Skip to content

Latest commit

 

History

History
executable file
·
95 lines (79 loc) · 4.41 KB

README.org

File metadata and controls

executable file
·
95 lines (79 loc) · 4.41 KB
Raw

Task Script

Each config file represents the single task, which might have several trainings according to the paper figures. You could simply run batch.sh to reproduce all the figures from our paper, or either go to each task script described in batch.sh for further research.

Easy life

Each run is according to the specific task generating figure in the paper. To reproduce all the experiments that we have in the paper, you could run our batch script by the following instruction:

# in bash
git clone git@gitlab.com:gladiator8072/hsic-bottleneck.git 
source env.sh
batch.sh

After the batch training, you’ll get the following result in your assets/exp folder, and please compare to our sample results under assets/sample folder: For more information, please visit the running procedure page (config/READMD.md [link](config/README.md)) and (bin/README.md [link](bin/README.md)) for more information.

# there's fig* context at beginning for convenience 
[mawand@machine HSIC-bottleneck]$ ls -l assets/exp | awk '{print $9}'
fig2a-varied-activation-hsic_xz-mnist.pdf
fig2b-varied-activation-hsic_yz-mnist.pdf
fig2c-varied-activation-acc-mnist.pdf
fig2d-varied-depth-hsic_xz-mnist.pdf
fig2e-varied-depth-hsic_yz-mnist.pdf
fig2f-varied-depth-acc-mnist.pdf
fig3a-needle-1d-dist-backprop.pdf
fig3b-needle-1d-dist-hsictrain.pdf
fig4-hsic-solve-actdist-cifar10.pdf
fig4-hsic-solve-actdist-fmnist.pdf
fig4-hsic-solve-actdist-mnist.pdf
fig5-hsic-solve-cifar10-train-acc.pdf
fig5-hsic-solve-fmnist-train-acc.pdf
fig5-hsic-solve-mnist-train-acc.pdf
fig6a-varied-epoch-acc-mnist.pdf
fig6b-varied-epoch-loss-mnist.pdf
fig7a-varied-dim-acc-mnist.pdf
fig7b-sigma-combined-mnist-sigmacomb-train-acc.pdf

Difficult life

Procedure

HSIC-bottleneck is task oriented project that make each tasks following the similar action. In general, you’ll take the following steps to achieve your specific task

  • go to the project root directory
  • setting the environment source env.sh
  • run command prefixed with task_
  • load the given config [config_path]
  • training,
  • save the logs under `./assets/logs` (optional)
  • save figures in `./assets/exp` (with prefixed “fig”, corresponding to the paper figures)

Tasks

Here we’re showing all the tasks that produced in the paper.

varied-activation (fig2a-c)

  • nHSIC monitoring in backprop training. It shows $\text{nHSIC}(X,Z_L)$, $\text{nHSIC}(Y,Z_L)$ would decrease and increase

respectively. As the evidence the network learn the dependency from output and reduce it from input

  • task_varied-act.sh - link

varied-depth (fig2d-f)

  • nHSIC monitoring in backprop training. It shows $nHSIC(X,Z_L)$, $nHSIC(Y,Z_L)$ would decrease and increase

respectively. As the evidence the network learn the dependency from output and reduce it from input

  • task_varied-dim.sh - link

needle [fig3a-b]

  • Originally placed in our first arxiv in fig3. This experiment shows how well the HSIC-trained network

separates the classed signals in the scalared network output (tanh case)

  • task_needle.sh - link

hsicsolve (fig3, fig4)

  • Unformat-training, where we use HSIC-trained network to solve the classification problem. The output

of the network is no longer ordered vector as in one-hot label matrix. The particular image category might go to some other entry.

  • task_hsicsolve.sh - link

varied-epoch (fig5a-b)

  • The boosting test which the long HSIC-trained network allows faster convergence of format-train.
  • task_varied-ep.sh - link

varied-dim (fig6a)

  • The capacity of HSIC-trained network experiments. The aim of this task is we hope large network can hold

more information from the input, making the format-training better.

  • task_varied-dim.sh - link

sigma-combined (fig6b)

  • The capacity of HSIC-trained network experiments. First of all process 3 HSIC-trained network with different

sigma scale and produce 3 format-training results. Then load and average those 3 networks for format-training, should be better than those individually

  • task_combsig.sh - link

conv-based resnet (fig7)

  • The HSIC-bottlneck on the ResNet based architecture, where the objective are applied on each residual block output
  • task_resconv.sh - link