Creating an Image Classifier using transfer learning as a part of the Deep Learning Course. A pretained 19 layer VGG model with Batch Normalization is used. The classifier is modified and the feature layer is kept frozen.
The data set consists of 3726 images divided among 8 classes with slight variations in number. All images are of different resolutions and PIL library is used in resizing them to 3 x 224 x 224 as it is the preferred input dim for the vgg modles. The prediciton set consists of 160 images in total. The train test data split is in random with a ratio of 80:20.
Using vgg19 with batch normalization resulted in a prediciton score of 92.5%.
The corresponding model has the below stats,
- Train Loss: 0.0012
- Train Acc: 97.58
- Validation Loss: 0.006552678989820762
- Validation Acc: 88.73994638069705