deepforge-dev / deepforge Goto Github PK

We should have a fully functioning 60 min blitz

When the basic language is imported using the plugin, the meta nodes should be positioned reasonably so they are more legible

Unfortunately, the zip file format is transitioning out (of webgme v2.x) and the library support isn't quite right in the imported zip files.

Operations should be viewable using a text editor (codemirror)

Users should be able to select the option to automatically wrap the model in a Java class for use in a production environment.

The pipelines could use more tests. This should include mocking the executors and testing more complex pipelines

the GenerateArchitecture plugin needs to be extended to support concat layers

When executing a pipeline, the pipeline should be duplicated and shown in an "execution view"

DeepForge should have a metamodel defining the Torch visual language

When I run CreateTorchMeta, it generates a read-only meta because the nodes are not given positions on the meta sheet

I need to add automated tests w/ travis ci

Users should be able to generate the Torch architecture code for download

ForgeActionButton should be secondary viz

Custom layers should be able to be tested quickly with some stock sample data. This could be happening while the user is creating the layer to provide quick feedback.

Users should be able to test trained models through the web UI with varied data sets

We need more tests for architecture generation. This includes tests for:

• concat layers
• parallel layers
• split tables

Layers should be colored by their base's base type

Users should be able to download the Torch trained models

Users should be able to fine tune trained models (creating new trained models).

When viewing the trained model, users should be able to view the architecture used.

DeepForge should have a plugin for automatically creating the "boilerplate" metamodel from the layers in the nn library.

There should be a demo project w/ example pipelines and architectures

Currently, data in libraries can't be used w/o subclassing it since the data types have to have src and dst references to their source/destination operation types. This is a problem for two reasons (which are debatably the same reason):

• Relationship is stored on both the operation and the data type
• Tightly coupled data and operations

The generated architecture code should be modular in the sense that it should not define globals and should return the network. This should prevent collisions with other lua code

When training and testing, the custom layer definitions should be packaged as modules and shipped to the workers performing the training/testing

Users should be able to use machine learning pipelines to make predictions on input data

Users should be able to train Torch models and interact with the resulting model through the web UI

The dimensionality information should be moved to the metamodel from src/common/dimensionality. The dimensionality information should be stored in the layers as a setting (maybe storing lua code that calculates the dimensionality from the previous dimensions and the attributes of the given layer.

Additional information from the layers.yml file should be moved to the meta. Most significantly, this will entail more complex attribute schemas. It will also need to replace layers.yml in the layer-args.js file

Ace isn't always loading in time (requirejs "thinks" it returns undefined)

MNIST Example using Twitter Dataset

Example/Test for Twitter's Torch-Dataset package:
MNIST Example

Users should be able to view a visualization of layer weights when inspecting layers of trained models

DeepForge should be able to import existing Torch projects into the web ui

Users should be able to create ensembles of models in which the entire workflow is managed by DeepForge

Users should be able to test and score machine learning pipelines

DeepForge should have a simplified user interface using:

• webgme-chflayout
• webgme-fab
• webgme-easydag
• webgme-autoviz

data types should be able to have attributes such as dimensionality

Users should be able to deploy their models to a REST API in one (maybe 2 if you consider the settings) clicks.

Free alphanumeric dataset from CMU.

http://www.speech.cs.cmu.edu/databases/an4/index.html

You'll want to convert to .wav files. You can use Sox (http://sox.sourceforge.net/) for this. Sox should handle the .sph files.

The graphical architecture editor should extend EasyDAG and have an automatic layout, etc

This is due to a bug in one of the dependencies: tdzl2003/lua.js#12

DeepForge should have an lua code editor for defining custom layers. This should have lua syntax highlighting

Add support for random sampling for hyper-parameters during training.