dk4120/neural_network_parameter_dataset_lenet5_binary

Neural Networks Dataset for Hypernetworks Research

Summary

This repository contains a dataset of neural networks, designed for the purpose of hypernetworks research. The dataset includes 10,610 neural networks trained for binary image classification separated into 10 classes, such that each class contains 1,061 different neural networks that can identify a certain ImageNette V2 class from all other classes. The classification models used a LeNet-5 framework with each model containing 91,481 parameters. A computing cluster of over 10,000 cores was used to generate the dataset. Basic classification results show that the neural networks can be classified with accuracy of 87%, indicating that the differences between the neural networks can be identified by supervised machine learning algorithms in accuracy better than mere chance. The ultimate purpose of the dataset is to enable hypernetworks research.

Specifications

This repository contains the individual model files, consolidated listing of parameters, and training/validation loss and accuracy plots for each model.

Directory Listing

Model Weights

Contains individual model parameters saved as hdf5 files

Plots

Contains plots for training/validation loss and accuracy for each model

Pickle Files

Consolidated parameters saved into a single file for all 10,610 models.

Modelwise

Parameters by model in dictionary format

• Keys: model, parameters_by_layer, parameters_flat
• Example:

{
  'model': 'church_649', 
  'parameters_by_layer': {
    'conv2d/conv2d/bias:0': tensor([ 0.0572, 0.0594, -0.0007, -0.0440, 0.0227, 0.0155]), 
    'conv2d/conv2d/kernel:0': tensor([[[[ 2.0242e-01, -8.1340e-02, -1.1604e-01, 6.9871e-02, -1.4720e-01,...
  }, 
  'parameters_flat': tensor([ 0.0572, 0.0594, -0.0007, ..., 0.2816, 0.3210, -0.1104])
}

Layerwise

Parameters by layer for all models in a given class

• Keys: class, parameters
• Example:

{
  'class': 'cassette_player', 
  'parameters': {
    'conv2d': [array([ 2.81625651e-02, -7.09968358e-02, -1.40507936e-01, -9.32385959e-03, -1.10361770e-01, 1.00062557e-01, -3.23645794e-03, 1.01349339e-01, 5.19516990e-02, -8.61838460e-02, 1.42671123e-01, -1.31018981e-01,...
    'dense_1': [array([-0.04802695, -0.26561895, -0.13131969,  0.13551578, -0.14921758, -0.15271764,  0.226575...
  }
}

Performance Metrics

Class	Accuracy Min	Accuracy Max	Accuracy Avg	Precision Min	Precision Max	Precision Avg	Recall Min	Recall Max	Recall Avg	F1 Min	F1 Max	F1 Avg
tench	0.932	0.949	0.942	0.663	0.872	0.769	0.478	0.672	0.587	0.604	0.713	0.665
english_springer	0.894	0.920	0.911	0.473	0.779	0.619	0.134	0.496	0.315	0.219	0.522	0.412
cassette_player	0.916	0.937	0.928	0.544	0.845	0.675	0.272	0.569	0.408	0.407	0.581	0.506
chain_saw	0.897	0.908	0.903	0.349	0.933	0.576	0.008	0.127	0.068	0.015	0.214	0.120
church	0.901	0.921	0.911	0.533	0.844	0.666	0.134	0.438	0.301	0.226	0.518	0.411
french_horn	0.886	0.907	0.900	0.300	0.634	0.507	0.008	0.353	0.186	0.015	0.406	0.265
garbage_truck	0.892	0.927	0.917	0.464	0.846	0.645	0.193	0.584	0.395	0.303	0.565	0.484
gas_pump	0.870	0.901	0.892	0.295	0.684	0.480	0.062	0.234	0.151	0.109	0.306	0.228
golf_ball	0.898	0.919	0.912	0.496	0.836	0.658	0.128	0.434	0.298	0.222	0.494	0.407
parachute	0.920	0.944	0.937	0.583	0.875	0.773	0.313	0.674	0.532	0.448	0.685	0.626
All Classes	0.870	0.949	0.915	0.295	0.933	0.637	0.008	0.674	0.324	0.015	0.713	0.412