[Dataset] Change the Data Type of the Node Features in GINDataset/TUDataset...

[Dataset] Change the Data Type of the Node Features in GINDataset/TUDataset from Float64 to Float32 (#2592)

* Update

* Update

* Update

* update
Co-authored-by: Jinjing Zhou <VoVAllen@users.noreply.github.com>
Co-authored-by: Tong He <hetong007@gmail.com>

docs/source/guide/training-graph.rst

@@ -217,7 +217,7 @@ updating the model.
     opt = torch.optim.Adam(model.parameters())
     for epoch in range(20):
         for batched_graph, labels in dataloader:
-            feats = batched_graph.ndata['attr'].float()
+            feats = batched_graph.ndata['attr']
             logits = model(batched_graph, feats)
             loss = F.cross_entropy(logits, labels)
             opt.zero_grad()

docs/source/guide_cn/training-graph.rst

@@ -165,7 +165,7 @@ DGL内置了常见的图读出函数，例如 :func:`dgl.readout_nodes` 就实
     opt = torch.optim.Adam(model.parameters())
     for epoch in range(20):
         for batched_graph, labels in dataloader:
-            feats = batched_graph.ndata['attr'].float()
+            feats = batched_graph.ndata['attr']
             logits = model(batched_graph, feats)
             loss = F.cross_entropy(logits, labels)
             opt.zero_grad()

examples/mxnet/gin/main.py

@@ -21,7 +21,7 @@ def train(args, net, trainloader, trainer, criterion, epoch):
     for pos, (graphs, labels) in zip(bar, trainloader):
         # batch graphs will be shipped to device in forward part of model
         labels = labels.as_in_context(args.device)
-        feat = graphs.ndata['attr'].astype('float32').as_in_context(args.device)
+        feat = graphs.ndata['attr'].as_in_context(args.device)
 
         with mx.autograd.record():
             graphs = graphs.to(args.device)
@@ -52,7 +52,7 @@ def eval_net(args, net, dataloader, criterion):
     for data in dataloader:
         graphs, labels = data
         labels = labels.as_in_context(args.device)
-        feat = graphs.ndata['attr'].astype('float32').as_in_context(args.device)
+        feat = graphs.ndata['attr'].as_in_context(args.device)
 
         total += len(labels)
         graphs = graphs.to(args.device)

examples/pytorch/gin/main.py

@@ -6,7 +6,7 @@ import torch
 import torch.nn as nn
 import torch.optim as optim
 
-from dgl.data.gindt import GINDataset
+from dgl.data import GINDataset
 from dataloader import GINDataLoader
 from parser import Parser
 from gin import GIN
@@ -22,11 +22,9 @@ def train(args, net, trainloader, optimizer, criterion, epoch):
 
     for pos, (graphs, labels) in zip(bar, trainloader):
         # batch graphs will be shipped to device in forward part of model
-        for key in graphs.node_attr_schemes().keys():
-            graphs.ndata[key] = graphs.ndata[key].float()
         labels = labels.to(args.device)
-        feat = graphs.ndata.pop('attr').to(args.device)
         graphs = graphs.to(args.device)
+        feat = graphs.ndata.pop('attr')
         outputs = net(graphs, feat)
 
         loss = criterion(outputs, labels)
@@ -55,11 +53,9 @@ def eval_net(args, net, dataloader, criterion):
 
     for data in dataloader:
         graphs, labels = data
-        for key in graphs.node_attr_schemes().keys():
-            graphs.ndata[key] = graphs.ndata[key].float()
-        feat = graphs.ndata.pop('attr').to(args.device)
         graphs = graphs.to(args.device)
         labels = labels.to(args.device)
+        feat = graphs.ndata.pop('attr')
         total += len(labels)
         outputs = net(graphs, feat)
         _, predicted = torch.max(outputs.data, 1)

examples/pytorch/gxn/main.py

@@ -48,8 +48,6 @@ def train(model:torch.nn.Module, optimizer, trainloader,
     for batch in trainloader:
         optimizer.zero_grad()
         batch_graphs, batch_labels = batch
-        for (key, value) in batch_graphs.ndata.items():
-            batch_graphs.ndata[key] = value.float()
         batch_graphs = batch_graphs.to(device)
         batch_labels = batch_labels.long().to(device)
         out, l1, l2 = model(batch_graphs, 
@@ -74,8 +72,6 @@ def test(model:torch.nn.Module, loader, device):
     for batch in loader:
         batch_graphs, batch_labels = batch
         num_graphs += batch_labels.size(0)
-        for (key, value) in batch_graphs.ndata.items():
-            batch_graphs.ndata[key] = value.float()
         batch_graphs = batch_graphs.to(device)
         batch_labels = batch_labels.long().to(device)
         out, _, _ = model(batch_graphs, batch_graphs.ndata["feat"])

examples/pytorch/gxn/main_early_stop.py

@@ -4,7 +4,7 @@ from time import time
 from datetime import datetime
 
 import dgl
-from dgl.data import LegacyTUDataset, TUDataset
+from dgl.data import LegacyTUDataset
 from torch.utils.data import random_split
 import torch
 from torch import Tensor
@@ -48,8 +48,6 @@ def train(model:torch.nn.Module, optimizer, trainloader,
     for batch in trainloader:
         optimizer.zero_grad()
         batch_graphs, batch_labels = batch
-        for (key, value) in batch_graphs.ndata.items():
-            batch_graphs.ndata[key] = value.float()
         batch_graphs = batch_graphs.to(device)
         batch_labels = batch_labels.long().to(device)
         out, l1, l2 = model(batch_graphs, 
@@ -74,8 +72,6 @@ def test(model:torch.nn.Module, loader, device):
     for batch in loader:
         batch_graphs, batch_labels = batch
         num_graphs += batch_labels.size(0)
-        for (key, value) in batch_graphs.ndata.items():
-            batch_graphs.ndata[key] = value.float()
         batch_graphs = batch_graphs.to(device)
         batch_labels = batch_labels.long().to(device)
         out, _, _ = model(batch_graphs, batch_graphs.ndata["feat"])
@@ -98,8 +94,6 @@ def validate(model:torch.nn.Module, loader, device,
     for batch in loader:
         batch_graphs, batch_labels = batch
         num_graphs += batch_labels.size(0)
-        for (key, value) in batch_graphs.ndata.items():
-            batch_graphs.ndata[key] = value.float()
         batch_graphs = batch_graphs.to(device)
         batch_labels = batch_labels.long().to(device)
         out, l1, l2 = model(batch_graphs, batch_graphs.ndata["feat"])

examples/pytorch/hgp_sl/main.py

@@ -90,8 +90,6 @@ def train(model:torch.nn.Module, optimizer, trainloader, device):
     for batch in trainloader:
         optimizer.zero_grad()
         batch_graphs, batch_labels = batch
-        for (key, value) in batch_graphs.ndata.items():
-            batch_graphs.ndata[key] = value.float()
         batch_graphs = batch_graphs.to(device)
         batch_labels = batch_labels.long().to(device)
         out = model(batch_graphs, batch_graphs.ndata["feat"])
@@ -113,8 +111,6 @@ def test(model:torch.nn.Module, loader, device):
     for batch in loader:
         batch_graphs, batch_labels = batch
         num_graphs += batch_labels.size(0)
-        for (key, value) in batch_graphs.ndata.items():
-            batch_graphs.ndata[key] = value.float()
         batch_graphs = batch_graphs.to(device)
         batch_labels = batch_labels.long().to(device)
         out = model(batch_graphs, batch_graphs.ndata["feat"])

examples/pytorch/sagpool/main.py

@@ -82,8 +82,6 @@ def train(model:torch.nn.Module, optimizer, trainloader, device):
     for batch in trainloader:
         optimizer.zero_grad()
         batch_graphs, batch_labels = batch
-        for (key, value) in batch_graphs.ndata.items():
-            batch_graphs.ndata[key] = value.float()
         batch_graphs = batch_graphs.to(device)
         batch_labels = batch_labels.long().to(device)
         out = model(batch_graphs)
@@ -105,8 +103,6 @@ def test(model:torch.nn.Module, loader, device):
     for batch in loader:
         batch_graphs, batch_labels = batch
         num_graphs += batch_labels.size(0)
-        for (key, value) in batch_graphs.ndata.items():
-            batch_graphs.ndata[key] = value.float()
         batch_graphs = batch_graphs.to(device)
         batch_labels = batch_labels.long().to(device)
         out = model(batch_graphs)

new-tutorial/5_graph_classification.py

@@ -182,7 +182,7 @@ optimizer = torch.optim.Adam(model.parameters(), lr=0.01)
 
 for epoch in range(20):
     for batched_graph, labels in train_dataloader:
-        pred = model(batched_graph, batched_graph.ndata['attr'].float())
+        pred = model(batched_graph, batched_graph.ndata['attr'])
         loss = F.cross_entropy(pred, labels)
         optimizer.zero_grad()
         loss.backward()
@@ -191,7 +191,7 @@ for epoch in range(20):
 num_correct = 0
 num_tests = 0
 for batched_graph, labels in test_dataloader:
-    pred = model(batched_graph, batched_graph.ndata['attr'].float())
+    pred = model(batched_graph, batched_graph.ndata['attr'])
     num_correct += (pred.argmax(1) == labels).sum().item()
     num_tests += len(labels)
 

python/dgl/data/gindt.py

@@ -49,7 +49,7 @@ class GINDataset(DGLBuiltinDataset):
     >>> g, label = data[128]
     >>> g
     Graph(num_nodes=13, num_edges=26,
-          ndata_schemes={'label': Scheme(shape=(), dtype=torch.int64), 'attr': Scheme(shape=(7,), dtype=torch.float64)}
+          ndata_schemes={'label': Scheme(shape=(), dtype=torch.int64), 'attr': Scheme(shape=(7,), dtype=torch.float32)}
           edata_schemes={})
     >>> label
     tensor(1)
@@ -61,7 +61,7 @@ class GINDataset(DGLBuiltinDataset):
     >>> batched_labels = torch.tensor(labels)
     >>> batched_graphs
     Graph(num_nodes=330, num_edges=748,
-          ndata_schemes={'label': Scheme(shape=(), dtype=torch.int64), 'attr': Scheme(shape=(7,), dtype=torch.float64)}
+          ndata_schemes={'label': Scheme(shape=(), dtype=torch.int64), 'attr': Scheme(shape=(7,), dtype=torch.float32)}
           edata_schemes={})
     """
 
@@ -175,7 +175,6 @@ class GINDataset(DGLBuiltinDataset):
                     if tmp == len(nrow):
                         # no node attributes
                         nrow = [int(w) for w in nrow]
-                        nattr = None
                     elif tmp > len(nrow):
                         nrow = [int(w) for w in nrow[:tmp]]
                         nattr = [float(w) for w in nrow[tmp:]]
@@ -208,10 +207,8 @@ class GINDataset(DGLBuiltinDataset):
 
                 if nattrs != []:
                     nattrs = np.stack(nattrs)
-                    g.ndata['attr'] = F.tensor(nattrs)
+                    g.ndata['attr'] = F.tensor(nattrs, F.float32)
                     self.nattrs_flag = True
-                else:
-                    nattrs = None
 
                 g.ndata['label'] = F.tensor(nlabels)
                 if len(self.nlabel_dict) > 1:
@@ -230,7 +227,6 @@ class GINDataset(DGLBuiltinDataset):
         if not self.nattrs_flag:
             if self.verbose:
                 print('there are no node features in this dataset!')
-            label2idx = {}
             # generate node attr by node degree
             if self.degree_as_nlabel:
                 if self.verbose:
@@ -265,7 +261,7 @@ class GINDataset(DGLBuiltinDataset):
                     g.number_of_nodes(), len(label2idx)))
                 attr[range(g.number_of_nodes()), [label2idx[nl]
                                                   for nl in F.asnumpy(g.ndata['label']).tolist()]] = 1
-                g.ndata['attr'] = F.tensor(attr)
+                g.ndata['attr'] = F.tensor(attr, F.float32)
 
         # after load, get the #classes and #dim
         self.gclasses = len(self.glabel_dict)

python/dgl/data/tu.py

 from __future__ import absolute_import
 import numpy as np
 import os
-import random
-
 
 from .dgl_dataset import DGLBuiltinDataset
 from .utils import loadtxt, save_graphs, load_graphs, save_info, load_info
 from .. import backend as F
-from ..utils import retry_method_with_fix
 from ..convert import graph as dgl_graph
 
 class LegacyTUDataset(DGLBuiltinDataset):
@@ -48,7 +45,7 @@ class LegacyTUDataset(DGLBuiltinDataset):
     >>> g, label = data[1024]
     >>> g
     Graph(num_nodes=88, num_edges=410,
-          ndata_schemes={'feat': Scheme(shape=(89,), dtype=torch.float64), '_ID': Scheme(shape=(), dtype=torch.int64)}
+          ndata_schemes={'feat': Scheme(shape=(89,), dtype=torch.float32), '_ID': Scheme(shape=(), dtype=torch.int64)}
           edata_schemes={'_ID': Scheme(shape=(), dtype=torch.int64)})
     >>> label
     tensor(1)
@@ -60,7 +57,7 @@ class LegacyTUDataset(DGLBuiltinDataset):
     >>> batched_labels = torch.tensor(labels)
     >>> batched_graphs
     Graph(num_nodes=9539, num_edges=47382,
-          ndata_schemes={'feat': Scheme(shape=(89,), dtype=torch.float64), '_ID': Scheme(shape=(), dtype=torch.int64)}
+          ndata_schemes={'feat': Scheme(shape=(89,), dtype=torch.float32), '_ID': Scheme(shape=(), dtype=torch.int64)}
           edata_schemes={'_ID': Scheme(shape=(), dtype=torch.int64)})
 
     Notes
@@ -121,7 +118,7 @@ class LegacyTUDataset(DGLBuiltinDataset):
             g.ndata['node_label'] = F.tensor(DS_node_labels)
             one_hot_node_labels = self._to_onehot(DS_node_labels)
             for idxs, g in zip(node_idx_list, self.graph_lists):
-                g.ndata['feat'] = F.tensor(one_hot_node_labels[idxs, :])
+                g.ndata['feat'] = F.tensor(one_hot_node_labels[idxs, :], F.float32)
             self.data_mode = "node_label"
         except IOError:
             print("No Node Label Data")
@@ -132,14 +129,15 @@ class LegacyTUDataset(DGLBuiltinDataset):
             if DS_node_attr.ndim == 1:
                 DS_node_attr = np.expand_dims(DS_node_attr, -1)
             for idxs, g in zip(node_idx_list, self.graph_lists):
-                g.ndata['feat'] = F.tensor(DS_node_attr[idxs, :])
+                g.ndata['feat'] = F.tensor(DS_node_attr[idxs, :], F.float32)
             self.data_mode = "node_attr"
         except IOError:
             print("No Node Attribute Data")
 
         if 'feat' not in g.ndata.keys():
             for idxs, g in zip(node_idx_list, self.graph_lists):
-                g.ndata['feat'] = np.ones((g.number_of_nodes(), self.hidden_size))
+                g.ndata['feat'] = F.ones((g.number_of_nodes(), self.hidden_size),
+                                         F.float32, F.cpu())
             self.data_mode = "constant"
             if self.verbose:
                 print("Use Constant one as Feature with hidden size {}".format(self.hidden_size))