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 OutcomeImplements
- model/gdbls_conv3block3.py for 32*32.
- model/gdbls_conv3block3_dogcatversion for any resolutions.
- resnet_fpn.py The model was used for comparison with the resnet-based feature pyramid structure to justify the feature blocks we designed.
- gdbls_conv3block3_noEB.py The model was used to demonstrate the feasibility and necessity of our feature enhancement strategy.
Provided experiments
- MNIST (1, 32, 32)
- SVHN (3, 32, 32)
- CIAFR10 (3, 32, 32)
- CIFAR100 (3, 32, 32)
- CATORDOG (3, 200, 200)
Steps to perform the experiment
To carry out experiment on specific dataset, just follow the folling phases:
- download dataset from this url: [url]
- extract files in ./dataset folder
- deploy the environment using pip for two sub-projects: GD_BLS and statserver.
- open ./statserver and run run.py using flask.
- 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.
- 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
- for each dataset, the configs are written to ./configs/[dataset_name].yaml
- ./logs/[dataset_name]/ saves the confusion matrix of the newest experiment.