rogalag / vlml

Machine learning in a visual language

VLML

Machine learning in the visual language VL, part of vvvv Gamma and vvvv Beta

---

VLML is a machine learning framework that aims to take some of the hassle out getting your hands dirty with machine learning, making it a relatively easy to get started. VLML uses CNTK as its back end, one of the most performant machine learning libraries out there. There is a short guide here on how to install VLML.

• For now I recommend installing and testing VLML with vvvv alpha f7d5bf1879 because later versions currently have some compile errors, I'll update this when they stabilize.
• There are now some demos available to download from here that go through some basic usecases for machine learning. A git repository can be found here.
• Finally there is also a toy to play with: pix2pix face painting!. This is an evaluator of a pix2pix (ConditionalGan) model was trained on photographs and segmentation data. The neural network will be released soon so you can train your own. See more about this subject here.

---

Setting up a basic network

Create and two flattened collections of data.

The first will be X (the kind of data you want to process) and second will be Y (the kind of data you expect the process to result in).

DataInput: Create DataIn nodes for each of your datasets

The sample size refers to the number of elements in each sample of your dataset. In this example X has 3 elements, Y has 1.

Design your network

Put in some layers. Here we have just one and without an activation function.

Pass your network and target data into a loss function.

There several to choose from that are best suited to different ML tasks. You can also write your own.

---

Set up the trainer

Pass your network and loss function into a trainer:

• Set your Minibatch Size This the size of a chunk of you data set to train consecutively each step. In this case it is set to zero so the whole batch of data (every sample of the data set) is used in each training step.

• Set the learning rate for your trainer. This defines the size of the learning steps the trainer can adjust the network with.

• Press train to run the training session

---

Evaluate your model

Method 1

• Create an DataIn node with a sample size that matches the one used for the network's 'x' input (3). This will have test data.
• Make an Evaluate(Debuggable) node Pass in the Network Pass in the test DataIn
• Use a DataOut node to output the result in a format of your needs, in this case float.

Method 2

• Use 'SetInput' node, taking your model and a DataIn node with a size matching the input of the your model.
• pass this into Evaluate node
• Use a DataOut node to output the result in a format of your needs.

The main difference between these two approaches is that the Debuggable Evaluator can use the latest model while it updates. The 'SetInput' node only updates the model when a new model is passed into in to it. It is best used with finished models.