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
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 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.pyto crop images to224, 224as in the paper implementation.
- Use the
- run
model-tiramasu-67-func-api.pyorpython model-tirmasu-56.pyfor now to generate each models file. - run
python train-tirmasu.pyto start training:- Saves best checkpoints for the model and
data_loaderincluded for theCamVidDataset
- Saves best checkpoints for the model and
helper.pycontains two methodsnormalizedandone_hot_it, currently for the CamVid Task
Dataset:
-
In a different directory run this to download the dataset from original Implementation.
git clone [email protected]:alexgkendall/SegNet-Tutorial.git- copy the
/CamVidto here, or change theDataPathindata_loader.pyto the above directory
-
The run
python data_loader.pyto 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 |  |
 |
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:
146 Dec 29, 2022
33 Dec 18, 2022
86 Dec 2, 2022
44 Dec 12, 2022
606 Dec 21, 2022
260 Jan 3, 2023
32 Sep 21, 2022
1 Feb 8, 2022
148 Dec 21, 2022
82 Dec 29, 2022
14 Sep 13, 2022
7 Jan 24, 2022
115 Dec 28, 2022
158 Dec 21, 2022
1.6k Jan 7, 2023
85 Dec 7, 2022
8 Nov 2, 2022
12 Jan 6, 2023