MNIST-classification-LeNet5

Goal

Classification of MNIST, FashionMNIST, KMNIST, and QMNIST datasets using the LeNet5 architecture.

Architecture

The famous LeNet5 architecture in implemented with Pytorch. It's composed of two convolutional layers (Conv + ReLU + MaxPool) followed by three fully connected layers (400-120-84-10) with ReLU and a Softmax as final activation layer. Each model is trained 10 times with different seeds (1,2,3,4,5,6,7,8,9,10)

Overall Performances

	MNIST	FashionMNIST	KMNIST	QMNIST
Train	99.362 ± 0.085%	89.701 ± 0.441%	98.518 ± 1.219%	99.356 ± 0.057
Test	99.445 ± 0.077%	89.495 ± 0.534%	96.255 ± 1.134%	99.328 ± 0.055

^{*Accuracy presented as mean ± 1.96 std}

---

MNIST

Digit recognition (0-9) from handwritten images (28x28 grayscale). The train set contains 60'000 examples and the test set contains 10'000 examples.

Setup

The model is trained 10 times with the following parameters:

N epoch	Learning rate	Batch Size	LR decay	Optimizer
150	0.002	128	0.98	ADAM

Sample Prediction

---

FashionMNIST

Cloth recognition from 28x28 grayscale images. There are 10 classes (T-Shirt/Top, Trouser, Pullover, Dress, Coat, Sandal, Shirt, Sneaker, Bag, Ankle boot). The train set contains 60'000 examples and the test set contains 10'000 examples.

Setup

The model is trained 10 times with the following parameters:

N epoch	Learning rate	Batch Size	LR decay	Optimizer
150	0.002	128	0.98	ADAM

Sample Prediction

---

KMNIST

Handwritten Japanese Hiragana recognition from 28x28 grayscale images. There are 10 Hiragana in this dataset: お (a), き (ki), す (su), つ (tu), な (na), は (ha), ま (ma), や (ya), れ (re), を (wo). The train set contains 60'000 examples and the test set contains 10'000 examples.

Setup

The model is trained 10 times with the following parameters:

N epoch	Learning rate	Batch Size	LR decay	Optimizer
150	0.002	128	0.98	ADAM

Sample Prediction

---

QMNIST

Same dataset of Handwritten digits as MNIST but with more test examples. The train set contains 60'000 examples and the test set contains 60'000 examples.

Setup

The model is trained 10 times with the following parameters:

N epoch	Learning rate	Batch Size	LR decay	Optimizer
150	0.002	128	0.98	ADAM

Sample Prediction