/gdbls-pytorch

an branch of original gdbls

Primary LanguagePython

Grand Descent Board Learning System (GDBLS)

Hello, this is the detailed implement for GDBLS, an efficient model structure for deeplearning using board-wise methods.

Model Structure

Model Structure

Outcome

Outcome

Implements

  1. model/gdbls_conv3block3.py for 32*32.
  2. model/gdbls_conv3block3_dogcatversion for any resolutions.
  3. resnet_fpn.py The model was used for comparison with the resnet-based feature pyramid structure to justify the feature blocks we designed.
  4. gdbls_conv3block3_noEB.py The model was used to demonstrate the feasibility and necessity of our feature enhancement strategy.

Provided experiments

  1. MNIST (1, 32, 32)
  2. SVHN (3, 32, 32)
  3. CIAFR10 (3, 32, 32)
  4. CIFAR100 (3, 32, 32)
  5. CATORDOG (3, 200, 200)

Steps to perform the experiment

To carry out experiment on specific dataset, just follow the folling phases:

  1. download dataset from this url: [url]
  2. extract files in ./dataset folder
  3. deploy the environment using pip for two sub-projects: GD_BLS and statserver.
  4. open ./statserver and run run.py using flask.
  5. in main.py, select the dataset and model you need to validate, specifying the name of this experiment and the number of repetitions, let it run.
  6. After the experiment completed, you will see many records in ./statserver/saves/data.csv. In analyse.py, run an analyse with information provided from main.py and you can analyse the detailed experiment information.

Something to pay attention to

  1. for each dataset, the configs are written to ./configs/[dataset_name].yaml
  2. ./logs/[dataset_name]/ saves the confusion matrix of the newest experiment.