namKolorfuL/torchseg

Segmentation models with pretrained backbones. PyTorch.

TorchSeg

TorchSeg is an actively maintained and up-to-date fork of the Segmentation Models PyTorch (smp) library.

Updates

The goal of this fork is to 1) provide maintenance support for the original library and 2) add features relevant to modern semantic segmentation. Since the fork, this library has added some features which can be summarized below:

• Improved PyTorch Image Models (timm) for models with feature extraction functionality (852/1017=84% of timm models). This includes the typical CNN models such as ResNet, EfficientNet, etc., but now extends to include modern architectures like ConvNext, Swin, PoolFormer, MaxViT and more!
• Support for pretrained Vision Transformer (ViT) encoders. Currently timm ViTs do not support feature extraction out of the box. However we have added support for extracting intermediate transformer encoder layer feature maps to obtain this functionality. We support 100+ ViT based models including ViT, DeiT, FlexiViT!

Additionally we have performed the following for improved software standards:

• More thorough testing and CI
• Formatting using black, isort, flake8, mypy
• Reduction of dependence on unmaintained libraries (now depends only on torch, timm, and einops)
• Reduce lines of code to maintain (removed custom utils, metrics, encoders) in favor of newer libraries such as torchmetrics and timm

Features

The main features of this library are:

• High level API (just two lines to create a neural network)
• 9 segmentation architectures for binary and multi class segmentation (including U-Net, DeepLabV3)
• Support for 852/1017 (~84%) of available encoders from timm
• All encoders have pre-trained weights for faster and better convergence
• Popular segmentation loss functions

Example Usage

TorchSeg models at their base are just torch nn.Modules. They can be created as follows:

import torchseg

model = torchseg.Unet(
    encoder_name="resnet50",
    encoder_weights=True,
    in_channels=3
    classes=3,
)

TorchSeg has an encoder_params feature which passes additional parameters to timm.create_model() when defining an encoder backbone. One can specify different activitions, normalization layers, and more like below.

You can also define a functools.partial callable as an activation/normalization layer. See the timm docs for more information on available activations and normalization layers. You can even used pretrained weights while changing the activations/normalizations!

model = torchseg.Unet(
    encoder_name="resnet50",
    encoder_weights=True,
    in_channels=3
    classes=3,
    encoder_params={
      "act_layer": "prelu",
      "norm_layer": "layernorm"
    }
)

Some models like Swin and ConvNext perform a downsampling of scale=4 in the first block (stem) and then downsample by 2 afterwards with only depth=4 blocks. This results in an output size of half after the decoder. To get the same output size as the input you can pass head_upsampling=2 which will upsample once more prior to the segmentation head.

model = torchseg.Unet(
    "convnextv2_tiny",
    in_channels=3,
    classes=2,
    encoder_weights=True,
    encoder_depth=4,
    decoder_channels=(256, 128, 64, 32),
    head_upsampling=2
)

model = torchseg.Unet(
    "swin_tiny_patch4_window7_224",
    in_channels=3,
    classes=2,
    encoder_weights=True,
    encoder_depth=4,
    decoder_channels=(256, 128, 64, 32),
    head_upsampling=2,
    encoder_params={"img_size": 256}  # need to define img size since swin is a ViT hybrid
)

model = torchseg.Unet(
    "maxvit_small_tf_224",
    in_channels=3,
    classes=2,
    encoder_weights=True,
    encoder_depth=5,
    decoder_channels=(256, 128, 64, 32, 16),
    encoder_params={"img_size": 256}
)

TorchSeg supports pretrained ViT encoders from timm by extracting intermediate transformer block features specified by the encoder_indices and encoder_depth arguments.

You will also need to define scale_factors for upsampling the feature layers to the resolutions expected by the decoders. For U-Net depth=5 this would be scales=(8, 4, 2, 1, 0.5). For depth=4 this would be scales=(4, 2, 1, 0.5), for depth=3 this would be scales=(2, 1, 0.5) and so on.

Another benefit of using timm is that by passing in a new img_size, timm automatically interpolates the ViT positional embeddings to work with your new image size which creates a different number of patch tokens.

import torch
import torchseg


model = torchseg.Unet(
    "vit_small_patch16_224",
    in_channels=8,
    classes=2,
    encoder_depth=5,
    encoder_indices=(2, 4, 6, 8, 10),  # which intermediate blocks to extract features from
    encoder_weights=True,
    decoder_channels=(256, 128, 64, 32, 16),
    encoder_params={  # additional params passed to timm.create_model and the vit encoder
        "scale_factors": (8, 4, 2, 1, 0.5), # resize scale_factors for patch size 16 and 5 layers
        "img_size": 256,  # timm automatically interpolates the positional embeddings to your new image size
    },
)

Models

Architectures (Decoders)

• Unet [paper]
• Unet++ [paper]
• MAnet [paper]
• Linknet [paper]
• FPN [paper]
• PSPNet [paper]
• PAN [paper]
• DeepLabV3 [paper]
• DeepLabV3+ [paper]

Encoders

TorchSeg relies entirely on the timm library for pretrained encoder support. This means that TorchSeg supports any timm model which has features_only feature extraction functionality. Additionally we support any ViT models with a get_intermediate_layers method. This results in a total of 852/1017 (~84%) encoders from timm including ResNet, Swin, ConvNext, ViT, and more!

To list the following supported encoders:

import torchseg

torchseg.list_encoders()

We have additionally pulled the the feature extractor metadata of each model with features_only support from timm at output_stride=32. This metadata provides information such as the number of intermediate layers, channels for each layer, layer name, and downsampling reduction.

import torchseg

metadata = torchseg.encoders.TIMM_ENCODERS["convnext_base"]
print(metadata)

"""
{
   'channels': [128, 256, 512, 1024],
   'indices': (0, 1, 2, 3),
   'module': ['stages.0', 'stages.1', 'stages.2', 'stages.3'],
   'reduction': [4, 8, 16, 32],
}
"""

metadata = torchseg.encoders.TIMM_ENCODERS["resnet50"]
print(metadata)

"""
{
   'channels': [64, 256, 512, 1024, 2048],
   'indices': (0, 1, 2, 3, 4),
   'module': ['act1', 'layer1', 'layer2', 'layer3', 'layer4'],
   'reduction': [2, 4, 8, 16, 32]
}
"""

Models API

• model.encoder - pretrained backbone to extract intermediate features
• model.decoder - network for processing the intermediate features to the original image resolution (Unet, DeepLabv3+, FPN)
• model.segmentation_head - final block producing the mask output (includes optional upsampling and activation)
• model.classification_head - optional block which create classification head on top of encoder
• model.forward(x) - sequentially pass x through model`s encoder, decoder and segmentation head (and classification head if specified)

Input channels

Timm encoders supports the use of pretrained weights with arbitrary input channels by repeating weights for channels if > 3. For example, if in_channels=6, RGB ImageNet pretrained weights in the initial layer would be repeated RGBRGB to avoid random initialization. For in_channels=7 this would result in RGBRGBR. Below is a diagram to visualize this method.

Auxiliary Classifier

All models support an optional auxiliary classifier head through the use of aux_params. If aux_params != None then the model will produce the a label output in addition to the mask output with shape (N, C). Classification head consists of GlobalPooling->Dropout(optional)->Linear->Activation(optional) layers, which can be configured by aux_params as follows:

aux_params=dict(
    pooling='avg',             # one of 'avg', 'max'
    dropout=0.5,               # dropout ratio, default is None
    activation=nn.Sigmoid(),   # activation function, default is Identity
    classes=4,                 # define number of output labels
)
model = torchseg.Unet('resnet18', classes=4, aux_params=aux_params)
mask, label = model(x)

Depth

Depth represents the number of downsampling operations in the encoder, so you can make your model lighter by specifying less depth. Defaults to depth=5.

Note that some models like ConvNext and Swin only have 4 intermediate feature blocks. Therefore, to use these encoders set encoder_depth=4. This can be found in the metadata above.

model = torchseg.Unet('resnet50', encoder_depth=4)