pna.py

from itertools import product
from typing import Optional, List

import torch
from torch import Tensor
import torch.nn.functional as F
from torch.nn import ModuleList, Linear, BatchNorm1d
from torch_sparse import SparseTensor
from torch_geometric.nn import MessagePassing

from torch_geometric_autoscale.models import ScalableGNN

EPS = 1e-5


class PNAConv(MessagePassing):
    def __init__(self, in_channels: int, out_channels: int,
                 aggregators: List[str], scalers: List[str], deg: Tensor,
                 **kwargs):
        super().__init__(aggr=None, **kwargs)

        self.in_channels = in_channels
        self.out_channels = out_channels
        self.aggregators = aggregators
        self.scalers = scalers

        deg = deg.to(torch.float)
        self.avg_deg = {
            'lin': deg.mean().item(),
            'log': (deg + 1).log().mean().item(),
        }

        self.pre_lins = torch.nn.ModuleList([
            Linear(in_channels, out_channels)
            for _ in range(len(aggregators) * len(scalers))
        ])
        self.post_lins = torch.nn.ModuleList([
            Linear(out_channels, out_channels)
            for _ in range(len(aggregators) * len(scalers))
        ])

        self.lin = Linear(in_channels, out_channels)

        self.reset_parameters()

    def reset_parameters(self):
        for lin in self.pre_lins:
            lin.reset_parameters()
        for lin in self.post_lins:
            lin.reset_parameters()
        self.lin.reset_parameters()

    def forward(self, x: Tensor, adj_t):
        out = self.propagate(adj_t, x=x)
        out += self.lin(x)[:out.size(0)]
        return out

    def message_and_aggregate(self, adj_t: SparseTensor, x: Tensor) -> Tensor:
        deg = adj_t.storage.rowcount().to(x.dtype).view(-1, 1)

        out = 0
        for (aggr, scaler), pre_lin, post_lin in zip(
                product(self.aggregators, self.scalers), self.pre_lins,
                self.post_lins):
            h = pre_lin(x).relu_()
            h = adj_t.matmul(h, reduce=aggr)
            h = post_lin(h)
            if scaler == 'amplification':
                h *= (deg + 1).log() / self.avg_deg['log']
            elif scaler == 'attenuation':
                h *= self.avg_deg['log'] / ((deg + 1).log() + EPS)

            out += h

        return out


class PNA(ScalableGNN):
    def __init__(self, num_nodes: int, in_channels: int, hidden_channels: int,
                 out_channels: int, num_layers: int, aggregators: List[int],
                 scalers: List[int], deg: Tensor, dropout: float = 0.0,
                 drop_input: bool = True, batch_norm: bool = False,
                 residual: bool = False, pool_size: Optional[int] = None,
                 buffer_size: Optional[int] = None, device=None):
        super().__init__(num_nodes, hidden_channels, num_layers, pool_size,
                         buffer_size, device)

        self.in_channels = in_channels
        self.out_channels = out_channels
        self.dropout = dropout
        self.drop_input = drop_input
        self.batch_norm = batch_norm
        self.residual = residual

        self.convs = ModuleList()
        for i in range(num_layers):
            in_dim = in_channels if i == 0 else hidden_channels
            out_dim = out_channels if i == num_layers - 1 else hidden_channels
            conv = PNAConv(in_dim, out_dim, aggregators=aggregators,
                           scalers=scalers, deg=deg)
            self.convs.append(conv)

        self.bns = ModuleList()
        for i in range(num_layers - 1):
            bn = BatchNorm1d(hidden_channels)
            self.bns.append(bn)

    @property
    def reg_modules(self):
        return ModuleList(list(self.convs[:-1]) + list(self.bns))

    @property
    def nonreg_modules(self):
        return self.convs[-1:]

    def reset_parameters(self):
        super().reset_parameters()
        for conv in self.convs:
            conv.reset_parameters()
        for bn in self.bns:
            bn.reset_parameters()

    def forward(self, x: Tensor, adj_t: SparseTensor, *args) -> Tensor:
        if self.drop_input:
            x = F.dropout(x, p=self.dropout, training=self.training)

        for conv, bn, hist in zip(self.convs[:-1], self.bns, self.histories):
            h = conv(x, adj_t)
            if self.batch_norm:
                h = bn(h)
            if self.residual and h.size(-1) == x.size(-1):
                h += x[:h.size(0)]
            x = h.relu_()
            x = self.push_and_pull(hist, x, *args)
            x = F.dropout(x, p=self.dropout, training=self.training)

        x = self.convs[-1](x, adj_t)
        return x

    @torch.no_grad()
    def forward_layer(self, layer, x, adj_t, state):
        if layer == 0 and self.drop_input:
            x = F.dropout(x, p=self.dropout, training=self.training)

        h = self.convs[layer](x, adj_t)
        if layer < self.num_layers - 1:
            if self.batch_norm:
                h = self.bns[layer](h)
            if self.residual and h.size(-1) == x.size(-1):
                h += x[:h.size(0)]
            h = h.relu_()
            h = F.dropout(h, p=self.dropout, training=self.training)

        return h