README

DConvNet-CIFAR100

Delphi ConvNet - convolutional neural network for CIFAR100

This is a ConvNet to Delphi 10 (object pascal) translation.

A Convolution Neural Network in Delphi.

This net learns to categorize images to 100 different categories. The pretrained has an acc of 70%.

Download the CIFAR-100 binaries (Images) from "https://www.cs.toronto.edu/~kriz/cifar-100-binary.tar.gz" and move them to "Data" directory.

The "Common" directory contains all CNN algorithm sources. The CIFAR_100/work Directory cointais a test project for CIFAR-100. What CIFAR-100 is, you can read at https://www.cs.toronto.edu/~kriz/cifar.html

Layer: input, fc, dropout, softmax, regression, conv, pool, relu, sigmoid, tanh, maxout, svm

Methods: adam, adagrad, windowgrad, adadelta, nesterov

Spec: Batch, Chunking

-- Layer: input, fc, dropout, softmax, regression, conv, pool, relu, sigmoid, tanh, maxout, svm

[Structure]

L001=Layer:Input, Name:input, Size_x:32, Size_y:32, Depth:3

L002=Layer:Conv, Name:conv 1, Filter_Size: 3, Filter_Count:20, Stride:0, Pad:0, Activation:relu

L003=Layer:Pool, Name:pool 1, Size:2, Stride:2

L004=Layer:Conv, Name:conv 2, Filter_Size: 3, Filter_Count:20, Stride:1, Pad:2, Activation:relu

L005=Layer:Pool, Name:pool 2, Size:2, Stride:2

L006=Layer:Conv, Name:conv 3, Filter_Size: 3, Filter_Count:20, Stride:1, Pad:2, Activation:relu

L007=Layer:Conv, Name:FC, Size:100, Activation:tanh

L008=Layer:softmax, Name:softmax,Size:100

[Options]

--Methods: adam, adagrad, windowgrad, adadelta, nesterov

method = adagrad

batch_size = 15

learning_rate = 0.01

momentum = 0.9

l1_decay = 0

l2_decay = 0.0001

ro = 0.95

eps = 0.0000001

[Chunk]

// ========================================================================= // DER CHUNK // ein Block an Infomationen, der sicher gelernt sein muss, bevor der nächste Chunk gelernt werden kann // wird eine Information nicht gelernt, so kommt sie in den nächsten Chank zum wiederholten Lernen // Chunklernen einschalten

ChunkEnabled = 0

// Anzahl der zu lernenden Bilder

ChunkSize = 150

// Anzahl der Wiederholungen

ChunkRepetitions = 4

// Mindest-W-keit, dass eine Information als gelernt akzeptiert wird

ChunkAccLikeliHood = 0.8

// Mindest-W-keit, dass eine Information als nicht gelernt akzeptiert wird

ChunkNonAccLikeliHood = 0.2

[Data_Filenames]

Labels=.\Data\cifar-100-binary\coarse_label_names.txt

DetailLabels=.\Data\cifar-100-binary\fine_label_names.txt

TrainData=.\Data\cifar-100-binary\train.bin

TestData=.\Data\cifar-100-binary\test.bin

Weights= .\Weights.net

Results= .\Results.csv