Techniques of generic and optimized Image Classification using python and Tensorflow-GPU
| Dataset Name | Description | Training Set (size) | Test Set (size) |
|---|---|---|---|
| DeepSat-6 | Satellite images | 324k | 81k |
| Dataset Name | Description | Traning Accuracy | Test Accuracy | Type of Norm |
|---|---|---|---|---|
| DeepSat-6 | Slope is less for b-norm | 0.9729 | 0.9692 | Batch Norm |
| g-norm converges faster for training | 0.9704 | 0.9657 | Group Norm |
- Linux Ubuntu 18.04
- Tensorflow 1.12 with GPU enabled
- CUDA 10 Toolkit with corresponding NVDIA drivers need to be installed
- I use a 1080Ti Nvidia GPU - currently state of the art, I do not use SLI, just 1 of these
- Current Implementation uses data from the DeepSat-6 from kaggle to classify images into the 6 classes used
- All images are used , with a training set of 324k and test set of 81k
- .h5 files of training and test sets are saved as numpy arrays using paperspace.com as I have limited RAM to handle the large dataset
- We use Batch Normalization and Group Normalization techniques to classify faster
- Try with bigger and varied datasets, such as imagenet also and compare performance
- I am integrating the COCO and CIFAR-100 dataset, to train