MNIST-LeNet

MNIST ("Modified National Institute of Standards and Technology") is the de facto “hello world” dataset of computer vision. Since its release in 1999, this classic dataset of handwritten images has served as the basis for benchmarking classification algorithms. As new machine learning techniques emerge, MNIST remains a reliable resource for researchers and learners alike.

Validation Accuracy Achieved: 99.32%

---

MNIST dataset

The MNIST database is available at https://www.kaggle.com/c/digit-recognizer/data

The MNIST database is a dataset of handwritten digits. Each image is represented by 28x28 pixels, each containing a value 0 - 255 with its grayscale value.

The digits have been size-normalized and centered in a fixed-size image.
Kaggle Link

LeNet:

LeNet's design:

Tweaks that I made:

• Two stacked 3x3 filters replace the single 5x5 filters. These become nonlinear 5x5 convolutions
• A convolution with stride 2 replaces pooling layers. These become learnable pooling layers.
• ReLU activation replaces sigmoid.
• Batch normalization is added
• Dropout is added
• RMSProp used in place of Adam optimizer

Use saved model:

Download this
Extract model_data.zip

from tensorflow import keras
model = keras.models.load_model($model_path)

You can also use this sample

References:

• LeNet Architecture
• Keras optimizer