neurlink 0.1.0

a compact neuron network model definition grammar

NEURLINK

A compact grammar for neural network definition based on PyTorch.

Basics

• Nodes:
  • (c, s, o): tensor (channel_size, downsample_scale, operation),
  • ((c1, s1), (c2, s2), ..., o) or ((c, s, o),)*n: parallel tensors,
  • [(c, s, o)]*n: sequential tensors,
• Ordered nodes form a graph. Each node has an operation (also known as transform function) that converts one or more previous tensors to (a) new tensor(s). Defining new operations requires some specific arguments and a decorator.

Example (resnet)

def resnet50(num_classes: int = 1000, **block_keywords):
    block = BottleNeck(**block_keywords)
    expansion = 4
    return build(
        [
            (3, 1, None),
            (64, 2, ConvLayer(7)),
            (64, 4, MaxPool2d(3)),
            [(64  * expansion, 4, block)] * 3,
            [(128 * expansion, 8, block)] * 4,
            [(256 * expansion, 16, block)] * 6,
            [(512 * expansion, 32, block)] * 3,
            (512 * expansion, "(1, 1)", AvgPool2d()),
            (num_classes, None, ConvLayer(1, norm=nn.Identity, act=nn.Identity)),
        ]
    )