SVHN is a popular real-world image dataset which shares some similarities with MNIST dataset. It requires minimal data pre-processing and formatting. In this dataset, all the images are RGB and in fixed shape of 32-by-32 pixels. The dataset consists of 73,257 and 26,032 digits with total 10 classes for training and testing respectively.
- SVHN_CNN.ipynb : with convolution neural network
- los_acc.jpg : in-sample and out of sample accuracy with loss
- prediction.png : predictions based on the testing data
- SVHN_CNN.h5 : we saved our model
- logs.zip : logs generated by tensorboard for analyzing the model
- We have evaluated our model with testing data and plotted several images with their predictions with percentage (out of 100) as follows:
- We also have plotted some graphs of in-sample and out of sample accuracy with loss based on the epoch value as follows:
