medgeconv 1.0.1

No project description provided

An efficient tensorflow 2 implementation of the edge-convolution layer EdgeConv used in e.g. ParticleNet.

The structure of the layer is as described in ‘ParticleNet: Jet Tagging via Particle Clouds’ https://arxiv.org/abs/1902.08570. Graphs often have a varying number of nodes. By making use of the disjoint union of graphs in a batch, memory intensive operations in this implementation are done only on the actual nodes (and not the padded ones).

Instructions

Install via:

pip install medgeconv

Use e.g. like this:

import medgeconv

nodes = medgeconv.DisjointEdgeConvBlock(
    units=[64, 64, 64],
    next_neighbors=16,
    to_disjoint=True,
    pooling=True,
)((nodes, is_valid, coordinates))

Inputs to EdgeConv are 3 dense tensors: nodes, is_valid and coordinates

• nodes, shape (batchsize, n_nodes_max, n_features)

  Node features of the graph, padded to fixed size. Valid nodes have to come first, then the padded nodes.

• is_valid, shape (batchsize, n_nodes_max)

  1 for actual node, 0 for padded node.

• coordinates, shape (batchsize, n_nodes_max, n_coords)

  Features of each node used for calculating nearest neighbors.

Examples

Example for batchsize = 2, n_nodes_max = 4, n_features = 2:

nodes = np.array([
   [[2., 4.],
    [2., 6.],
    [0., 0.],  # <-- these nodes are padded, their
    [0., 0.]],  #           value doesn't matter

   [[0., 2.],
    [3., 7.],
    [4., 0.],
    [1., 2.]],
])

is_valid = np.array([
    [1, 1, 0, 0],  # <-- 0 defines these nodes as padded
    [1, 1, 1, 1],
])

coordinates = nodes

By using to_disjoint = True, the dense tensors get transformed to the disjoint union. The output is also disjoint, so this only needs to be done once. pooling = True will attach a node-wise global average pooling layer in the end, producing dense tensors again.

A full model could look like this:

import tensorflow as tf
import medgeconv

inp = (nodes, is_valid, coordinates)
x = medgeconv.DisjointEdgeConvBlock(
    units=[64, 64, 64],
    to_disjoint=True,
    batchnorm_for_nodes=True,
)(inp)

x = medgeconv.DisjointEdgeConvBlock(
    units=[128, 128, 128],
)(x)

x = medgeconv.DisjointEdgeConvBlock(
    units=[256, 256, 256],
    pooling=True,
)(x)

output = tf.keras.layers.Dense(2)(x)
model = tf.keras.Model(inp, output)

To load models, use the custom_objects:

import medgeconv

model = load_model(path, custom_objects=medgeconv.custom_objects)

knn_graph kernel

This package includes a cuda kernel for calculating the k nearest neighbors on a batch of graphs. It comes with a precompiled kernel for the version of tensorflow specified in requirements.txt.

To compile it locally, e.g. for a different version of tensorflow, go to medgeconv/tf_ops and run make clean and then make. This will produce the file medgeconv/tf_ops/python/ops/_knn_graph_ops.so. For details on how to setup the docker environment for compiling, see https://github.com/tensorflow/custom-op .