TF Semantic Segmentation

Quick Start

See GETTING_STARTED, or the Colab Notebook.

Learn more at our documentation.
See upcoming features on our roadmap.

Features

• Distributed Training on Multiple GPUs
• Hyper Parameter Optimization using WandB
• WandB Integration
• Easily create TFRecord from Directory
• Tensorboard visualizations
• Ensemble inference

Datasets

• Ade20k
• Camvid
• Cityscapes
• MappingChallenge
• MotsChallenge
• Coco
• PascalVoc2012
• Taco
• Shapes (randomly creating triangles, rectangles and circles)
• Toy (Overlaying TinyImageNet with MNIST)
• ISIC2018
• CVC-ClinicDB

Models

• Unet
• Erfnet
• MultiResUnet
• PSP (experimental)
• FCN (experimental)
• NestedUnet (Unet++) (experimental)
• U2Net / U2NetP (experimental)
• SatelliteUnet
• MobilenetUnet (unet with mobilenet encoder pre-trained on imagenet)
• InceptionResnetV2Unet (unet with inception-resnet v2 encoder pre-trained on imagenet)
• ResnetUnet (unet with resnet50 encoder pre-trained on imagenet)
• AttentionUnet

Losses

• Catagorical Crossentropy
• Binary Crossentropy
• Crossentropy + SSIM
• Dice
• Crossentropy + Dice
• Tversky
• Focal
• Focal + Tversky

Activations

• mish
• swish
• relu6

Optimizers

• Ranger
• RAdam

Normalization

• Instance Norm
• Batch Norm

On the fly Augmentations

• flip left/right
• flip up/down
• rot 180
• color

Getting Started

Requirements

sudo apt-get install libsm6 libxext6 libxrender-dev libyaml-dev libpython3-dev

Tensorflow (2.x) & Tensorflow Addons (optional)

pip install tensorflow-gpu==2.4.0 --upgrade
pip install tensorflow-addons==0.12.0 --upgrade

Installation

pip install tf-semantic-segmentation

Run tensorboard

• Hint: To see train/test/val images you have to start tensorboard like this

tensorboard --logdir=logs/ --reload_multifile=true

Train on inbuild datasets (generator)

python -m tf_semantic_segmentation.bin.train -ds 'tacobinary' -bs 8 -e 100 \
    -logdir 'logs/taco-binary-test' -o 'adam' -lr 5e-3 --size 256,256 \
    -l 'binary_crossentropy' -fa 'sigmoid' \
    --train_on_generator --gpus='0' \
    --tensorboard_train_images --tensorboard_val_images

Create a tfrecord from a dataset

# create a tfrecord from the toy dataset and resize to 128x128
tf-semantic-segmentation-tfrecord-writer -d 'toy' -c /hdd/datasets/ -s '128,128'

Train using a fixed record path

python -m tf_semantic_segmentation.bin.train --record_dir=records/cityscapes-512x256-rgb/ \
    -bs 4 -e 100 -logdir 'logs/cityscapes-bs8-e100-512x256' -o 'adam' -lr 1e-4 -l 'categorical_crossentropy' \
    -fa 'softmax' -bufsize 50 --metrics='iou_score,f1_score' -m 'erfnet' --gpus='0' -a 'mish' \
    --tensorboard_train_images --tensorboard_val_images

Multi GPU training

python -m tf_semantic_segmentation.bin.train --record_dir=records/cityscapes-512x256-rgb/ \
    -bs 4 -e 100 -logdir 'logs/cityscapes-bs8-e100-512x256' -o 'adam' -lr 1e-4 -l 'categorical_crossentropy' \
    -fa 'softmax' -bufsize 50 --metrics='iou_score,f1_score' -m 'erfnet' --gpus='0,1,2,3' -a 'mish'

Using Code

from tf_semantic_segmentation.bin.train import train_test_model, get_args
# get the default args
args = get_args({})
# change some parameters
# !rm -r logs/
args.model = 'erfnet'
# args['color_mode'] = 0
args.batch_size = 8
args.size = [128, 128] # resize input dataset to this size
args.epochs = 10
args.learning_rate = 1e-4
args.optimizer = 'adam' # ['adam', 'radam', 'ranger']
args.loss = 'dice'
args.logdir = 'logs'
args.record_dir = "datasets/shapes/records"
args.final_activation = 'softmax'
# train and test
results, model = train_test_model(args)

Models

• Erfnet
• Unet

from tf_semantic_segmentation import models
# print all available models
print(list(modes.models_by_name.keys()))
# returns a model (without the final activation function)
model = models.get_model_by_name('erfnet', {"input_shape": (128, 128, 3), "num_classes": 5})
# call models directly
model = models.erfnet(input_shape=(128, 128), num_classes=5)

Use your own dataset

• Accepted file types are: jpg(jpeg) and png

If you already have a train/test/val split then use the following data structure:

dataset/
    labels.txt
    test/
        images/
        masks/
    train/
        images/
        masks/
    val/
        images/
        masks/

or use

dataset/
    labels.txt
    images/
    masks/

The labels.txt should contain a list of labels separated by newline [/n]. For instance it looks like this:

background
car
pedestrian

• To create a tfrecord using the original image size and color use the script like this:

INPUT_DIR = ...
tf-semantic-segmentation-tfrecord-writer -dir $INPUT_DIR -r $INPUT_DIR/records

There are the following addition arguments:

• -s [—size] ‘$width,$height’ (f.e. “512,512”)
• -rm [—resize_method] (‘resize’, ‘resize_with_pad’, ‘resize_with_crop_or_pad)
• cm [—color_mode] (0=RGB, 1=GRAY, 2=NONE (default))

Datasets

from tf_semantic_sementation.datasets import get_dataset by name, datasets_by_name, DataType, get_cache_dir
# print availiable dataset names
print(list(datasets_by_name.keys()))
# get the binary (waste or not) dataset
data_dir = '/hdd/data/'
name = 'tacobinary'
cache_dir = get_cache_dir(data_dir, name.lower())
ds = get_dataset_by_name(name, cache_dir)
# print labels and classes
print(ds.labels)
print(ds.num_classes)
# print number of training examples
print(ds.num_examples(DataType.TRAIN))
# or simply print the summary
ds.summary()

Debug datasets

python -m tf_semantic_segmentation.debug.dataset_vis -d ade20k

TFRecords

This library simplicifies the process of creating a tfrecord dataset for faster training.

Write tfrecords:

from tf_semantic_segmentation.datasets import TFWriter
ds = ...
writer = TFWriter(record_dir)
writer.write(ds)
writer.validate(ds)

or use simple with this script (will be save with size 128 x 128 (width x height)):

tf-semantic-segmentation-tfrecord-writer -d 'toy' -c /hdd/datasets/ -s '128,128'

Analyse already written tfrecord (with mean)

python -m tf_semantic_segmentation.bin.tfrecord_analyser -r records/ --mean

Docker

docker build -t tf_semantic_segmentation -f docker/Dockerfile ./

or pull the latest release

docker pull baudcode/tf_semantic_segmentation:latest

Prediction

pip install matplotlib

Using Code

from tensorflow.keras.models import load_model
import numpy as np
from tf_semantic_segmentation.processing import dataset
from tf_semantic_segmentation.visualizations import show, masks
model = load_model('logs/model-best.h5', compile=False)
# model parameters
size = tuple(model.input.shape[1:3])
depth = model.input.shape[-1]
color_mode = dataset.ColorMode.GRAY if depth == 1 else dataset.ColorMode.RGB
# define an image
image = np.zeros((256, 256, 3), np.uint8)
# preprocessing
image = image.astype(np.float32) / 255.
image, _ = dataset.resize_and_change_color(image, None, size, color_mode, resize_method='resize')
image_batch = np.expand_dims(image, axis=0)
# predict (returns probabilities)
p = model.predict(image_batch)
# draw segmentation map
num_classes = p.shape[-1] if p.shape[-1] > 1 else 2
predictions_rgb = masks.get_colored_segmentation_mask(p, num_classes, images=image_batch, binary_threshold=0.5)
# show images using matplotlib
show.show_images([predictions_rgb[0], image_batch[0]])

Using scripts

• On image

python -m tf_semantic_segmentation.evaluation.predict -m model-best.h5  -i image.png

• On TFRecord (data type ‘val’ is default)

python -m tf_semantic_segmentation.evaluation.predict -m model-best.h5 -r records/camvid/

• On TFRecord (with export to directory)

python -m tf_semantic_segmentation.evaluation.predict -m model-best.h5 -r records/cubbinary/ -o out/ -rm 'resize_with_pad'

• On Video

python -m tf_semantic_segmentation.evaluation.predict -m model-best.h5 -v video.mp4

• On Video (with export to out/p-video.mp4)

python -m tf_semantic_segmentation.evaluation.predict -m model-best.h5 -v video.mp4 -o out/

Prediction using Tensorflow Model Server

• Installation

# install
echo "deb [arch=amd64] http://storage.googleapis.com/tensorflow-serving-apt stable tensorflow-model-server tensorflow-model-server-universal" | sudo tee /etc/apt/sources.list.d/tensorflow-serving.list && \
curl https://storage.googleapis.com/tensorflow-serving-apt/tensorflow-serving.release.pub.gpg | sudo apt-key add -
sudo apt-get update && apt-get install tensorflow-model-server

• Start Model Server

### using a single model
tensorflow_model_server --rest_api_port=8501 --model_base_path=/home/user/models/mymodel/saved_model
### or using an ensamble of multiple models
# helper to write the ensamble config yaml file (models/ contains multiple logdirs/, logdir must contain the name 'unet')
python -m tf_semantic_segmentation.bin.model_server_config_writer -d models/ -c 'unet'
# start model server with written models.yaml
tensorflow_model_server --model_config_file=models.yaml --rest_api_port=8501

Compare models and ensemnble

python -m tf_semantic_segmentation.evaluation.compare_models -i logs/ -c 'taco' -data /hdd/datasets/ -d 'tacobinary'

Parameters:

• -i (directory containing models)
• -c (model name (directory name) must contain this value)
• -data (data directory)
• -d (dataset name)

Use —help to get more help

Using Code

from tf_semantic_segmentation.serving import predict, predict_on_batch, ensamble_prediction, get_models_from_directory
from tf_semantic_segmentation.processing.dataset import resize_and_change_color
image = np.zeros((128, 128, 3))
image_size = (256, 256)
color_mode = 0  # 0=RGB, 1=GRAY
resize_method = 'resize'
scale_mask = False # only scale mask when model output is scaled using sigmoid activation
num_classes = 3
# preprocess image
image = image.astype(np.float32) / 255.
image, _ = resize_and_change_color(image, None, image_size, color_mode, resize_method='resize')
# prediction on 1 image
p = predict(image.numpy(), host='localhost', port=8501, input_name='input_1', model_name='0')
#############################################################################################################
# if the image size should not match, the color mode does not match or the model_name does not match
# you'll most likely get a `400 Client Error: Bad Request for url: http://localhost:8501/v1/models/0:predict`
# hint: if you only started 1 model try using model_name 'default'
#############################################################################################################
# prediction on batch (for faster prediction of multiple images)
p = predict_on_batch([image], host='localhost', port=8501, input_name='input_1', model_name='0')
# ensamble prediction (average the predictions of multiple models)
# either specify models like this:
models = [
    {
        "name": "0",
        "path": "/home/user/models/mymodel/saved_model/",
        "version": 0, # optional
        "input_name": "input_1"
    },
    {
        "name": "1",
        "path": "/home/user/models/mymodel2/saved_model/",
        "input_name": "input_1"
    }
]
# or load from models in directory (models/) that contain the name 'unet'
models = get_models_from_directory('models/', contains='unet')
# returns the ensamble and all predictions made
ensamble, predictions = ensamble_prediction(models, image.numpy(), host='localhost', port=8501)

TFLite support

Convert the model

python -m tf_semantic_segmentation.bin.convert_tflite -i logs/mymodel/saved_model/0/ -o model.tflite

Test inference on the model

python -m tf_semantic_segmentation.debug.tflite_test -m model.tflite -i Harris_Sparrow_0001_116398.jpg