This repo contains the implementation of Denoising Autoencoder (DAE) using TensorFlow and a series of experiments. It starts with blackecho's gist.
- Python 3
- TensorFlow 1.0
- NumPy
- Pickle
- Improve original code.
- Update to Python 3 and TensorFlow 1.0.
- Modify the default values of hyper-parameters.
- Export Origin / Corrupted / Decoded image.
- Miscellaneous (see commits).
- Add Gaussian noise to the corruption module.
- Add ReLU as one activation function, in addition to sigmoid and tanh.
- Compare the performance between different activation functions.
- Compare the performance between cross entropy and squared loss.
- Compare different learning approaches.
- Feed the output from the hidden layer into SVM. Feed the feture map from LeNet-5 into SVM. Then compare their classification performance. (This gist helps to extract features from LeNet-5)
- Use the weight trained from DAE to initialize LeNet-5.
This project is licensed under the MIT License - see the LICENSE file for details