Litematica Schematic Generation
Table of contents
General representation standard
nn-to-litematica module converts the binary weights from the pytorch checkpoint into litematica schematics. Every litematica is arranged so that the first 2 dimensions lie in (x, z) plane. Moreover, the loading point of schematic (the block where player is standing when loading schematic) always represents the very first number in weights (i.e. index [0, 0, 0] in weights array). The first dimension of weight array is from loading point towards positive X axis. The second dimension of weight array is from loading point towards positive Z axis. Any dimensions more than 3 are reshaped into third axis (Y axis).
Representation of conv layers
We follow the general representation rules as stated above. Here, since conv layers are stored as a 4 dimensional array, we need to squeeze extra dimension to make it 3 dimensional. Typical shape of conv layer is \([c_2, c_1, f_h, f_w]\). Here,
- \(c_2\) = Number of channels output activations should have
- \(c_1\) = Number of channels input activations should have
- \((f_h, f_w)\) = Height and width of convolution filter
To convert this into 3 dimensional representation, scarpet-nn squeezes \((f_h, f_w)\) into a single dimension of size \(f_h \times f_w\). Therefore, the new shape of conv layer for schematic generation would be \([c_2, c_1, f_h \times f_w]\).
To understand this visually, please take a look at following image. It displays the block representation of conv3 layer weights from the demo world. The shape of conv3 layer weights is [16, 8, 3, 3] and is then reshaped into [16, 8, 9]
Representation of fc layers
fc layers are also inline to general representation standard defined by scarpet-nn. Typical shape of fc layer is \([n, k]\). fc layer takes in input of shape \([m, n]\) and produces output of shape \([m, k]\). Since all the operations are 2 dimensional, we don’t need any dimension adjustments here. Since there is no third dimensions, the whole weight array would just lie in X-Z plane without ever going into Y direction.
From the demo world, fc1 layer is visualized below. The shape of this layer is [8, 16]
