Skip to content
Segmentation models with pretrained backbones. Keras and TensorFlow Keras.
Python
Branch: master
Clone or download
Latest commit a6fbf2d Jan 10, 2020
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
.github Update FUNDING.yml Dec 26, 2019
docs Fix wrong parameter name in docs (#235) Oct 30, 2019
examples Release v1.0.0b1 (tf.keras) (#162) Aug 9, 2019
images fix-images Feb 17, 2019
segmentation_models Release v1.0.1 (#283) Jan 10, 2020
tests Release v1.0.0b1 (tf.keras) (#162) Aug 9, 2019
.gitignore Release v1.0.0b1 (tf.keras) (#162) Aug 9, 2019
.gitmodules submodule removed Dec 21, 2018
.travis.yml Release v1.0.0 (#228) Oct 15, 2019
CHANGELOG.md Release v1.0.0 (#228) Oct 15, 2019
LICENSE
MANIFEST.in Release v1.0.0 (#228) Oct 15, 2019
README.rst Update README.rst (#275) Dec 15, 2019
__init__.py Update __init__.py Sep 5, 2018
requirements.txt Release v1.0.0 (#228) Oct 15, 2019
setup.py Release v1.0.0b1 (tf.keras) (#162) Aug 9, 2019

README.rst

Python library with Neural Networks for Image Segmentation based on Keras and TensorFlow.

The main features of this library are:

  • High level API (just two lines of code to create model for segmentation)
  • 4 models architectures for binary and multi-class image segmentation (including legendary Unet)
  • 25 available backbones for each architecture
  • All backbones have pre-trained weights for faster and better convergence
  • Helpful segmentation losses (Jaccard, Dice, Focal) and metrics (IoU, F-score)

Important note

Some models of version 1.* are not compatible with previously trained models, if you have such models and want to load them - roll back with:

$ pip install -U segmentation-models==0.2.1

Table of Contents

Quick start

Library is build to work together with Keras and TensorFlow Keras frameworks

import segmentation_models as sm
# Segmentation Models: using `keras` framework.

By default it tries to import keras, if it is not installed, it will try to start with tensorflow.keras framework. There are several ways to choose framework:

  • Provide environment variable SM_FRAMEWORK=keras / SM_FRAMEWORK=tf.keras before import segmentation_models
  • Change framework sm.set_framework('keras') / sm.set_framework('tf.keras')

You can also specify what kind of image_data_format to use, segmentation-models works with both: channels_last and channels_first. This can be useful for further model conversion to Nvidia TensorRT format or optimizing model for cpu/gpu computations.

import keras
# or from tensorflow import keras

keras.backend.set_image_data_format('channels_last')
# or keras.backend.set_image_data_format('channels_first')

Created segmentation model is just an instance of Keras Model, which can be build as easy as:

model = sm.Unet()

Depending on the task, you can change the network architecture by choosing backbones with fewer or more parameters and use pretrainded weights to initialize it:

model = sm.Unet('resnet34', encoder_weights='imagenet')

Change number of output classes in the model (choose your case):

# binary segmentation (this parameters are default when you call Unet('resnet34')
model = sm.Unet('resnet34', classes=1, activation='sigmoid')
# multiclass segmentation with non overlapping class masks (your classes + background)
model = sm.Unet('resnet34', classes=3, activation='softmax')
# multiclass segmentation with independent overlapping/non-overlapping class masks
model = sm.Unet('resnet34', classes=3, activation='sigmoid')

Change input shape of the model:

# if you set input channels not equal to 3, you have to set encoder_weights=None
# how to handle such case with encoder_weights='imagenet' described in docs
model = Unet('resnet34', input_shape=(None, None, 6), encoder_weights=None)

Simple training pipeline

import segmentation_models as sm

BACKBONE = 'resnet34'
preprocess_input = sm.get_preprocessing(BACKBONE)

# load your data
x_train, y_train, x_val, y_val = load_data(...)

# preprocess input
x_train = preprocess_input(x_train)
x_val = preprocess_input(x_val)

# define model
model = sm.Unet(BACKBONE, encoder_weights='imagenet')
model.compile(
    'Adam',
    loss=sm.losses.bce_jaccard_loss,
    metrics=[sm.metrics.iou_score],
)

# fit model
# if you use data generator use model.fit_generator(...) instead of model.fit(...)
# more about `fit_generator` here: https://keras.io/models/sequential/#fit_generator
model.fit(
   x=x_train,
   y=y_train,
   batch_size=16,
   epochs=100,
   validation_data=(x_val, y_val),
)

Same manipulations can be done with Linknet, PSPNet and FPN. For more detailed information about models API and use cases Read the Docs.

Examples

Models training examples:
  • [Jupyter Notebook] Binary segmentation (cars) on CamVid dataset here.
  • [Jupyter Notebook] Multi-class segmentation (cars, pedestrians) on CamVid dataset here.

Models and Backbones

Models

Unet Linknet
unet_image linknet_image
PSPNet FPN
psp_image fpn_image

Backbones

Type Names
VGG 'vgg16' 'vgg19'
ResNet 'resnet18' 'resnet34' 'resnet50' 'resnet101' 'resnet152'
SE-ResNet 'seresnet18' 'seresnet34' 'seresnet50' 'seresnet101' 'seresnet152'
ResNeXt 'resnext50' 'resnext101'
SE-ResNeXt 'seresnext50' 'seresnext101'
SENet154 'senet154'
DenseNet 'densenet121' 'densenet169' 'densenet201'
Inception 'inceptionv3' 'inceptionresnetv2'
MobileNet 'mobilenet' 'mobilenetv2'
EfficientNet 'efficientnetb0' 'efficientnetb1' 'efficientnetb2' 'efficientnetb3' 'efficientnetb4' 'efficientnetb5' efficientnetb6' efficientnetb7'
All backbones have weights trained on 2012 ILSVRC ImageNet dataset (encoder_weights='imagenet').

Installation

Requirements

  1. python 3
  2. keras >= 2.2.0 or tensorflow >= 1.13
  3. keras-applications >= 1.0.7, <=1.0.8
  4. image-classifiers == 1.0.*
  5. efficientnet == 1.0.*

PyPI stable package

$ pip install -U segmentation-models

PyPI latest package

$ pip install -U --pre segmentation-models

Source latest version

$ pip install git+https://github.com/qubvel/segmentation_models

Documentation

Latest documentation is avaliable on Read the Docs

Change Log

To see important changes between versions look at CHANGELOG.md

Citing

@misc{Yakubovskiy:2019,
  Author = {Pavel Yakubovskiy},
  Title = {Segmentation Models},
  Year = {2019},
  Publisher = {GitHub},
  Journal = {GitHub repository},
  Howpublished = {\url{https://github.com/qubvel/segmentation_models}}
}

License

Project is distributed under MIT Licence.

You can’t perform that action at this time.