This is an implementation of a convolutional neural network for recognizing hand written digits using the MNIST dataset. This model attains a validation accuracy of about 99.2% is obtained after training for 12 epochs. The model architecture and weights are saved in the files model_architecture.json and model_weights.h5. Note that these weights are compatible only with the Tensorflow backed.
To train the model run train.py. The file test.py generates a file predictions.csv which contains the predicted labels to the images in the test set. This file can be used for submission at Kaggle. display_random.py displays 25 random images from the test set along with their predicted labels. Here is an example:
- Python 2.7
- Tensorflow
- Keras
- h5py
- numpy
- matplotlib
- pandas
-
The model is trained on the MNIST dataset downloaded from Kaggle.
-
The file
train.csvcontains pixel intensity values as flattened vectors for 42000 images and their corresponding labels. Similarly,test.csvhas pixel intensity values for 28000 unlabelled images.
