/One-Hundred-Layers-Tiramisu

Keras Implementation of The One Hundred Layers Tiramisu: Fully Convolutional DenseNets for Semantic Segmentation by (Simon Jégou, Michal Drozdzal, David Vazquez, Adriana Romero, Yoshua Bengio)

Primary LanguageJupyter NotebookMIT LicenseMIT

The One Hundred Layers Tiramisu: Fully Convolutional DenseNets for Semantic Segmentation:

Work In Progress, Results can't be replicated yet with the models here

  • UPDATE: April 28th: Skip_Connection added thanks to the reviewers, check model model-tiramasu-67-func-api.py

feel free to open issues for suggestions:)

  • Keras2 + TF used for the recent updates, which might cause with some confilict from previous version I had in here

What is The One Hundred Layers Tiramisu?

  • A state of art (as in Jan 2017) Semantic Pixel-wise Image Segmentation model that consists of a fully deep convolutional blocks with downsampling, skip-layer then to Upsampling architecture.
  • An extension of DenseNets to deal with the problem of semantic segmentation.

Fully Convolutional DensNet = (Dense Blocks + Transition Down Blocks) + (Bottleneck Blocks) + (Dense Blocks + Transition Up Blocks) + Pixel-Wise Classification layer

model

The One Hundred Layers Tiramisu: Fully Convolutional DenseNets for Semantic Segmentation (Simon Jégou, Michal Drozdzal, David Vazquez, Adriana Romero, Yoshua Bengio) arXiv:1611.09326 cs.CV

Requirements:

  • Keras==2.0.2
  • tensorflow-gpu==1.0.1
  • or just go ahead and do: pip install -r requirements.txt

Model Strucure:

  • DenseBlock: BatchNormalization + Activation [ Relu ] + Convolution2D + Dropout

  • TransitionDown: BatchNormalization + Activation [ Relu ] + Convolution2D + Dropout + MaxPooling2D

  • TransitionUp: Deconvolution2D (Convolutions Transposed)

model-blocks

Model Params:

  • RMSprop is used with Learnining Rete of 0.001 and weight decay 0.995
    • However, using those got me nowhere, I switched to SGD and started tweaking the LR + Decay myself.
  • There are no details given about BatchNorm params, again I have gone with what the Original DenseNet paper had suggested.
  • Things to keep in mind perhaps:
    • the weight inti: he_uniform (maybe change it around?)
    • the regualzrazation too agressive?

Repo (explanation):

  • Download the CamVid Dataset as explained below:
    • Use the data_loader.py to crop images to 224, 224 as in the paper implementation.
  • run model-tiramasu-67-func-api.py or python model-tirmasu-56.py for now to generate each models file.
  • run python train-tirmasu.py to start training:
    • Saves best checkpoints for the model and data_loader included for the CamVidDataset
  • helper.py contains two methods normalized and one_hot_it, currently for the CamVid Task

Dataset:

  1. In a different directory run this to download the dataset from original Implementation.

    • git clone git@github.com:alexgkendall/SegNet-Tutorial.git
    • copy the /CamVid to here, or change the DataPath in data_loader.py to the above directory

  2. The run python data_loader.py to generate these two files:

    • /data/train_data.npz/ and /data/train_label.npz
    • This will make it easy to process the model over and over, rather than waiting the data to be loaded into memory.
  • Experiments:
Models Acc Loss Notes
FC-DenseNet 67 model-results model-results 150 Epochs, RMSPROP

To Do:

[x] FC-DenseNet 103
[x] FC-DenseNet 56
[x] FC-DenseNet 67
[ ] Replicate Test Accuracy CamVid Task
[ ] Replicate Test Accuracy GaTech Dataset Task
[ ] Requirements

  • Original Results Table:

    model-results