-
Notifications
You must be signed in to change notification settings - Fork 5
/
train.py
231 lines (183 loc) · 11.2 KB
/
train.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
import argparse
import os
import time
from functools import partial
from neptuneml_toolkit.train import get_training_config, get_train_nids, get_valid_nids, get_node_labels
from neptuneml_toolkit.metrics import classification_report, save_eval_metrics
from neptuneml_toolkit.transform import save_node_prediction_model_artifacts, normalize_hyperparameter_keys
from neptuneml_toolkit.utils import get_device_type
from neptuneml_toolkit.graphloader import GraphLoader
from neptuneml_toolkit.modelzoo import RGCNEncoder, MLP, MLPFeatureTransformer
from sklearn.metrics import roc_auc_score
import dgl
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
class RGCNNodeClassification(nn.Module):
def __init__(self, etypes, target_ntype, in_sizes, hidden_size, num_bases, out_size, num_encoder_layers, num_decoder_layers):
super(RGCNNodeClassification, self).__init__()
self.target_ntype = target_ntype
self.feature_transformer = MLPFeatureTransformer(in_sizes, hidden_size, per_feat_name=False)
self.encoder = RGCNEncoder(etypes, hidden_size, hidden_size, num_encoder_layers, num_bases=num_bases)
self.decoder = MLP(hidden_size, out_size, num_layers=num_decoder_layers)
def forward(self, g, x):
h = self.feature_transformer(x)
embeddings = self.encoder(g, h)
return self.decoder(embeddings[self.target_ntype])
def get_embeddings(self, g, x, batch_size, device='cpu', num_workers=0):
h = self.feature_transformer(x)
return self.encoder.batch_inference(g, h, batch_size, device=device, num_workers=num_workers)
def save(self, model_file):
torch.save({'model_state_dict': self.state_dict()}, model_file)
def get_model(g, in_size, out_size, hyperparameters, model_file=None, device="cpu"):
model = RGCNNodeClassification(g.etypes,
hyperparameters['target_ntype'],
in_size,
int(hyperparameters['hidden_size']),
int(hyperparameters['num_bases']),
out_size,
int(hyperparameters['num_encoder_layers']),
int(hyperparameters['num_decoder_layers']))
if model_file is not None:
model_dict = torch.load(model_file, map_location=torch.device('cpu'))
model.load_state_dict(model_dict['model_state_dict'])
model = model.to(device)
return model
def evaluate(model, features, labels, loss_fn, dataloader, target_ntype, device="cpu"):
model.eval()
predictions = []
node_ids = []
for i, (input_nodes, output_nodes, subgraphs) in enumerate(dataloader):
batch_features = {ntype: {feat_name: feat[node_idx].to(device) for feat_name, feat in features[ntype].items()}
for ntype, node_idx in input_nodes.items()}
subgraphs = [subgraph.to(device) for subgraph in subgraphs]
logits = model(subgraphs, batch_features)
predictions.append(logits)
node_ids.append(output_nodes[target_ntype])
predictions = torch.cat(predictions, dim=0).detach()
labels = labels[torch.cat(node_ids, dim=0)].to(device)
loss = loss_fn(predictions, labels).item()
metric = eval_metric(labels.cpu().numpy(), torch.sigmoid(predictions).cpu().numpy())
report = classification_report(labels, (torch.sigmoid(predictions) > 0.5).float())
return metric, loss, report
def eval_metric(labels, predicted_labels, weighted=True):
rocauc_list = []
support_weights = [] if weighted else None
for i in range(labels.shape[1]):
# AUC is only defined when there is at least one positive data.
if np.sum(labels[:, i] == 1) > 0 and np.sum(labels[:, i] == 0) > 0:
if weighted:
support_weights.append(sum(labels[:, i]) / labels.shape[0])
rocauc_list.append(roc_auc_score(labels[:, i], predicted_labels[:, i]))
if len(rocauc_list) == 0:
# For imbalance labels, this may happen during training
print('No positively labeled data available. Cannot compute ROC-AUC.')
return 0
_roc_auc_score = np.average(rocauc_list, weights=support_weights)
return _roc_auc_score
def train_n_epochs(model, optimizer, features, labels, loss_fn, train_dataloader, validation_dataloader, n_epochs,
target_ntype, device, model_path, model_file="model.pt", train_log_freq=5):
best_eval_metric = 0
for epoch in range(n_epochs):
t1 = time.time()
for i, (input_nodes, output_nodes, subgraphs) in enumerate(train_dataloader):
batch_features = {ntype: {feat_name: feat[node_idx].to(device) for feat_name, feat in features[ntype].items()}
for ntype, node_idx in input_nodes.items()}
subgraphs = [subgraph.to(device) for subgraph in subgraphs]
logits = model(subgraphs, batch_features)
loss = loss_fn(logits, labels[output_nodes[target_ntype]].to(device))
if (i+1)%train_log_freq == 0:
print("Train Loss: {:.4f}".format(loss.item()))
optimizer.zero_grad()
loss.backward()
optimizer.step()
print("Epoch {:05d}:{:05d} | Epoch Time(s) {:.4f}".format(epoch + 1, n_epochs, time.time() - t1))
metric, val_loss, report = evaluate(model, features, labels, loss_fn, validation_dataloader, target_ntype, device=device)
if metric > best_eval_metric:
print("Validation ROC AUC Score: {:.4f} | Validation loss: {:.4f}".format(metric, val_loss))
model.save(os.path.join(model_path, model_file))
report["roc_auc_score"] = metric
save_eval_metrics(report, model_path)
best_eval_metric = metric
def train(data_path, model_path, devices, hyperparameters):
print("Training config: data_path: {}, model_path: {}, devices: {} hyperparameters: {}".format(data_path,
model_path,
devices,
hyperparameters))
device_type = get_device_type(devices)
graphloader = GraphLoader(data_path)
g = graphloader.graph
print("Loaded graph: {}".format(g))
target_ntype = hyperparameters["target_ntype"]
train_nids = {target_ntype: get_train_nids(g, target_ntype)}
sampler = dgl.dataloading.MultiLayerNeighborSampler([args.num_neighbors] * args.num_encoder_layers)
train_dataloader = dgl.dataloading.NodeDataLoader(g, train_nids, sampler, batch_size=args.batch_size, shuffle=True,
num_workers=0)
valid_nids = {target_ntype: get_valid_nids(g, target_ntype)}
val_sampler = dgl.dataloading.MultiLayerNeighborSampler([args.num_neighbors] * args.num_encoder_layers)
val_dataloader = dgl.dataloading.NodeDataLoader(g, valid_nids, val_sampler, batch_size=args.batch_size, shuffle=False,
num_workers=0)
features_dict = graphloader.get_node_features()
input_sizes = {ntype: {feat_name: features[feat_name].shape[1] for feat_name in features}
for ntype, features in features_dict.items()}
print("Got input features with shape graph: {}".format(input_sizes))
labels = get_node_labels(g, target_ntype).float()
output_size = len(graphloader.label_map[target_ntype])
print("Got training labels with shape graph: {}".format(labels.shape))
loss_fn = F.binary_cross_entropy_with_logits
model = get_model(g, input_sizes, output_size, hyperparameters, device=device_type)
print("Created model: {}".format(model))
optimizer = torch.optim.Adam(model.parameters(), lr=hyperparameters["lr"],
weight_decay=hyperparameters["weight_decay"])
print("Starting model training")
train_n_epochs(model, optimizer, features_dict, labels, loss_fn, train_dataloader, val_dataloader,
hyperparameters["n_epochs"], hyperparameters["target_ntype"], device_type, model_path)
def transform(data_path, model_path, devices, hyperparameters):
hyperparameters = normalize_hyperparameter_keys(hyperparameters)
print("Transform config: data_path: {}, model_path: {}, devices: {} hyperparameters: {}".format(data_path,
model_path,
devices,
hyperparameters))
device_type = get_device_type(devices)
graphloader = GraphLoader(data_path)
g = graphloader.graph
print("Loaded graph: {}".format(g))
target_ntype = hyperparameters["target_ntype"]
features_dict = graphloader.get_node_features()
input_sizes = {ntype: {feat_name: features[feat_name].shape[1] for feat_name in features}
for ntype, features in features_dict.items()}
output_size = len(graphloader.label_map[target_ntype])
print("Got input features with shape graph: {}".format(input_sizes))
model = get_model(g, input_sizes, output_size, hyperparameters, device=device_type, model_file=os.path.join(model_path, "model.pt"))
print("Created model with saved parameters: {}".format(model))
print("Getting model embeddings")
node_embeddings = model.get_embeddings(g, features_dict, batch_size=hyperparameters['batch_size'], device=device_type, num_workers=0)
predictions = torch.sigmoid(model.decoder(node_embeddings[target_ntype]))
print("Saving model artifacts")
save_node_prediction_model_artifacts(model_path, predictions, graphloader, hyperparameters, embeddings=node_embeddings)
def parse_args():
parser = argparse.ArgumentParser()
parser.add_argument("--local", action='store_true', default=False, help='Whether script is running locally')
parser.add_argument("--name", type=str, default='rgcn-node-class')
parser.add_argument("--model", type=str, default='custom')
parser.add_argument("--task", type=str, default='node_class')
parser.add_argument("--property", type=str, default='label')
parser.add_argument("--target_ntype", type=str, default='movie')
parser.add_argument("--num-neighbors", type=int, default=30)
parser.add_argument("--batch-size", type=int, default=1024)
parser.add_argument("--lr", type=float, default=1e-2)
parser.add_argument("--weight-decay", type=float, default=0.)
parser.add_argument("--n-epochs", type=int, default=2)
parser.add_argument("--hidden-size", type=int, default = 128)
parser.add_argument("--num-bases", type=int, default=2)
parser.add_argument("--num-encoder-layers", type=int, default=2)
parser.add_argument("--num-decoder-layers", type=int, default=1)
return parser.parse_args()
if __name__ == "__main__":
args = parse_args()
if args.local:
data_path, model_path, devices = './data', './output', [-1]
else:
data_path, model_path, devices = get_training_config()
train(data_path, model_path, devices, vars(args))