Experiments with Fashion-MNIST
Dependencies
python 3.6.5, argparse 1.1, numpy 1.14, tensorflow 1.10, Pillow 5.2.0
Or simply run pip3 install -r requirements.txt
Running the script
Training from scratch
Run python3 model.py --model-dir=<MODEL_DIR>. This will train the model from scratch if <MODEL_DIR> is empty.
Model checkpoints and logs will be created in this folder
Resuming training
Run python3 model.py --model-dir=<MODEL_DIR>. This will continue training from the latest checkpoint in <MODEL_DIR>
Running inference on an image
Run python3 model.py --model-dir=<MODEL_DIR> --input-img=<IMAGE_FILE_PATH>. Pretrained checkpoint is given in pretrained directory.
E.g
python3 model.py --input-img=test_imgs/2.png
would output
Predicted class: Pullover
Approach taken
Model architecture
| Layer | Output Shape |
|---|---|
| Input layer | (BATCH_SIZE, 28, 28, 1) |
| conv1 (with elu) | (BATCH_SIZE, 28, 28, 64) |
| pool1 | (BATCH_SIZE, 14, 14, 64) |
| conv2 (with elu) | (BATCH_SIZE, 14, 14, 64) |
| pool2 | (BATCH_SIZE, 7, 7, 64) |
| pool2_flat | (BATCH_SIZE, 3136) |
| dense | (BATCH_SIZE, 1024) |
| logits | (BATCH_SIZE, 10) |
Training
Model is trained with batch size of 250 over 4 epochs using Adam optimizer, with default hyper-parameters (learning_rate=0.001, beta1=0.9, beta2=0.999, epsilon=1e-08). Softmax cross entropy is used as the loss function.
Logging
Training accuracy, loss & precision are logged using the standard interval. Evaluation accuracy, loss & precision are logged every 40 training steps.
All the metrics can be inspected by accessing Tensorboard using tensorboard --logdir=<MODEL_DIR>
Analysis
Training loss vs Evaluation loss over time
Blue line indicates validation loss, while orange line indicate training loss. The final evaluation loss after training is 0.28, while final training loss is 0.17.
Training accuracy vs Evaluation accuracy over time
Training accuracy over time:
Evaluation accuracy over time:
The final training accuracy is 0.93, versus evaluation accuracy of 0.90. The accuracies appear close enough so there is few sign of overfitting.