Image Classification using Transfer Learning

Project Description:

In this project, I have implemented "Transfer Learning" to train a deep learning model for the task where training data are very limited. Here you are given a data set with ten classes (Ten different monkey species) with only 140 images per class. First, I trained a simple convolutional neural network using these images and test the accuracy of the model using the validation set. Then, I used the pre-trained VGG16 as a feature extractor to perform the transfer learning to boost the efficiency of the model. Transfer learning significantly improved the efficiency of the model.

Pipeline

• Here you are given a data set with ten classes (ten different monkey species) with only 140 images per class.
• First, I designed Convolutional Nueral Network and trained it using this data. Because of the low number of training samples, the test accuracy was lower than expected.
• This is where transfer learning can help. Transfer Learning can be used for deep learning applications where either data or computational power are restricted.
• Next, I used pretrained ResNet34 model as a feature extractor to perform the transfer learning.
• To do this, I removed the last fully connected layers of the pretrained model and replace it with the untrained fully connected layer s to classify the monkey species.
• During training, I freezed the convolutional layer parameters so that they remain the same and only update the fully connected layers at the end. In this way, the convolutional layers act as generalized feature extractors that have already been pretrained on millions of other images (that werent necessarily all monkeys) while the fully connected layers are able to take these features and classify our images.
• This showed substential boost in accuracy despite having the same amount of training sample.
• To further boost the performance of the network, I unfreezed the convolutional layers and trained for a few more epochs with a small step size to fine tune the network to extract even more predictive power.

Dataset

10 monkey species

Results

1. Performance of the CNN without Transfer Learning

   Sr No	Hyperparameters	Values/Type
   1	Optimizer	Adam
   2	Learning Rate	10-5
   3	Number of Epochs	100
   4	Batch Size	16

• Testing Accuracy: 45.58%

2. Performance with the Transfer learning with only Fully connected layer tuned

   Sr No	Hyperparameters	Values/Type
   1	Optimizer	Adam
   2	Learning Rate	10-5
   3	Number of Epochs	100
   4	Batch Size	16

• Testing Accuracy: 95.58%

3. Performance with the Transfer learning with all the layer tuned

   Sr No	Hyperparameters	Values/Type
   1	Optimizer	Adam
   2	Learning Rate	10-5
   3	Number of Epochs	100
   4	Batch Size	16

• Testing Accuracy: 98.38%

Requirement

Python 2.0 or above

License

Copyright (c) Feb 2023 Abhijit Mahalle