[Model][MXNet] neighbor sampling & skip connection & control variate …

…& graphsage (dmlc#322)

* neighbor sampling draft

* val/test acc

* control variate draft

* control variate

* update

* fix new_history

* maintain aggregated history while updating new history

* preprocess the first layer, change push to pull

* update

* fix subg_degree

* nodeflow

* clear

* readme

* doc and unittest for self loop

* address comments

* rename

* update

* fix

* Update node_flow.py

* Update node_flow.py

examples/mxnet/sampling/README.md

@@ -0,0 +1,66 @@
+# Stochastic Training for Graph Convolutional Networks
+
+* Paper: [Control Variate](https://arxiv.org/abs/1710.10568)
+* Paper: [Skip Connection](https://arxiv.org/abs/1809.05343)
+* Author's code: [https://github.com/thu-ml/stochastic_gcn](https://github.com/thu-ml/stochastic_gcn)
+
+### Dependencies
+
+- MXNet nightly build
+
+```bash
+pip install mxnet --pre
+```
+
+### Neighbor Sampling & Skip Connection
+cora: test accuracy ~83% with `--num-neighbors 2`, ~84% by training on the full graph
+```
+DGLBACKEND=mxnet python gcn_ns_sc.py --dataset cora --self-loop --num-neighbors 2 --batch-size 1000000 --test-batch-size 1000000 --gpu 0
+```
+
+citeseer: test accuracy ~69% with `--num-neighbors 2`, ~70% by training on the full graph
+```
+DGLBACKEND=mxnet python gcn_ns_sc.py --dataset citeseer --self-loop --num-neighbors 2 --batch-size 1000000 --test-batch-size 1000000 --gpu 0
+```
+
+pubmed: test accuracy ~76% with `--num-neighbors 3`, ~77% by training on the full graph
+```
+DGLBACKEND=mxnet python gcn_ns_sc.py --dataset pubmed --self-loop --num-neighbors 3 --batch-size 1000000 --test-batch-size 1000000 --gpu 0
+```
+
+reddit: test accuracy ~91% with `--num-neighbors 2` and `--batch-size 1000`, ~93% by training on the full graph
+```
+DGLBACKEND=mxnet python gcn_ns_sc.py --dataset reddit-self-loop --num-neighbors 2 --batch-size 1000 --test-batch-size 500 --n-hidden 64
+```
+
+
+### Control Variate & Skip Connection
+cora: test accuracy ~84% with `--num-neighbors 1`, ~84% by training on the full graph
+```
+DGLBACKEND=mxnet python gcn_cv_sc.py --dataset cora --self-loop --num-neighbors 1 --batch-size 1000000 --test-batch-size 1000000 --gpu 0
+```
+
+citeseer: test accuracy ~69% with `--num-neighbors 1`, ~70% by training on the full graph
+```
+DGLBACKEND=mxnet python gcn_cv_sc.py --dataset citeseer --self-loop --num-neighbors 1 --batch-size 1000000 --test-batch-size 1000000 --gpu 0
+```
+
+pubmed: test accuracy ~77% with `--num-neighbors 1`, ~77% by training on the full graph
+```
+DGLBACKEND=mxnet python gcn_cv_sc.py --dataset pubmed --self-loop --num-neighbors 1 --batch-size 1000000 --test-batch-size 1000000 --gpu 0
+```
+
+reddit: test accuracy ~93% with `--num-neighbors 1` and `--batch-size 1000`, ~93% by training on the full graph
+```
+DGLBACKEND=mxnet python gcn_cv_sc.py --dataset reddit-self-loop --num-neighbors 1 --batch-size 1000 --test-batch-size 500 --n-hidden 64
+```
+
+### Control Variate & GraphSAGE-mean
+
+Following [Control Variate](https://arxiv.org/abs/1710.10568), we use the mean pooling architecture GraphSAGE-mean, two linear layers and layer normalization per graph convolution layer.
+
+reddit: test accuracy 96.1% with `--num-neighbors 1` and `--batch-size 1000`, ~96.2% in [Control Variate](https://arxiv.org/abs/1710.10568) with `--num-neighbors 2` and `--batch-size 1000`
+```
+DGLBACKEND=mxnet python graphsage_cv.py --batch-size 1000 --test-batch-size 500 --n-epochs 50 --dataset reddit --num-neighbors 1 --n-hidden 128 --dropout 0.2 --weight-decay 0
+```
+

examples/mxnet/sampling/gcn_cv_sc.py

@@ -0,0 +1,321 @@
+import argparse, time, math
+import numpy as np
+import mxnet as mx
+from mxnet import gluon
+import dgl
+import dgl.function as fn
+from dgl import DGLGraph
+from dgl.data import register_data_args, load_data
+
+
+class NodeUpdate(gluon.Block):
+    def __init__(self, layer_id, in_feats, out_feats, dropout, activation=None, test=False, concat=False):
+        super(NodeUpdate, self).__init__()
+        self.layer_id = layer_id
+        self.dropout = dropout
+        self.test = test
+        self.concat = concat
+        with self.name_scope():
+            self.dense = gluon.nn.Dense(out_feats, in_units=in_feats)
+            self.activation = activation
+
+    def forward(self, node):
+        h = node.data['h']
+        if self.test:
+            norm = node.data['norm']
+            h = h * norm
+        else:
+            agg_history_str = 'agg_h_{}'.format(self.layer_id-1)
+            agg_history = node.data[agg_history_str]
+            # control variate
+            h = h + agg_history
+            if self.dropout:
+                h = mx.nd.Dropout(h, p=self.dropout)
+        h = self.dense(h)
+        if self.concat:
+            h = mx.nd.concat(h, self.activation(h))
+        elif self.activation:
+            h = self.activation(h)
+        return {'activation': h}
+
+
+
+class GCNSampling(gluon.Block):
+    def __init__(self,
+                 in_feats,
+                 n_hidden,
+                 n_classes,
+                 n_layers,
+                 activation,
+                 dropout,
+                 **kwargs):
+        super(GCNSampling, self).__init__(**kwargs)
+        self.dropout = dropout
+        self.n_layers = n_layers
+        with self.name_scope():
+            self.layers = gluon.nn.Sequential()
+            # input layer
+            self.dense = gluon.nn.Dense(n_hidden, in_units=in_feats)
+            self.activation = activation
+            # hidden layers
+            for i in range(1, n_layers):
+                skip_start = (i == self.n_layers-1)
+                self.layers.add(NodeUpdate(i, n_hidden, n_hidden, dropout, activation, concat=skip_start))
+            # output layer
+            self.layers.add(NodeUpdate(n_layers, 2*n_hidden, n_classes, dropout))
+
+    def forward(self, nf):
+        h = nf.layers[0].data['preprocess']
+        if self.dropout:
+            h = mx.nd.Dropout(h, p=self.dropout)
+        h = self.dense(h)
+
+        skip_start = (0 == self.n_layers-1)
+        if skip_start:
+            h = mx.nd.concat(h, self.activation(h))
+        else:
+            h = self.activation(h)
+
+        for i, layer in enumerate(self.layers):
+            new_history = h.copy().detach()
+            history_str = 'h_{}'.format(i)
+            history = nf.layers[i].data[history_str]
+            h = h - history
+
+            nf.layers[i].data['h'] = h
+            nf.block_compute(i,
+                             fn.copy_src(src='h', out='m'),
+                             lambda node : {'h': node.mailbox['m'].mean(axis=1)},
+                             layer)
+            h = nf.layers[i+1].data.pop('activation')
+            # update history
+            if i < nf.num_layers-1:
+                nf.layers[i].data[history_str] = new_history
+
+        return h
+
+
+class GCNInfer(gluon.Block):
+    def __init__(self,
+                 in_feats,
+                 n_hidden,
+                 n_classes,
+                 n_layers,
+                 activation,
+                 **kwargs):
+        super(GCNInfer, self).__init__(**kwargs)
+        self.n_layers = n_layers
+        with self.name_scope():
+            self.layers = gluon.nn.Sequential()
+            # input layer
+            self.dense = gluon.nn.Dense(n_hidden, in_units=in_feats)
+            self.activation = activation
+            # hidden layers
+            for i in range(1, n_layers):
+                skip_start = (i == self.n_layers-1)
+                self.layers.add(NodeUpdate(i, n_hidden, n_hidden, 0, activation, True, concat=skip_start))
+            # output layer
+            self.layers.add(NodeUpdate(n_layers, 2*n_hidden, n_classes, 0, None, True))
+
+
+    def forward(self, nf):
+        h = nf.layers[0].data['preprocess']
+        h = self.dense(h)
+
+        skip_start = (0 == self.n_layers-1)
+        if skip_start:
+            h = mx.nd.concat(h, self.activation(h))
+        else:
+            h = self.activation(h)
+
+        for i, layer in enumerate(self.layers):
+            nf.layers[i].data['h'] = h
+            nf.block_compute(i,
+                             fn.copy_src(src='h', out='m'),
+                             fn.sum(msg='m', out='h'),
+                             layer)
+            h = nf.layers[i+1].data.pop('activation')
+
+        return h
+
+
+def main(args):
+    # load and preprocess dataset
+    data = load_data(args)
+
+    if args.gpu >= 0:
+        ctx = mx.gpu(args.gpu)
+    else:
+        ctx = mx.cpu()
+
+    if args.self_loop and not args.dataset.startswith('reddit'):
+        data.graph.add_edges_from([(i,i) for i in range(len(data.graph))])
+
+    train_nid = mx.nd.array(np.nonzero(data.train_mask)[0]).astype(np.int64)
+    test_nid = mx.nd.array(np.nonzero(data.test_mask)[0]).astype(np.int64)
+
+    num_neighbors = args.num_neighbors
+    n_layers = args.n_layers
+
+    features = mx.nd.array(data.features).as_in_context(ctx)
+    labels = mx.nd.array(data.labels).as_in_context(ctx)
+    train_mask = mx.nd.array(data.train_mask).as_in_context(ctx)
+    val_mask = mx.nd.array(data.val_mask).as_in_context(ctx)
+    test_mask = mx.nd.array(data.test_mask).as_in_context(ctx)
+    in_feats = features.shape[1]
+    n_classes = data.num_labels
+    n_edges = data.graph.number_of_edges()
+
+    n_train_samples = train_mask.sum().asscalar()
+    n_test_samples = test_mask.sum().asscalar()
+    n_val_samples = val_mask.sum().asscalar()
+
+    print("""----Data statistics------'
+      #Edges %d
+      #Classes %d
+      #Train samples %d
+      #Val samples %d
+      #Test samples %d""" %
+          (n_edges, n_classes,
+              n_train_samples,
+              n_val_samples,
+              n_test_samples))
+
+    # create GCN model
+    g = DGLGraph(data.graph, readonly=True)
+
+    g.ndata['features'] = features
+
+    norm = mx.nd.expand_dims(1./g.in_degrees().astype('float32'), 1)
+    g.ndata['norm'] = norm.as_in_context(ctx)
+
+    g.update_all(fn.copy_src(src='features', out='m'),
+                 fn.sum(msg='m', out='preprocess'),
+                 lambda node : {'preprocess': node.data['preprocess'] * node.data['norm']})
+
+    for i in range(n_layers):
+        g.ndata['h_{}'.format(i)] = mx.nd.zeros((features.shape[0], args.n_hidden), ctx=ctx)
+
+    g.ndata['h_{}'.format(n_layers-1)] = mx.nd.zeros((features.shape[0], 2*args.n_hidden), ctx=ctx)
+
+    model = GCNSampling(in_feats,
+                        args.n_hidden,
+                        n_classes,
+                        n_layers,
+                        mx.nd.relu,
+                        args.dropout,
+                        prefix='GCN')
+
+    model.initialize(ctx=ctx)
+
+    loss_fcn = gluon.loss.SoftmaxCELoss()
+
+    infer_model = GCNInfer(in_feats,
+                           args.n_hidden,
+                           n_classes,
+                           n_layers,
+                           mx.nd.relu,
+                           prefix='GCN')
+
+    infer_model.initialize(ctx=ctx)
+
+    # use optimizer
+    print(model.collect_params())
+    trainer = gluon.Trainer(model.collect_params(), 'adam',
+                            {'learning_rate': args.lr, 'wd': args.weight_decay},
+                            kvstore=mx.kv.create('local'))
+
+    # initialize graph
+    dur = []
+    for epoch in range(args.n_epochs):
+        for nf, aux in dgl.contrib.sampling.NeighborSampler(g, args.batch_size,
+                                                            num_neighbors,
+                                                            neighbor_type='in',
+                                                            shuffle=True,
+                                                            num_hops=n_layers,
+                                                            seed_nodes=train_nid):
+            for i in range(n_layers):
+                agg_history_str = 'agg_h_{}'.format(i)
+                g.pull(nf.layer_parent_nid(i+1), fn.copy_src(src='h_{}'.format(i), out='m'),
+                       fn.sum(msg='m', out=agg_history_str),
+                       lambda node : {agg_history_str: node.data[agg_history_str] * node.data['norm']})
+
+            node_embed_names = [['preprocess', 'h_0']]
+            for i in range(1, n_layers):
+                node_embed_names.append(['h_{}'.format(i), 'agg_h_{}'.format(i-1)])
+            node_embed_names.append(['agg_h_{}'.format(n_layers-1)])
+
+            nf.copy_from_parent(node_embed_names=node_embed_names)
+            # forward
+            with mx.autograd.record():
+                pred = model(nf)
+                batch_nids = nf.layer_parent_nid(-1).as_in_context(ctx)
+                batch_labels = labels[batch_nids]
+                loss = loss_fcn(pred, batch_labels)
+                loss = loss.sum() / len(batch_nids)
+
+            loss.backward()
+            trainer.step(batch_size=1)
+
+            node_embed_names = [['h_{}'.format(i)] for i in range(n_layers)]
+            node_embed_names.append([])
+
+            nf.copy_to_parent(node_embed_names=node_embed_names)
+
+        infer_params = infer_model.collect_params()
+
+        for key in infer_params:
+            idx = trainer._param2idx[key]
+            trainer._kvstore.pull(idx, out=infer_params[key].data())
+
+        num_acc = 0.
+
+        for nf, aux in dgl.contrib.sampling.NeighborSampler(g, args.test_batch_size,
+                                                            g.number_of_nodes(),
+                                                            neighbor_type='in',
+                                                            num_hops=n_layers,
+                                                            seed_nodes=test_nid):
+            node_embed_names = [['preprocess']]
+            for i in range(n_layers):
+                node_embed_names.append(['norm'])
+
+            nf.copy_from_parent(node_embed_names=node_embed_names)
+            pred = infer_model(nf)
+            batch_nids = nf.layer_parent_nid(-1).as_in_context(ctx)
+            batch_labels = labels[batch_nids]
+            num_acc += (pred.argmax(axis=1) == batch_labels).sum().asscalar()
+
+        print("Test Accuracy {:.4f}". format(num_acc/n_test_samples))
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser(description='GCN')
+    register_data_args(parser)
+    parser.add_argument("--dropout", type=float, default=0.5,
+            help="dropout probability")
+    parser.add_argument("--gpu", type=int, default=-1,
+            help="gpu")
+    parser.add_argument("--lr", type=float, default=3e-2,
+            help="learning rate")
+    parser.add_argument("--n-epochs", type=int, default=200,
+            help="number of training epochs")
+    parser.add_argument("--batch-size", type=int, default=1000,
+            help="train batch size")
+    parser.add_argument("--test-batch-size", type=int, default=1000,
+            help="test batch size")
+    parser.add_argument("--num-neighbors", type=int, default=2,
+            help="number of neighbors to be sampled")
+    parser.add_argument("--n-hidden", type=int, default=16,
+            help="number of hidden gcn units")
+    parser.add_argument("--n-layers", type=int, default=1,
+            help="number of hidden gcn layers")
+    parser.add_argument("--self-loop", action='store_true',
+            help="graph self-loop (default=False)")
+    parser.add_argument("--weight-decay", type=float, default=5e-4,
+            help="Weight for L2 loss")
+    args = parser.parse_args()
+
+    print(args)
+
+    main(args)
+