tf-lambda 0.3.0

A TensorFlow 2 implementation of LambdaNetworks.

Lambda Networks

This is a TensorFlow 2 implementation of a Lambda Layer from: LambdaNetworks: Modeling Long-Range Interactions Without Attention. LambdaNetworks achieve SOTA on ImageNet. For now, this library provides a basic UNTESTED version of the Lambda Layer based off of this repository.

Usage

Global Context

import tensorflow as tf

layer = LambdaLayer(
    dim_out = 32,   # channels out
    n = 64 * 64,    # number of input pixels (64 x 64 image)
    dim_k = 16,     # key dimension
    heads = 4,      # number of heads, for multi-query
    dim_u = 1       # 'intra-depth' dimension
)

x = tf.random.normal(shape=(1, 64, 64, 32))
layer(x)

Local Context

layer = LambdaLayer(
    dim_out = 32,
    r = 23,         # the receptive field for relative positional encoding (23 x 23)
    dim_k = 16,
    heads = 4,
    dim_u = 4
)

x = tf.random.normal(shape=(1, 64, 64, 32))
layer(x)