Understanding shapes

[hmf]

I am looking at chapter 5.2 of the online book. We have the following code:

NDManager manager = NDManager.newBaseManager();

NDArray x = manager.randomUniform(0, 1, new Shape(2, 4));

Model model = Model.newInstance("lin-reg");

SequentialBlock net = new SequentialBlock();

net.add(Linear.builder().setUnits(8).build());
net.add(Activation.reluBlock());
net.add(Linear.builder().setUnits(1).build());
net.setInitializer(new NormalInitializer());
net.initialize(manager, DataType.FLOAT32, x.getShape());

model.setBlock(net);

Predictor<NDList, NDList> predictor = model.newPredictor(new NoopTranslator());

predictor.predict(new NDList(x)).singletonOrThrow(); // forward computation

I am assuming that given the NoopTranslator has no Batchifier the input may consist of one or more samples. The input is 8 and the output 1, so for an input of Shape(1,8) I should get a single value.

However, looking at the example the input is Shape(2,4). I don't see any Blocks.batchFlattenBlock(8). In addition to this the output is Shape(2,1). What transformations are occurring here?

Apologies if this is simpleton's question.

TIA

[zachgk]

The input is not specified when creating the linear blocks in net. Because inputs can be computed, we thought it would be too much work to have to type them for every block. So, the 8 you are looking at is the output of the first linear block.

The input is not determined until the call to net.initialize. At this point, it determines the input is of size 4 (with batch size 2) using x.getShape()

[hmf]

Silly me. This was "explained" in chapter 5.1. Thank you.