FPC Neural Networks + Places2: 1.8 million training images - TESTING HAS STARTED

[schuler]

  Hello everyone
I think that the importance of the moment deserves a new topic.

Previous methods of the API load the full set of training images into memory. This was ok or even stretched with:

• CIFAR 10 and CIFAR100 (60k images): https://www.cs.toronto.edu/~kriz/cifar.html
• Tiny image net 200 (100k images): http://cs231n.stanford.edu/tiny-imagenet-200.zip
• Plant Village (54k images): https://data.mendeley.com/datasets/tywbtsjrjv/1/files/d5652a28-c1d8-4b76-97f3-72fb80f94efc

For a dataset with 1.8 million images, keeping everything into RAM wouldn't work. Places2 standard has 1.8 million training images:
http://places2.csail.mit.edu/download.html

For my own tests, I'm downloading "Small images (256 * 256) with easy directory structure with 21GB".

A solution for this is currently in testing. The folder structure is loaded into RAM with:

Code: Pascal  [Select][+][-]

1.    
2.     FTrainingFileNames, FValidationFileNames, FTestFileNames: TFileNameList;
3. ...
4.     ProportionToLoad := 1;
5.     CreateFileNameListsFromImagesFromFolder(
6.       FTrainingFileNames, FValidationFileNames, FTestFileNames,
7.       {FolderName=}'places_folder/train', {pImageSubFolder=}'',
8.       {TrainingProp=}0.9*ProportionToLoad,
9.       {ValidationProp=}0.05*ProportionToLoad,
10.       {TestProp=}0.05*ProportionToLoad
11.     );

There is a new fitting class capable of working with the above:

Code: Pascal  [Select][+][-]

1.     NeuralFit := TNeuralImageLoadingFit.Create;
2.     ...
3.     NeuralFit.FitLoading({NeuralNetworkModel}NN, {ImageSizeX}256, {ImageSizeY}256, FTrainingFileNames, FValidationFileNames, FTestFileNames, {BatchSize}256, {Epochs}100);

Is it bug free? I don't know. It's currently in testing. My current test case is modifying a source code intended for Plant Village dataset by just switching the folder name:
https://github.com/joaopauloschuler/neural-api/blob/master/examples/SimplePlantLeafDisease/SimplePlantLeafDiseaseLoadingAPI.pas

When creating the model (the neural network), the number of classes comes straight from the dataset:

Code: Pascal  [Select][+][-]

1. NN.AddLayer([
2.       TNNetInput.Create(FSizeX, FSizeY, 3),
3.       ...
4.       TNNetFullConnectLinear.Create(FTrainingFileNames.ClassCount),
5.       TNNetSoftMax.Create()
6.     ]);

Wish everyone happy pascal coding. May the source be with you.

[Editor]

Hi. it looks fantastic. Could yo tell more about the whole code?

[Laksen]

Really cool library, and excellent information on github

I'm curious about whether you have support for fixed point datatypes of any kind?