Image Classification using a CNN on Flowers🌷

In this notebook, we will train a CNN-based image classifier to classify a couple thousand images of flowers, loaded from a directory on google drive.

We will compare three pretrained model: VGG16, ResNet50, Inception

The training will take data augmentation.

• Convnet without data augmentation: 63%
• Convnet with data augmentation: 70%
• VGG: 80% (Winner)
• ResNet50: 67%
• InceptionV3: 75%

• The model nearly overfitting at 5 epoch.
• The avg accuracy is 62%

• Accuracy: 70%

• Accuracy: 80%

• Accuracy: 67%

• Accuracy: 65%

Let's take a look at some of the images

• VGG16 converges quicker than ResNet50
• The training data and validation data of VGG16 improves steadly, while ResNet50 is hard to improve validation accuracy at first, but imporves quickly after a certain epoch
• The InceptionV3 has very low loss but it's validation accuary still not improve