PieNet is a Convolutional Neural Network (CNN) model built with TensorFlow for classifying digits from the popular MNIST dataset. This CNN architecutre is based off of the LeNet-5 architecture developed by Yann LeCun (see http://yann.lecun.com/).
- python3
- Tensorflow
- Numpy
| Layer | Num_Maps | Size | Kernel Size | Stride | Activation |
|---|---|---|---|---|---|
| Input | 1 | 32x32 | - | - | - |
| Convolution 1 | 6 | 28x28 | 5x5 | 1 | ReLu |
| Max Pooling | 6 | 14x14 | 2x2 | 2 | - |
| Convolution 2 | 16 | 10x10 | 5x5 | 1 | ReLu |
| Max Pooling | 16 | 5x5 | 2x2 | 2 | - |
| Convolution 3 | 120 | 1x1 | 5x5 | 1 | ReLu |
| Fully Conn. | - | 84 | - | - | ReLu |
| Fully Conn. | - | 10 | - | - | Softmax |
Although I have included a pretrained model, it is simple to perform accuracte training in a short amound of time (<1hr) on a standard CPU. Currently, the default training uses TensorFlow's MNIST example dataset which contains 55,000 training images of digits. With Mini-batch training, with batchs sizes of 100 and using a Nesterov Acclerated Gradient Optimizer, the network achieved ~97% accuracy on the validation set over only 40 epochs. Here is how to perform the training.(Note: You can change the batch size and epochs in main.py).
python main.py train
The Accuracy data and Cross Entropy Cost Function data will automatically be saved to be view through TensorBoard. The default location of the tf_logs directory is [src code directory]/tf_logs. To view the graph architecture and data graphs on TensorBoard, run:
tensorboard --logdir /tf_logs
The model parameters and graph structure will also be saved and able to be restored for both inferencing and/or re-training. By default the final model save is located in directory C:\Machine_Learning\ML Projects\MNIST_CNN_PieNet\PieNet_final_model.ckpt. This can be changed by editing the MNISTPieNet class attribute self.default_model_path. Note: Documentation in the code is provided to easily save and restore models.
After having a pretrained model(I include an already trained model), it is easily to do a simpe inference on a number of MNIST images provide in the Tensorflow example set. Simply run:
python main.py inference [number_of_images_to_inference]
Example: python main.py inference 100. This inferences 100 MNIST images.