Image Classification on FashionMNIST dataset CNN based model for image classification on Fashion-MNIST dataset. Best test accuracy - 94.55 percent. Regularization methods like Dropout, L2 regularization, Data Augmentation and Batch Normalization have been explored. This code supports python3, keras and tensorflow-v1 and runs on GPU and CPU machines. Please note that checkpoint of model trained on GPU cannot be used for inference on CPU.
For SmallNet model training,
python3 main.py --use_data_aug True --use_dropout True --dropout_p 0.2 --lr 0.001 --l2_lambda 0.0 --data_aug random_erasing
For BigNet model training,
python3 main.py --use_data_aug True --use_dropout True --dropout_p 0.2 --lr 0.001 --l2_lambda 0.0 --model BigNet --use_bn True --batch_size 96
Trained checkpoints can be evaluated by adding : --ckpt checkpoint_path --inference_only True