https://github.com/gallo13/pattern-rec-neural-networks-image-classification_2

Image Classification with Convolutional Neural Networks: Imablance, Performance Evaluation, Ensembling (Final 2)
https://github.com/gallo13/pattern-rec-neural-networks-image-classification_2

cnn convolutional-neural-networks ensemble-machine-learning image-classification image-processing jupyter-notebook keras tensorflow

Last synced: 2 months ago
JSON representation

Image Classification with Convolutional Neural Networks: Imablance, Performance Evaluation, Ensembling (Final 2)

• Host: GitHub
• URL: https://github.com/gallo13/pattern-rec-neural-networks-image-classification_2
• Owner: Gallo13
• Created: 2020-05-22T18:48:21.000Z (about 6 years ago)
• Default Branch: master
• Last Pushed: 2020-12-30T19:35:43.000Z (over 5 years ago)
• Last Synced: 2025-01-26T00:34:27.582Z (over 1 year ago)
• Topics: cnn, convolutional-neural-networks, ensemble-machine-learning, image-classification, image-processing, jupyter-notebook, keras, tensorflow
• Language: Jupyter Notebook
• Homepage:
• Size: 2.7 MB
• Stars: 1
• Watchers: 2
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README

Awesome Lists containing this project

README

          
Final Project Part 2:

Problem 1:

You can continue with the dataset you used for FinalProjectPart1 – Problem 2 or you can pick-up another dataset. Requirements to the 

dataset if you pick-up a new one are the same: not less than 4 classes, not less than 1000 images for each of the classes in the training 

set, not less than 300 images for each of the classes in the validation set, not less than 60 images for each of the classes in the test 

set). Datasets as MNIST, CIFAR-10, CIFAR-100 or MNIST-like and CIFAR-like are excluded. Use dataset normalization. As the datasets are 

usually bigger than the specified size above, you can use only part of it as in dogs-vs-cats.

When displaying numerical results use instead of only two digits after the decimal point, four digits after the decimal point. When 

possible use tabular and graphics for displaying all results what you have done in Part 1 of Final Project.

A) Cut the original dataset if necessary and split the new one into three subsets: traing set, valdiation set, and test set. Kepp an 

   apploximation ratio 70%-20%-10%

   Make a display of the three classes like a graph plus the numerical way:

   

   Number of Samples Train: 3091

   Number of Samples ValidationL 616

   Number of Samples Test: 155

   

   *graph*

B) Use Sequential model and build the topology of your CNN with not less than four convolutional layers. Use data normalization, batches 

   and batch normalization, Glorot initialization, RELU activation function, shuffling, number of epochs not less than 50. All other 

   elements as well as parameters of the CNN are up to you.

C) Display the number of patterns in each class for the training and validation sets like in the example shown below. Do the same for 

   the test set as well. In this case both the training and validation sets are imbalanced.

D) Perform augmentation for only training and validation sets and for only the underrepresented classes. Make the same display as above 

   only for training and validation sets after augmentation. Do not implement augmentation for the test set.

E) Make experiments with three different optimizers by your choice (exclude Gradient Descent and SGD with or without momentum) and plot 

   the train and validation accuracy and train and validation loss (see the example below). Pick up the thickness of the lines so that 

   the graphics are well visible. Show the optimizer on the figures what is not done below. Increase the number of epochs if the learning 

   accuracy shows the tendency of further increase.

F) Evaluate the performance of each implemented by you optimizer using the test set. Display the normalized confusion matrix, precision, 

   recall, F1-score, support. Draw ROC and AUC.

G) Visualize several channels in some intermediate activations and display at least 12 images from the test set with correct and wrong 

   predictions.

H) Build ensemble based on the results of implemented by you optimizers and evaluate the ensemble performance.

=======================================================================================================================================================

DATA DESCRIPTION

Original dataset is from https://www.kaggle.com/chrisfilo/fruit-recognition and consists of 44406 fruit images. These images are from 15 different kind of fruits. Images were made in a lab environment, with different angles, lightning, variation and etc. Each image has the size of 320x358. However, for this lab we took only 

part of the data. We collected around 1000 pictures from each class for our training set, around 300 from each class for our validation class and around 60 from each class for our testing set.

- 12526 images in training set

- 3922 images in testing set

- 827 images in testing set

- 12 classes (apple, banana, carambola, guava, kiwi, mango, muskmelon, orange, peach, pear, persimmon and pitaya)