... | ... | @@ -33,26 +33,25 @@ TODO: |
|
|
- ggf. mit Handbüchern anderer Programme vergleichen und anpassen
|
|
|
- ggf. Video (Ersatz für Live-Demo aus der Präsentation)
|
|
|
|
|
|
## Introduction / Abstract: Alex S
|
|
|
## Introduction / Abstract: Alex S **WIP**, but ideas, annotations and additions welcome
|
|
|
**Barista** is a graphical user interface for **caffe**, a deep neural network framework. http://caffe.berkeleyvision.org/
|
|
|
|
|
|
Its intention is to faciliate the creation and training of neural networks.
|
|
|
Its intention is to facilitate the creation and training of neural networks.
|
|
|
|
|
|
To achieve this, the network is visualized as a graph and can be convientiently altered.
|
|
|
To achieve this, the network is visualized as a graph and can be conveniently altered.
|
|
|
|
|
|
The network definition of caffe relies on the prototxt format, which defines a file where the network defining parameter are written in plain text. Due to manifold parameters and the possibility to create extensive networks, these files tend to grow very large. The user has to manually write and alter these files to manipulate the network, an error-prone and confusing process.
|
|
|
The network definition of **caffe** relies on the **prototxt** format, which defines a file where the network defining parameter are written in plain text. Due to manifold parameters and the possibility to create extensive networks, these files tend to grow very large. The user has to manually write and alter these files to manipulate the network, an error-prone and confusing process.
|
|
|
|
|
|
A solution to that problem is to show the interconnectivity of the layers in the network as a graph, which can be directly manipulated by the user. Thes nodes are representing the layers, which inner parameters can now be accessed seperately. The position of the layers in the network can be altered by drawing connection with the mouse, new layers can be added per drag-and-drop.
|
|
|
A solution to that problem is to show the interconnectivity of the layers in the network as a graph, which can be directly manipulated by the user. The nodes are representing the layers, which inner parameters can now be accessed separately. The position of the layers in the network can be altered by drawing connection with the mouse, new layers can be added per drag-and-drop.
|
|
|
|
|
|
Parameters of the layers and the solver can be convieniently added, removed and altered. **Barista** reduces the possibility of errors by checking the network for consistency and only allowing consistent parameters to be added.
|
|
|
Parameters of the layers and the solver can be conveniently added, removed and edited. **Barista** reduces the possibility of errors by checking the network for consistency and only allowing consistent parameters to be added.
|
|
|
|
|
|
|
|
|
Furthermore, the caffe training process can be controlled solely over **Barista**, with live plotting of the accumulated training data and storing of previous training sessions.
|
|
|
Furthermore, the **caffe** training process can be controlled solely over **Barista**, with live plotting of the accumulated training data and storing of previous training sessions.
|
|
|
|
|
|
Additionally, **Barista** provides further functionality, like a prototxt-editor with syntax highlighting, the specification of training- and testdata in HDF5-format, the visualization of trained convolutional layers and more.
|
|
|
|
|
|
To failiate the access to **Barista**, existing caffe projects can be added to by importing existing prototxt files.
|
|
|
To facilitate the access to **Barista**, existing caffe projects can be added to by importing existing prototxt files.
|
|
|
|
|
|
Since caffe officially only supports Linux and MacOS, **Barista** is targeted mostly for Linux.
|
|
|
Since **caffe** officially only supports Linux and MacOS, **Barista** is targeted mostly for Linux.
|
|
|
|
|
|
|