Use of brackets vs braces

[embeddetech]

Hello, I'm currently trying to slowly digest the brainscript syntax. Is it correct that the use of the bracket in the example below (beginning "BrainScriptNetworkBuilder = [") is deprecated, and the use of braces is recommended (i.e. "BrainScriptNetworkBuilder = {")? And the best explanation for this is that "BrainScriptNetworkBuilder = " is a record declaration as opposed to a function/macro declaration?

https://github.com/Microsoft/CNTK/wiki/Train-a-regression-model-on-images

[frankseide]

Yes, it is deprecated; yes, it is a record declaration.

However, many functions also create a record and either return the entire record, or select one element to return (this in a way of using local variables).

[embeddetech]

Thank you Frank! So from the example above, this is the definition of a function that returns a record:

model (features) = {...}

Then inside that, this line below calls the LinearLayer function with arguments in the braces, and that function returns a function pointer/lambda which is then immediately called with (featNorm) as its argument, and this result is assigned to h1?

h1 = LinearLayer {100, init="gaussian", initValueScale=1.5} (featNorm)

Think I'm slowly getting there... thanks for helping the slow kids in class...

[frankseide]

Typically, it is

model (features) = { a = ... ; b = ...;  z = ... }.z

which technically creates a record but then immediately pulls out z. You can think of a and b etc as variables in a local scope.

h1 = LinearLayer {100, init="gaussian", initValueScale=1.5} (featNorm)

Here, the curly braces mean something entirely different--function calls can use ( ) or { }, and by convention, functions that create model parameters use { }. The above is actually two function calls:

• LinearLayer{...} is a function that creates a function object with baked in learnable model parameters. This function object is returned as a callable function.
• h1 = ... (featNorm) then invokes that callable function immediately.

This separation was done so that it is always clear where new parameters are created. It also has additional benefits:

• You can create a layer or an entire model segment, and apply it multiple times. E.g. processing a query and a text passage with the same LSTM encoder, where all parameters are shared and jointly updated.
• You can use the Sequential() function to form neural networks as concatenations of layers.