Skip to content

Latest commit

 

History

History
 
 

tf-deeplab

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
readme.md
readme.md
 
 
tf_deeplab_frozen_graph_transforms.py
tf_deeplab_frozen_graph_transforms.py
 
 

readme.md

Tensorflow Deeplab Models

Model Details

  • Tensorflow Deeplab has scripts and pre-trained models for several popular model architectures. Several datasets such as Cityscapes, Pascal VOC and ADE20K are supported.

Graph transformation

  • We have slightly modified these models for faster execution.

  • Following changes were made from the existing frozen graph without any further training. These changes resulted in slight degradation in accuracy.

    • Input resolution changed to power of two. (E.g. 513x513 is changed to 512x512).
    • Resize factor changed to power of two. (All different resizing are made pow of 2).
    • Avg-pooling changed to power of two. (E.g 65x65 is changed to 64x64).
    • In these models, depthwise convolutions with dilation comes as {SpaceToBatch, depthwise convolutions, BatchToSpace} layers. To accommodate above changes, modifications were made in these layers.

  • In order to make the above changes, we have provided a script in tf_deeplab_frozen_graph_transforms.py. Follow the steps below to run the script. The script uses graph transformation tool that was readily available in TF 1.XX.

    pip install tensorflow==1.15 

    wget http://download.tensorflow.org/models/deeplabv3_mnv2_dm05_pascal_trainval_2018_10_01.tar.gz
tar xvf deeplabv3_mnv2_dm05_pascal_trainaug_2018_10_01.tar.gz
python  frozen_graph_transforms.py --pb_path ./deeplabv3_mnv2_dm05_pascal_trainaug/frozen_inference_graph.pb --input_nodes sub_7 --output_nodes ArgMax --input_shape 512 512

  • The above script has been validated for several models.

Results

Dataset Model Name Input Size GigaMACs MeanIoU%
- Our modified models
VOC2012 deeplabv3_mnv2_dm05 512x512 2.77 66.94
VOC2012 deeplabv3_mnv2 512x512 8.60 72.66
VOC2012 deeplabv3_xception 512x512 171.77 81.74
- Original DeepLab models
Cityscapes MobileNetV2+DeepLab 769x769 21.27 70.71
Cityscapes MobileNetV3+DeepLab 769x769 15.95 72.41
Cityscapes Xception65+DeepLab 769x769 418.64 78.79