cnn using numpy but vectorized

Vectorized implementation of the Layers/Building Blocks of a Convolutional Neural Network for faster training and prediction used to Classify Handwritten Bengali Digits.

Dataset used from a Kaggle Competition:

Bengali Handwritten Digit Recognition Competition by Bengal.ai

Layers implemented:

1. Convolution: Four (hyper)parameters: a. Number of output channels b. Filter dimension c. Stride d. Padding
2. ReLU Activation
3. Max-pooling: Two parameters: a. Filter dimension b. Stride
4. Flattening layer: Converts a (series of) convolutional filter maps to a column vector.
5. Fully-connected layer: a dense layer. One parameter: output dimension.
6. Softmax: Converts final layer projections to normalized probabilities.

Full Specifications of build: cnn from scratch

Model and Data

• The Architecture used was the standard LeNet-5
• The images were preprocessed as follows:
  1. resize to 64 by 64
  2. invert from colors of pixel
  3. transpose to have the shape (batch, channels, image_dimension, image_dimension)

Training

Training was done using combined images from 'training-a', 'training-b' and 'training-c'. Train-Test was split 80-20. Training was done using mini-batch gradient descent.

• Epochs: 24
• batch size: 32
• Learning Rate: 0.1 (tuned)
• batch size of mini-batch: 32

Validation

• Accuracy: 0.9525473399458972
• Validation Loss: -7.247408912811393
• Macro f1 score: 0.9523544234682374 ,
• Training Loss: 2.0793054652897394e-05

Confusion Matrix:

Metric Graphs:

Independent Testing

Testing was done on 'training-d' as per instructions of project supervisor

• Accuracy: 0.9540704070407041,
• Macro f1 score: 0.9541251631006855

Full Report on the Results: Training Results