Coding exercise with FashionMNIST

See how basic image processing affect inferencing quality.

Training

The model was constructed with MobileNet and ImageNet pretrained weights. You can train the model with:

python train.py

Inferencing

The prediction accuracy on the original testing data was 93.67%. After some basic image processing (complementary value, Gaussian blur) on the testing images, the accuracy decreases.

Model checkpoint

The model checkpoint can be obtained in the above link. Please put the files in /weight folders, then you can run the experiment with:

python infer.py

It can be seen that both process have a huge influence on the CNN.

Accuracy vs percentage of the pixels converted to complementary value

Accuracy vs std of the Gaussian 3x3 kernel