Digit Recognizer

Python Jupyter Notebook with Convolutional Neural Network digit recognizer implemented in Keras. It's Google Colab ready.

Part of Kaggle competition.

Submitted Kernel with 0.995 score.

Data

Dataset: MNIST Handwritten digits

Description: Classification of handwritten digits, 10 classes (0-9).

Training: 37.8k (0.9) images

Validation: 4.2k (0.1) images

Testing: 28k images

Model

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d_22 (Conv2D)           (None, 28, 28, 32)        832       
_________________________________________________________________
conv2d_23 (Conv2D)           (None, 28, 28, 32)        25632     
_________________________________________________________________
max_pooling2d_11 (MaxPooling (None, 14, 14, 32)        0         
_________________________________________________________________
dropout_7 (Dropout)          (None, 14, 14, 32)        0         
_________________________________________________________________
conv2d_24 (Conv2D)           (None, 14, 14, 64)        18496     
_________________________________________________________________
conv2d_25 (Conv2D)           (None, 14, 14, 64)        36928     
_________________________________________________________________
max_pooling2d_12 (MaxPooling (None, 7, 7, 64)          0         
_________________________________________________________________
dropout_8 (Dropout)          (None, 7, 7, 64)          0         
_________________________________________________________________
flatten_4 (Flatten)          (None, 3136)              0         
_________________________________________________________________
dense_8 (Dense)              (None, 8192)              25698304  
_________________________________________________________________
dropout_9 (Dropout)          (None, 8192)              0         
_________________________________________________________________
dense_9 (Dense)              (None, 2048)              16779264  
_________________________________________________________________
dropout_10 (Dropout)         (None, 2048)              0         
_________________________________________________________________
dense_10 (Dense)             (None, 10)                20490     
=================================================================
Total params: 42,579,946
Trainable params: 42,579,946
Non-trainable params: 0
_________________________________________________________________

Training

Results

Kaggle score: 0.995