FashionMNIST

Installation

• Create and activate new virtual env with python3
• pip install -r requirements.txt

Usage

• Set parameters (see main.py)
• ./run_train.sh

Experiments

Architectures

CNN

2 Conv Blocks (Conv, ReLU, max pool) + 2 FC. This was meant as simple baseline, although given its very small footprint accuracy is still competetive.

CNNwithBN

5 Conv Blocks + 1 FC (Conv, ReLU6, BN with max pooling in later layers). Very standard architecture these days, ReLU6 improves resilience when quantizing the network, BN improves training stability. In a second experiment separable convolutions are used as drop in replacement for regular Conv2d. This reduces the number of parameters by a factor of 8 and number of ops by 5, while reducing accuracy by 2%.

MobileNetV2

Architecture optimized for mobile devices and is commonly used as backend for image classification. While number of parameters is higher, the number of operations is greatly reduced compared to naive CNNwithBN Seperable. Comparing it to the CNN baseline it may still be overkill given the simplicity of this task (only 1 channel 28x28 images). https://arxiv.org/abs/1801.04381

MobileNetV3

Optimized version of MobileNetV2. While smaller and more compute efficient, accuracy seems to fall short significantly. https://arxiv.org/abs/1905.02244

Data Augmentation

We use horizontal flips as a regularization technique. Here we don't do a detailed breakdown on its effect on accuracy, but experiments showed no significant boost.

Results

Architecture	Accuracy	GFlops	MParams
CNN	0.899	0.00027	0.032998
CNNwithBN	0.934	0.17586	1.575114
CNNwithBN Separable	0.914	0.03603	0.187028
MobileNetV2	0.924	0.00561	2.236106
MobileNetV3	0.907	0.00346	1.66989