before pre-processing by the data transforms),

• id: The image id of the loaded image data.

Thanks to the modular design in SSSegmentation, we can simply modify the configs in SEGMENTOR_CFG['dataset'] to train one segmentor on various datasets.

Dataset Config Structure

An example of dataset config is as follows,

import os

DATASET_CFG_ADE20k_512x512 = {
    'type': 'ADE20kDataset',
    'rootdir': os.path.join(os.getcwd(), 'ADE20k'),
    'train': {
        'set': 'train',
        'data_pipelines': [
            ('Resize', {'output_size': (2048, 512), 'keep_ratio': True, 'scale_range': (0.5, 2.0)}),
            ('RandomCrop', {'crop_size': (512, 512), 'one_category_max_ratio': 0.75}),
            ('RandomFlip', {'flip_prob': 0.5}),
            ('PhotoMetricDistortion', {}),
            ('Normalize', {'mean': [123.675, 116.28, 103.53], 'std': [58.395, 57.12, 57.375]}),
            ('ToTensor', {}),
            ('Padding', {'output_size': (512, 512), 'data_type': 'tensor'}),
        ],
    },
    'test': {
        'set': 'val',
        'data_pipelines': [
            ('Resize', {'output_size': (2048, 512), 'keep_ratio': True, 'scale_range': None}),
            ('Normalize', {'mean': [123.675, 116.28, 103.53], 'std': [58.395, 57.12, 57.375]}),
            ('ToTensor', {}),
        ],
    }
}

where type denotes the dataset you want to train on. Now, SSSegmentation supports the following dataset types,

REGISTERED_MODULES = {
    'BaseDataset': BaseDataset, 'VOCDataset': VOCDataset, 'PascalContext59Dataset': PascalContext59Dataset, 'PascalContextDataset': PascalContextDataset,
    'COCOVOCSUBDataset': COCOVOCSUBDataset, 'COCOStuff10kDataset': COCOStuff10kDataset, 'COCOStuffDataset': COCOStuffDataset, 'CIHPDataset': CIHPDataset,
    'LIPDataset': LIPDataset, 'ATRDataset': ATRDataset, 'MHPv1Dataset': MHPv1Dataset, 'MHPv2Dataset': MHPv2Dataset, 'SuperviselyDataset': SuperviselyDataset,
    'HRFDataset': HRFDataset, 'ChaseDB1Dataset': ChaseDB1Dataset, 'STAREDataset': STAREDataset, 'DRIVEDataset': DRIVEDataset, 'SBUShadowDataset': SBUShadowDataset,
    'VSPWDataset': VSPWDataset, 'ADE20kDataset': ADE20kDataset, 'DarkZurichDataset': DarkZurichDataset, 'NighttimeDrivingDataset': NighttimeDrivingDataset,
    'CityScapesDataset': CityScapesDataset, 'MultipleDataset': MultipleDataset,
}

The keyword rootdir is used to specify the data set path. It is recommended to symlink the dataset root to $SSSEGMENTATION/ or directly run bash scripts/prepare_datasets.sh $DATASETNAME in $SSSEGMENTATION/ so that you don’t need to modify the default rootdir for each dataset.

The keyword train and test are used to specify the configs for model training and testing. And the value type of SEGMENTOR_CFG['dataset']['train'] and SEGMENTOR_CFG['dataset']['test'] is dict. Specifically, set means a certain division of the data set, usually including train set (set as train), validation set (set as val) and test set (set as test). data_pipelines is used to define data transforms to pre-process sample_meta before feeding into the models, more details please refer to Customize Data Pipelines.

The other arguments supported in SEGMENTOR_CFG['dataset'] is listed as follows,

• repeat_times: The default value is 1, if increase it in the config, the appeared times of one image in an epoch will increase accordingly,

• evalmode: local or server, server denotes seg_target will be set as None.

If the users want to learn more about this part, it is recommended that you could jump to the ssseg/modules/datasets directory in SSSegmentation to read the source codes of dataset classes.

Customize Data Pipelines

Constructing data pipelines is used to preprocess the input data (e.g., images and segmentation masks) for the following training and testing of the segmentors.

Specifically, it is defined at,

• SEGMENTOR_CFG['dataset']['train']['data_pipelines']: The constructed data pipelines for training,

• SEGMENTOR_CFG['dataset']['test']['data_pipelines']: The constructed data pipelines for testing.

The value of the data_pipelines should be a list like following,

SEGMENTOR_CFG['dataset']['train']['data_pipelines'] = [
    ('Resize', {'output_size': (2048, 512), 'keep_ratio': True, 'scale_range': (0.5, 2.0)}),
    ('RandomCrop', {'crop_size': (512, 512), 'one_category_max_ratio': 0.75}),
    ('RandomFlip', {'flip_prob': 0.5}),
    ('PhotoMetricDistortion', {}),
    ('Normalize', {'mean': [123.675, 116.28, 103.53], 'std': [58.395, 57.12, 57.375]}),
    ('ToTensor', {}),
    ('Padding', {'output_size': (512, 512), 'data_type': 'tensor'}),
]

And each item in the list should be a tuple or dict. For example, it could be,

# tuple
('Resize', {'output_size': (2048, 512), 'keep_ratio': True, 'scale_range': (0.5, 2.0)})
# dict
{'type': 'Resize', 'output_size': (2048, 512), 'keep_ratio': True, 'scale_range': (0.5, 2.0)}

where Resize means a data transform method defined in ssseg/modules/datasets/pipelines/transforms.py and other values denote for the arguments for instancing the corresponding data transform method.

Here is a list of supported data transform methods,

REGISTERED_MODULES = {
    'Resize': Resize, 'RandomCrop': RandomCrop, 'RandomFlip': RandomFlip, 'RandomRotation': RandomRotation, 'EdgeExtractor': EdgeExtractor,
    'PhotoMetricDistortion': PhotoMetricDistortion, 'Padding': Padding, 'ToTensor': ToTensor, 'ResizeShortestEdge': ResizeShortestEdge,
    'Normalize': Normalize, 'RandomChoiceResize': RandomChoiceResize, 'Rerange': Rerange, 'CLAHE': CLAHE, 'RandomCutOut': RandomCutOut, 
    'AlbumentationsWrapper': AlbumentationsWrapper, 'RGB2Gray': RGB2Gray, 'AdjustGamma': AdjustGamma, 'RandomGaussianBlur': RandomGaussianBlur,
    'RandomShortestEdgeResize': RandomShortestEdgeResize, 'PILRandomGaussianBlur': PILRandomGaussianBlur,
}

To learn the functions of each data transform method, please check the source codes in ssseg/modules/datasets/pipelines/transforms.py by yourselves.

It is worth mentioning that SSSegmentation provides AlbumentationsWrapper to make the users leverage the data augmentation algorithms implemented in albumentations. Here is an example of calling AlbumentationsWrapper,

SEGMENTOR_CFG['dataset']['train']['data_pipelines'] = [
    ('AlbumentationsWrapper', {'albu_cfg': {'type': 'RandomCrop', 'width': 256, 'height': 256}}),
    ('AlbumentationsWrapper', {'albu_cfg': {'type': 'HorizontalFlip', 'p': 0.5}}),
    ('AlbumentationsWrapper', {'albu_cfg': {'type': 'RandomBrightnessContrast', 'p': 0.2}}),
]

Finally, if you want to define the data transform method by yourselves during developing, you can first write the transform method like following,

class RGB2Gray(object):
    def __init__(self, out_channels=None, weights=(0.299, 0.587, 0.114)):
        # assert
        assert isinstance(weights, collections.abc.Sequence)
        assert out_channels is None or out_channels > 0
        for item in weights: assert isinstance(item, (float, int))
        # set attributes
        self.weights = weights
        self.out_channels = out_channels
    '''call'''
    def __call__(self, sample_meta):
        sample_meta = self.rgb2gray('image', sample_meta, self.weights, self.out_channels)
        return sample_meta
    '''rgb2gray'''
    @staticmethod
    def rgb2gray(key, sample_meta, weights, out_channels):
        if key not in sample_meta: return sample_meta
        # assert
        assert len(sample_meta[key].shape) == 3
        assert sample_meta[key].shape[2] == len(weights)
        # apply
        weights = np.array(weights).reshape((1, 1, -1))
        sample_meta[key] = (sample_meta[key] * weights).sum(2, keepdims=True)
        if out_channels is None:
            sample_meta[key] = sample_meta[key].repeat(weights.shape[2], axis=2)
        else:
            sample_meta[key] = sample_meta[key].repeat(out_channels, axis=2)
        # return
        return sample_meta

Then, import DataTransformBuilder and register this method,

from ssseg.modules import DataTransformBuilder

data_transformer_builder = DataTransformBuilder()
data_transformer_builder.register('RGB2Gray', RGB2Gray)

From this, you can call data_transformer_builder.build to build your own defined transform algorithms as well as the original supported data transform methods.

Add New Custom Dataset

SSSegmentation provides BaseDataset class to help the users quickly add a new custom dataset.

Specifically, you can directly inherit this class to define your own dataset class. Here is an example code to add SuperviselyDataset dataset in SSSegmentation ssseg/modules/datasets directory by using BaseDataset,

import os
import pandas as pd
from .base import BaseDataset

'''SuperviselyDataset'''
class SuperviselyDataset(BaseDataset):
    num_classes = 2
    classnames = ['__background__', 'person']
    palette = [(0, 0, 0), (255, 0, 0)]
    clsid2label = {255: 1}
    assert num_classes == len(classnames) and num_classes == len(palette)
    def __init__(self, mode, logger_handle, dataset_cfg):
        super(SuperviselyDataset, self).__init__(mode=mode, logger_handle=logger_handle, dataset_cfg=dataset_cfg)
        # obtain the dirs
        rootdir = dataset_cfg['rootdir']
        self.image_dir = os.path.join(rootdir, 'Images', dataset_cfg['set'])
        self.ann_dir = os.path.join(rootdir, 'Anno-Person', dataset_cfg['set'])
        # obatin imageids
        df = pd.read_csv(os.path.join(rootdir, dataset_cfg['set']+'.txt'), names=['imageids'])
        self.imageids = df['imageids'].values
        self.imageids = [str(_id) for _id in self.imageids]

In __init__, you are required to set some attributes including:

• image_dir: Image file directory,

• ann_dir: Annotation file directory (i.e., the directory used to save ground truth segmentation masks),

• image_ext: Image file extension, default value is .jpg,

• ann_ext: Annotation file extension, default value is .png,

• imageids: The image file names.

Besides, clsid2label could be set for transferring the label ids in ground truth segmentation mask to continuous training ids automatically during training. For example, clsid2label = {10: 1} means if the class id in ground truth segmentation mask is 10, they will be set as class id 1 before feeding into the models. Other attributes classnames is used to set the class names in the dataset, palette is used to set the colors for each class during Inference with Segmentors Integrated in SSSegmentation and num_classes represents the number of the classes existed in the dataset.

After that, you should add this custom dataset class in ssseg/modules/datasets/builder.py if you want to use it by simply modifying SEGMENTOR_CFG['dataset']. Of course, you can also register this custom dataset by the following codes,

from ssseg.modules import DatasetBuilder

dataset_builder = DatasetBuilder()
dataset_builder.register('SuperviselyDataset', SuperviselyDataset)

From this, you can also call dataset_builder.build to build your own defined dataset class as well as the original supported data classes.

Finally, the users could jump to the ssseg/modules/datasets directory in SSSegmentation to read more source codes of the supported dataset classes and thus better learn how to customize the dataset classes in SSSegmentation.

Customize Backbones

Backbone is the image encoder that transforms an image to feature maps, such as a ResNet-50 without the last fully connected layer.

Backbone Config Structure

An example of backbone config is as follows,

SEGMENTOR_CFG['backbone'] = {
    'type': 'ResNet', 'depth': 101, 'structure_type': 'resnet101conv3x3stem',
    'pretrained': True, 'outstride': 16, 'use_conv3x3_stem': True, 'selected_indices': (2, 3),
}

where type denotes the backbone network you want to employ. Now, SSSegmentation supports the following backbone types,

REGISTERED_MODULES = {
    'UNet': UNet, 'BEiT': BEiT, 'CGNet': CGNet, 'HRNet': HRNet, 'MobileViT': MobileViT, 'MobileViTV2': MobileViTV2,
    'ERFNet': ERFNet, 'ResNet': ResNet, 'ResNeSt': ResNeSt, 'PCPVT': PCPVT, 'MobileSAMTinyViT': MobileSAMTinyViT, 
    'SVT': SVT, 'FastSCNN': FastSCNN, 'ConvNeXt': ConvNeXt, 'BiSeNetV1': BiSeNetV1, 'MAE': MAE, 'SAMViT': SAMViT,
    'BiSeNetV2': BiSeNetV2, 'SwinTransformer': SwinTransformer, 'VisionTransformer': VisionTransformer,
    'MixVisionTransformer': MixVisionTransformer, 'TIMMBackbone': TIMMBackbone, 'ConvNeXtV2': ConvNeXtV2,
    'MobileNetV2': MobileNetV2, 'MobileNetV3': MobileNetV3, 
}

The other arguments in SEGMENTOR_CFG['backbone'] are set for instancing the corresponding backbone network.

Here we also list some common arguments and their explanation,

• structure_type: The structure type of the specified backbone network, e.g., resnet101conv3x3stem means ResNet-101 using three 3x3 convolutions as the stem layer, it is useful if you want to load the pretrained backbone weights automatically,

• pretrained: Whether to load the pretrained backbone weights,

• pretrained_model_path: If you set pretrained_model_path as None and pretrained as True, SSSegmentation will load the pretrained backbone weights automatically, otherwise, load the pretrained backbone weights from the path specified by pretrained_model_path,

• out_indices: Generally, a backbone network can be divided into several stages, out_indices is used to specify whether return the feature maps outputted by the corresponding backbone stage,

• norm_cfg: The config of normalization layer, it should be a dict, refer to customize-normalizations more details,

• act_cfg: The config of activation layer, it should be a dict, refer to customize-activations more details.

To learn more about how to set the specific arguments for each backbone type, you can jump to ssseg/modules/models/backbones directory to check the source codes of each backbone network.

Add New Custom Backbone

If the users want to add a new custom backbone, you should first create a new file in ssseg/modules/models/backbones directory, e.g., ssseg/modules/models/backbones/mobilenet.py.

Then, you can define the backbone module in this file by yourselves, e.g.,

import torch.nn as nn

'''MobileNet'''
class MobileNet(nn.Module):
    def __init__(self, arg1, arg2):
        pass
    def forward(self, x):
        pass

After that, you should add this custom backbone class in ssseg/modules/models/backbones/builder.py if you want to use it by simply modifying SEGMENTOR_CFG['backbone']. Of course, you can also register this custom backbone by the following codes,

from ssseg.modules import BackboneBuilder

backbone_builder = BackboneBuilder()
backbone_builder.register('MobileNet', MobileNet)

From this, you can also call backbone_builder.build to build your own defined backbone class as well as the original supported backbone classes.

Finally, the users could jump to the ssseg/modules/models/backbones directory in SSSegmentation to read more source codes of the supported backbone classes and thus better learn how to customize the backbone classes in SSSegmentation.

Customize Losses

Loss is utilized to define the objective functions for the segmentation framework, such as the Cross Entropy Loss.

Loss Config Structure

An example of loss config is as follows,

SEGMENTOR_CFG['losses'] = {
    'loss_aux': {'type': 'CrossEntropyLoss', 'scale_factor': 0.4, 'ignore_index': 255, 'reduction': 'mean'},
    'loss_cls': {'type': 'CrossEntropyLoss', 'scale_factor': 1.0, 'ignore_index': 255, 'reduction': 'mean'},
}

It is a dict including several keys like loss_aux and loss_cls, which is designed for distinguishing the loss config utilized in head or auxiliary head. The value corresponding to each key is also a dict or a list of dict and in each dict, type denotes the objective function type you want to adopt and the other arguments are set for instancing this objective function.

In the example of loss config above, SEGMENTOR_CFG['losses']['loss_cls'] is used to build losses in head and SEGMENTOR_CFG['losses']['loss_aux'] is used to build losses in auxiliary head. CrossEntropyLoss represents the objective function type you want to adopt and {'scale_factor': 1.0, 'ignore_index': 255, 'reduction': 'mean'} contains the arguments for instancing the CrossEntropyLoss.

And if SEGMENTOR_CFG['losses']['loss_aux'] or SEGMENTOR_CFG['losses']['loss_cls'] is a list of dict, the final loss will be calculated as,

loss = 0
for l_cfg in SEGMENTOR_CFG['losses']['loss_aux']:
    loss = loss + BuildLoss(l_cfg)(prediction, target)

Now, SSSegmentation supports the following loss types,

REGISTERED_MODULES = {
    'L1Loss': L1Loss, 'DiceLoss': DiceLoss, 'KLDivLoss': KLDivLoss, 'LovaszLoss': LovaszLoss,
    'CrossEntropyLoss': CrossEntropyLoss, 'SigmoidFocalLoss': SigmoidFocalLoss,
    'CosineSimilarityLoss': CosineSimilarityLoss, 'BinaryCrossEntropyLoss': BinaryCrossEntropyLoss,
}

Here we also list some common arguments and their explanation,

• scale_factor: The return loss value is equal to the original loss times scale_factor,

• ignore_index: Specifies a target value that is ignored and does not contribute to the input gradient,

• lowest_loss_value: If lowest_loss_value is set, the return loss value is equal to min(lowest_loss_value, scale_factor * original loss), this argument is designed according to the paper Do We Need Zero Training Loss After Achieving Zero Training Error? - ICML 2020.

To learn more about how to set the specific arguments for each loss function, you can jump to ssseg/modules/models/losses directory to check the source codes of each loss function.

Add New Custom Loss

If the users want to add a new custom loss, you should first create a new file in ssseg/modules/models/losses directory, e.g., ssseg/modules/models/losses/klloss.py.

Then, you can define the loss function in this file by yourselves, e.g.,

import torch.nn as nn

'''KLDivLoss'''
class KLDivLoss(nn.Module):
    def __init__(self, arg1, arg2):
        pass
    def forward(self, prediction, target):
        pass

After that, you should add this custom loss class in ssseg/modules/models/losses/builder.py if you want to use it by simply modifying SEGMENTOR_CFG['losses']. Of course, you can also register this custom loss function by the following codes,

from ssseg.modules import LossBuilder

loss_builder = LossBuilder()
loss_builder.register('KLDivLoss', KLDivLoss)

From this, you can also call loss_builder.build to build your own defined loss class as well as the original supported loss classes.

Finally, the users could jump to the ssseg/modules/models/losses directory in SSSegmentation to read more source codes of the supported loss classes and thus better learn how to customize the loss classes in SSSegmentation.

Customize Schedulers

Scheduler provides several methods to adjust the learning rate based on the number of epochs or iterations.

Scheduler Config Structure

An example of scheduler config is as follows,

SEGMENTOR_CFG['scheduler'] = {
    'type': 'PolyScheduler', 'max_epochs': 0, 'power': 0.9,
    'optimizer': {
        'type': 'SGD', 'lr': 0.01, 'momentum': 0.9, 'weight_decay': 5e-4, 'params_rules': {},
    }
}

where type denotes the scheduler you want to utilize during training. Now, SSSegmentation supports the following scheduler types,

REGISTERED_MODULES = {
    'PolyScheduler': PolyScheduler, 'CosineScheduler': CosineScheduler,
}

The other arguments in SEGMENTOR_CFG['scheduler'] are set for instancing the corresponding scheduler, where SEGMENTOR_CFG['scheduler']['optimizer'] is the optimizer config used to build a optimizer for model training. The detailed instruction about building optimizer in SSSegmentation please refer to Customize Optimizers.

To learn more about how to set the specific arguments for each scheduler, you can jump to ssseg/modules/models/schedulers directory to check the source codes of each scheduler class.

Customize Optimizers

Optimizer is used to define the process of adjusting model parameters to reduce model error in each training step, such as the Stochastic Gradient Descent.

Specifically, an example of optimizer config used to construct an optimizer for model training could be,

SEGMENTOR_CFG['scheduler']['optimizer'] = {
    'type': 'SGD', 'lr': 0.01, 'momentum': 0.9, 'weight_decay': 5e-4, 'params_rules': {},
}

where type denotes the optimizer you want to utilize during training. Here is a list of supported optimizer types,

REGISTERED_MODULES = {
    'SGD': optim.SGD, 'Adam': optim.Adam, 'AdamW': optim.AdamW, 'Adadelta': optim.Adadelta,
}

The other arguments in SEGMENTOR_CFG['scheduler']['optimizer'] are set for instancing the corresponding optimizer.

Among these arguments, params_rules could be set for implementing some training tricks. For example, if you want to train the backbone network and the decoder layer with different learning rates, you can set params_rules as following,

SEGMENTOR_CFG['scheduler']['optimizer']['params_rules'] = {
    'backbone_net': dict(lr_multiplier=0.1, wd_multiplier=1.0),
}

where lr_multiplier = 0.1 means the learning rate of backbone network is one-tenth of the decoder layer.

And if you want to set the weight decay of some layers in the segmentor as zeros, you can set params_rules as following,

SEGMENTOR_CFG['scheduler']['optimizer']['params_rules'] = {
    'absolute_pos_embed': dict(wd_multiplier=0.),
    'relative_position_bias_table': dict(wd_multiplier=0.),
    'norm': dict(wd_multiplier=0.),
}

You can refer to the ssseg/configs directory for more examples about implementing some training tricks by setting params_rules.

Finally, if you want to customize the optimizer by yourselves during developing, you should first create a new file in ssseg/modules/models/optimizers directory, e.g., ssseg/modules/models/optimizers/sgd.py.

Then, you can define the optimization algorithm in this file by yourselves, e.g.,

class SGD():
    def __init__(self, arg1, arg2):
        pass

After that, you should add this custom optimization algorithm in ssseg/modules/models/optimizers/builder.py if you want to use it by simply modifying SEGMENTOR_CFG['scheduler']['optimizer']. Of course, you can also register this custom optimization algorithm by the following codes,

from ssseg.modules import OptimizerBuilder

optimizer_builder = NormalizationBuilder()
optimizer_builder.register('SGD', SGD)

From this, you can also call optimizer_builder.build to build your own defined optimization algorithms as well as the original supported optimization algorithms.

Add New Custom Scheduler

SSSegmentation provides BaseScheduler class to help the users quickly add a new custom scheduler.

Specifically, if the users want to add a new custom scheduler, you should first create a new file in ssseg/modules/models/schedulers directory, e.g., ssseg/modules/models/schedulers/polyscheduler.py.

Then, you can define the scheduler in this file by inheriting BaseScheduler, e.g.,

from .basescheduler import BaseScheduler

'''PolyScheduler'''
class PolyScheduler(BaseScheduler):
    def __init__(self, arg1, arg2):
        pass
    def updatelr(self):
        pass

where updatelr function is used to define the learning rate adjust strategy.

After that, you should add this custom scheduler class in ssseg/modules/models/schedulers/builder.py if you want to use it by simply modifying SEGMENTOR_CFG['scheduler']. Of course, you can also register this custom scheduler by the following codes,

from ssseg.modules import SchedulerBuilder

scheduler_builder = SchedulerBuilder()
scheduler_builder.register('PolyScheduler', PolyScheduler)

From this, you can also call scheduler_builder.build to build your own defined schedulers as well as the original supported schedulers.

Finally, the users could jump to the ssseg/modules/models/schedulers directory in SSSegmentation to read more source codes of the supported schedulers and thus better learn how to customize the schedulers in SSSegmentation.

Customize Segmentors

Segmentor will process the feature maps outputted by the backbone network, and transforms the feature maps to a predicted segmentation mask using a decoder head (e.g., Deeplabv3 and IDRNet).

Here is an example of head config,

SEGMENTOR_CFG['head'] = {
    'in_channels_list': [1024, 2048], 'transform_channels': 256, 'query_scales': (1, ), 
    'feats_channels': 512, 'key_pool_scales': (1, 3, 6, 8), 'dropout': 0.1,
}

These arguments will be used during instancing the segmentor where the segmentor type is specified in SEGMENTOR_CFG['type']. Now, SSSegmentation supports the following segmentor types,

REGISTERED_MODULES = {
    'FCN': FCN, 'CE2P': CE2P, 'ICNet': ICNet, 'ISNet': ISNet, 'CCNet': CCNet, 'DANet': DANet,
    'GCNet': GCNet, 'DMNet': DMNet, 'ISANet': ISANet, 'ENCNet': ENCNet, 'APCNet': APCNet, 'SAM': SAM,
    'EMANet': EMANet, 'PSPNet': PSPNet, 'PSANet': PSANet, 'OCRNet': OCRNet, 'DNLNet': DNLNet,
    'ANNNet': ANNNet, 'SETRUP': SETRUP, 'SETRMLA': SETRMLA, 'FastFCN': FastFCN, 'UPerNet': UPerNet,
    'Segformer': Segformer, 'MCIBI': MCIBI, 'PointRend': PointRend, 'Deeplabv3': Deeplabv3,
    'LRASPPNet': LRASPPNet, 'MaskFormer': MaskFormer, 'MCIBIPlusPlus': MCIBIPlusPlus, 'SemanticFPN': SemanticFPN,
    'NonLocalNet': NonLocalNet, 'Deeplabv3Plus': Deeplabv3Plus, 'DepthwiseSeparableFCN': DepthwiseSeparableFCN,
    'MobileSAM': MobileSAM, 'IDRNet': IDRNet, 'Mask2Former': Mask2Former,
}

To learn more about how to set the specific arguments for each segmentor, you can jump to ssseg/modules/models/segmentors directory to check the source codes of each segmentor.

Also, SSSegmentation provides BaseSegmentor class to help the users quickly add a new custom segmentor.

Specifically, if the users want to add a new custom segmentor, you should first create a new file in ssseg/modules/models/segmentors directory, e.g., ssseg/modules/models/segmentors/fcn/fcn.py.

Then, you can define the segmentor in this file by inheriting BaseSegmentor, e.g.,

from ..base import BaseSegmentor

'''FCN'''
class FCN(BaseSegmentor):
    def __init__(self, cfg, mode):
        super(FCN, self).__init__(cfg, mode)
	'''forward'''
    def forward(self, x, targets=None):
        pass

After that, you should add this custom segmentor class in ssseg/modules/models/segmentors/builder.py if you want to use it by simply modifying SEGMENTOR_CFG. Of course, you can also register this custom segmentor by the following codes,

from ssseg.modules import SegmentorBuilder

segmentor_builder = SegmentorBuilder()
segmentor_builder.register('FCN', FCN)

From this, you can also call segmentor_builder.build to build your own defined segmentors as well as the original supported segmentors.

Finally, the users could jump to the ssseg/modules/models/segmentors directory in SSSegmentation to read more source codes of the supported segmentors and thus better learn how to customize the segmentors in SSSegmentation.

Customize Auxiliary Heads

Auxiliary head is the image decoder that transforms the shallow feature maps to a predicted segmentation mask. It is first introduced in PSPNet, which is used to help segmentation framework training.

Auxiliary Head Config Structure

An example of auxiliary head config is as follows,

SEGMENTOR_CFG['auxiliary'] = {'in_channels': 1024, 'out_channels': 512, 'dropout': 0.1}

With the arguments in SEGMENTOR_CFG['auxiliary'], the setauxiliarydecoder function defined in ssseg/modules/models/segmentors/base/base.py will be called to build the auxiliary head.

To disable auxiliary head, you can simply set SEGMENTOR_CFG['auxiliary'] as None.

Add New Custom Auxiliary Head

If the users want to add a new custom auxiliary head, you can first define a new segmentor inherited from BaseSegmentor class like following,

from ..base import BaseSegmentor

'''Deeplabv3'''
class Deeplabv3(BaseSegmentor):
    def __init__(self, arg1, arg2):
        pass
    def forward(self, x, targets=None):
        pass

After that, you can define the setauxiliarydecoder function in this class by yourselves, e.g.,

from ..base import BaseSegmentor

'''Deeplabv3'''
class Deeplabv3(BaseSegmentor):
    def __init__(self, arg1, arg2):
        pass
    def forward(self, x, targets=None):
        pass
    def setauxiliarydecoder(self, auxiliary_cfg):
        pass

And the arguments in SEGMENTOR_CFG['auxiliary'] could be changed according to the new defined setauxiliarydecoder function.

Customize Normalizations

Normalization layer transforms the inputs to have zero mean and unit variance across some specific dimensions.

Normalization Config Structure

An example of normalization config is as follows,

SEGMENTOR_CFG['norm_cfg'] = {'type': 'SyncBatchNorm'}

where type denotes the normalization layer you want to leverage. Now, SSSegmentation supports the following normalization types,

REGISTERED_MODULES = {
    'LayerNorm': nn.LayerNorm, 'LayerNorm2d': LayerNorm2d, 'GroupNorm': nn.GroupNorm, 'LocalResponseNorm': nn.LocalResponseNorm, 
    'BatchNorm1d': nn.BatchNorm1d, 'BatchNorm2d': nn.BatchNorm2d, 'BatchNorm3d': nn.BatchNorm3d, 'SyncBatchNorm': nn.SyncBatchNorm, 
    'InstanceNorm1d': nn.InstanceNorm1d, 'InstanceNorm2d': nn.InstanceNorm2d, 'InstanceNorm3d': nn.InstanceNorm3d, 'GRN': GRN,
}

The other arguments in SEGMENTOR_CFG['norm_cfg'] are set for instancing the corresponding normalization layer.

To learn more about how to set the specific arguments for each normalization layer, you can jump to ssseg/modules/models/backbones/bricks/normalization directory to check the source codes of each normalization layer.

Add New Custom Normalization

If the users want to add a new custom normalization layer, you should first create a new file in ssseg/modules/models/backbones/bricks/normalization directory, e.g., ssseg/modules/models/backbones/bricks/normalization/grn.py.

Then, you can define the normalization layer in this file by yourselves, e.g.,

import torch.nn as nn

'''GRN'''
class GRN(nn.Module):
    def __init__(self, arg1, arg2):
        pass
    def forward(self, x):
        pass

After that, you should add this custom normalization class in ssseg/modules/models/backbones/bricks/normalization/builder.py if you want to use it by simply modifying SEGMENTOR_CFG['norm_cfg'] or SEGMENTOR_CFG['backbone']['norm_cfg']. Of course, you can also register this custom normalization layer by the following codes,

from ssseg.modules import NormalizationBuilder

norm_builder = NormalizationBuilder()
norm_builder.register('GRN', GRN)

From this, you can also call norm_builder.build to build your own defined normalization layers as well as the original supported normalization layers.

Finally, the users could jump to the ssseg/modules/models/backbones/bricks/normalization directory in SSSegmentation to read more source codes of the supported normalization classes and thus better learn how to customize the normalization layers in SSSegmentation.

Customize Activations

Activation layer is linear or non linear equation which processes the output of a neuron and bound it into a limited range of values (e.g., [0, +∞]).

Activation Config Structure

An example of activation config is as follows,

SEGMENTOR_CFG['act_cfg'] = {'type': 'ReLU', 'inplace': True}

where type denotes the activation layer you want to employ. Now, SSSegmentation supports the following activation types,

REGISTERED_MODULES = {
    'ReLU': nn.ReLU, 'GELU': nn.GELU, 'ReLU6': nn.ReLU6, 'PReLU': nn.PReLU,
    'Sigmoid': nn.Sigmoid, 'HardSwish': HardSwish, 'LeakyReLU': nn.LeakyReLU,
    'HardSigmoid': HardSigmoid, 'Swish': Swish,
}

The other arguments in SEGMENTOR_CFG['act_cfg'] are set for instancing the corresponding activation layer.

To learn more about how to set the specific arguments for each activation function, you can jump to ssseg/modules/models/backbones/bricks/activation directory to check the source codes of each activation function.

Add New Custom Activation

If the users want to add a new custom activation layer, you should first create a new file in ssseg/modules/models/backbones/bricks/activation directory, e.g., ssseg/modules/models/backbones/bricks/activation/hardsigmoid.py.

Then, you can define the activation layer in this file by yourselves, e.g.,

import torch.nn as nn

'''HardSigmoid'''
class HardSigmoid(nn.Module):
    def __init__(self, arg1, arg2):
        pass
    def forward(self, x):
        pass

After that, you should add this custom activation class in ssseg/modules/models/backbones/bricks/activation/builder.py if you want to use it by simply modifying SEGMENTOR_CFG['act_cfg'] or SEGMENTOR_CFG['backbone']['act_cfg']. Of course, you can also register this custom activation layer by the following codes,

from ssseg.modules import ActivationBuilder

act_builder = ActivationBuilder()
act_builder.register('HardSigmoid', HardSigmoid)

From this, you can also call act_builder.build to build your own defined activation layers as well as the original supported activation layers.

Finally, the users could jump to the ssseg/modules/models/backbones/bricks/activation directory in SSSegmentation to read more source codes of the supported activation classes and thus better learn how to customize the activation layers in SSSegmentation.

Mixed Precision Training

Mixed precision methods combine the use of different numerical formats in one computational workload. It offers significant computational speedup by performing operations in half-precision format, while storing minimal information in single-precision to retain as much information as possible in critical parts of the network.

For more technical details, please refer to Mixed Precision Training.

SSSegmentation supports two types of mixed precision training, i.e.,

• apex: Mixed precision training implemented by using the third-party python package apex supported by NVIDIA,

• pytorch: Mixed precision training implemented by using torch.cuda.amp supported by Pytorch official API.

To turn on the mixed precision training in SSSegmentation, you could modify the corresponding config file with the following codes,

import torch

# use Mixed Precision (FP16) Training supported by Apex
SEGMENTOR_CFG['fp16_cfg'] = {'type': 'apex', 'initialize': {'opt_level': 'O1'}, 'scale_loss': {}}
# use Mixed Precision (FP16) Training supported by Pytorch
SEGMENTOR_CFG['fp16_cfg'] = {'type': 'pytorch', 'autocast': {'dtype': torch.float16}, 'grad_scaler': {}}

If you choose to use the mixed precision training supported by Apex, the following arguments could be given,

• initialize: Arguments dict for instancing apex.amp.initialize,

• scale_loss: Arguments dict for calling apex.amp.scale_loss.

The detailed usage and the explanations of each argument please refer to Apex Official Document.

Of course, you can also choose to adopt the mixed precision training supported by Pytorch and the following arguments could be given,

• autocast: Arguments dict for instancing torch.cuda.amp.autocast,

• grad_scaler: Arguments dict for instancing torch.cuda.amp.GradScaler.

The detailed usage and the explanations of each argument please refer to Pytorch Official Document.

Finally, if you want turn off the mixed precision training in SSSegmentation, just delete fp16_cfg in SEGMENTOR_CFG or set SEGMENTOR_CFG['fp16_cfg']['type'] as None.

Exponential Moving Average (EMA)

Exponential moving average is a neural network training trick that sometimes improves the model accuracy. Concretely, instead of using the optimized parameters from the final training iteration (parameter update step) as the final parameters for the model, the exponential moving average of the parameters over the course of all the training iterations are used.

When training a model, it is often beneficial to maintain moving averages of the trained parameters. Evaluations that use averaged parameters sometimes produce significantly better results than the final trained values.

You can refer to Exponential-Moving-Average for more technical details.

To turn on EMA in SSSegmentation, you could modify the corresponding config file with the following codes,

SEGMENTOR_CFG['ema_cfg'] = {'momentum': 0.0005, 'device': 'cpu'}

where device denotes perform EMA on CPU or GPU, momentum is the moving average weight which is used in the following codes,

ema_v * (1.0 - momentum)) + (momentum * cur_v)

Finally, if you want turn off EMA in SSSegmentation, just delete ema_cfg in SEGMENTOR_CFG or set SEGMENTOR_CFG['ema_cfg']['momentum'] as None.