A cuda implementation of a convolutional autoencoder which can operate in both standard coordinate space and momentum space.
Training of the network by backpropagation can also be performed in both standard coordinate space and momentum space (and due to the simplicity of convolution in momentum space, it turns out to be much faster when operating in such configuration).
The input is configured to be the real time feed from the camera (accessed through opencv), but can be modified to be a image or video file by tweaking the file autoencoder.cpp.
Output is the recostructed image produced by the autoencoder (which once trained should reproduce the input). Also shown are the learned kernels for each layer and the corresponding feature maps that they produce (only available in coordinate space).
Convolutional layers can be added/removed during execution and weights saved/loaded.
Various configuration parameters can also be adjusted in real time:
- learning rate
- use of active learning rate
- use of symmetric weights shared between encoder and decoder
- use of multiobjective optimization to maximaze the difference between the learned kernels
The program requires opencv and cuda toolkit to run.
On Ubuntu 22.04, the script install_dependencies.sh can be used to install opencv4 and cuda/cuda toolkit 11.7
To compile run make.
Note that it could be necessary to modify the Makefile to point to the correct libraries paths.
To run the program run the command ./AutoEnc
Following is the list of commands / options:
| Command | Action |
|---|---|
| 1 | Start/Stop backpropagation |
| 2 | Increase the scale of the backpropagation input |
| 3 | Decrease the scale of the backpropagation input |
| 4 | Increase the learning rate |
| 5 | Decrease the learning rate |
| 6 | Increase inertia value (Not av in FFT) |
| 7 | Decrease inertia value (Not av in FFT) |
| 9 | Switch active learning rate on and off (Not av in FFT) |
| 0 | Switch gpu processing on and off (0=off, 1=on) |
| f | Switch FFT processing on and off (i.e. convolution and backpropagation in momentum space) |
| g | Perform inverse FFT on each layer (useful to visualize the feature maps in momentum space) |
| q | itearate (forward) throught the feature maps of each layer |
| w | itearate (backward) throught the feature maps of each layer |
| m | Switch multiobjective optimization on and off |
| z | itearate (forward) throught the network layers |
| x | itearate (backward) throught the network layers |
| e | Initialize random weights on the selected layer |
| c | Clean FFT convolutional kernels vector |
| p | Activate/deactivate symmetric weights |
| s | Save convolutional weights |
| l | Load convolutional weights |
| n | Add new convolutional layer (the parameters of the layer are set in the file New_Layer_Param.txt) |
| d | Delete last/most internal layer |
| i | Print network structure |
| Esc | Quit |