This repository has been archived by the owner on Oct 7, 2024. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathtrain_supervised_graphsage.py
260 lines (229 loc) · 9.44 KB
/
train_supervised_graphsage.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
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
import argparse
import dgl
from dgl.dataloading import (
DataLoader,
NeighborSampler,
as_edge_prediction_sampler,
negative_sampler,
)
import numpy as np
import pandas as pd
import random
import sys
import time
import torch
import torch.nn as nn
import torch.nn.functional as F
import tqdm
from graphsage_model import Model
def evaluate(model, val_dataloader):
model.eval()
score_all = torch.Tensor()
labels_all = torch.Tensor()
with val_dataloader.enable_cpu_affinity():
with torch.no_grad():
for _, (input_nodes, pair_graph, neg_pair_graph, blocks) in enumerate(
tqdm.tqdm(val_dataloader)
):
x = input_nodes
pos_score, pos_label = model(pair_graph, neg_pair_graph, blocks, x)
score_all = torch.cat([score_all, pos_score], 0)
labels_all = torch.cat([labels_all, pos_label], 0)
assert(score_all.shape[0] == labels_all.shape[0])
labels_one_hot = torch.hstack((1.0 - labels_all, labels_all)).float()
bceloss = F.binary_cross_entropy(score_all, labels_one_hot)
return bceloss
def train(args, train_dataloader, val_dataloader, model):
best_model_path = args.model_out
best_val_bceloss = float("inf")
print("learning rate: ", args.lr)
opt = torch.optim.AdamW(model.parameters(), lr=args.lr)
with train_dataloader.enable_cpu_affinity():
all_losses = []
for epoch in range(args.num_epochs):
start = time.time()
step_time = []
model.train()
total_loss = []
for it, (input_nodes, pair_graph, neg_pair_graph, blocks) in enumerate(
train_dataloader
):
opt.zero_grad()
x = input_nodes #blocks[0].srcdata[dgl.NID] #this is equal to input_nodes
pos_score, pos_label = model(pair_graph, neg_pair_graph, blocks, x)
pos_label_oh = torch.hstack((1.0 - pos_label, pos_label)).float() # 0 -> [1, 0] and 1 -> [0, 1]
loss = F.binary_cross_entropy(pos_score, pos_label_oh)
total_loss += loss.detach().item(),
loss.backward()
opt.step()
step_t = time.time() - start
step_time.append(step_t)
start = time.time()
all_losses += total_loss
print("Epoch {:02d} | Train BCELoss {:.4f}".format(epoch, sum(total_loss) / len(train_dataloader)))
model.eval()
if (epoch % args.eval_every == 0) or (epoch == args.num_epochs -1):
bceloss = evaluate(model, val_dataloader)
# update best model if needed
if best_val_bceloss > bceloss:
print("Updating best model")
best_val_bceloss = bceloss
torch.save(model.state_dict(), best_model_path)
print("Epoch {:05d} | val bceloss {:.4f}".format(epoch, bceloss))
def main(arglist):
# random seeds for testing
seed=7
print("random seed set to: ", seed)
random.seed(seed)
np.random.seed(seed)
dgl.seed(seed)
dgl.random.seed(seed)
torch.random.manual_seed(seed)
torch.manual_seed(seed)
parser = argparse.ArgumentParser()
parser.add_argument(
"--mode",
default="cpu",
choices=["cpu", "gpu", "mixed"],
help="Training mode. 'cpu' for CPU training, 'gpu' for pure-GPU training, "
"'mixed' for CPU-GPU mixed training.",
)
parser.add_argument(
"--CSVDataset_dir",
default="/localdisk/akakne/recsys2023/data/supGNN_graph_data/train_graph/recsys_graph",
help="Path to CSVDataset",
)
parser.add_argument(
"--model_out",
default="/localdisk/akakne/recsys2023/data/supGNN_graph_data/train_graph/gnn_model.pt",
type=str,
help="output for model /your_path/model.pt",
)
parser.add_argument("--num_epochs", type=int, default=10)
parser.add_argument("--num_hidden_gnn", type=int, default=128)
parser.add_argument("--num_hidden_mlp", type=int, default=512)
parser.add_argument("--num_layers", type=int, default=2)
parser.add_argument("--fan_out", type=str, default="25,25")
parser.add_argument("--batch_size", type=int, default=40000)
parser.add_argument("--batch_size_eval", type=int, default=100000)
parser.add_argument("--eval_every", type=int, default=1)
parser.add_argument("--lr", type=float, default=1e-4)
# step 1 : parse input arguments
args = parser.parse_args(arglist)
if not torch.cuda.is_available():
args.mode = "cpu"
print(f"Training in {args.mode} mode.")
# step 2 : load CSVDataset files into a heterogeneous graph dataset
print("Loading data")
start = time.time()
# note : set force_reload=False if no changes on input graph (much faster otherwise ingestion ~30min)
dataset = dgl.data.CSVDataset(args.CSVDataset_dir, force_reload=False)
print("time to load dataset from CSVs: ", time.time() - start)
# step 3 : extract the heterogenous graph
train_val_hg = dataset[0]
print(train_val_hg)
print("etype to read train/test/val from: ", train_val_hg.canonical_etypes[0][1])
print("-"*100)
train_val_g = dgl.to_homogeneous(train_val_hg,
edata=["feat", "label", "train_mask", "val_mask", "test_mask"])
# g.edata['prob'] = 1.0 - g.edata['test_mask']
print(train_val_g)
print("-"*100)
# step 4 : create masks and collect edge IDs from
def get_mask_and_eids(graph, key):
forward_mask = graph.edges["e"].data[key]
reverse_mask = graph.edges["sym_e"].data[key]
mask = torch.cat((forward_mask, reverse_mask))
eids = torch.nonzero(mask, as_tuple=False).squeeze()
return mask, eids
full_val_mask, full_val_eids = get_mask_and_eids(train_val_hg, "val_mask")
print("number of validation edges = {}".format(torch.sum(full_val_mask)))
print("-"*100)
# step 5 : create homogenous split graphs by removing edges val and/or test edges
def remove_edges(graph, eids_to_remove):
new_g = dgl.remove_edges(graph, eids_to_remove, store_ids=True)
eidx_orig = new_g.edata[dgl.EID]
assert len(torch.unique(eidx_orig)) == eidx_orig.shape[0] == new_g.num_edges()
return new_g
train_g = remove_edges(train_val_g, full_val_eids)
num_train_edges = train_g.num_edges()
num_val_edges = train_val_g.num_edges() - train_g.num_edges()
print("train edges = {}, val edges = {}".format(num_train_edges, num_val_edges))
print("-"*100)
# step 6 : create sampler & dataloaders
def get_reverse_mapping(graph):
"""
pass homogeneous graph,
assumes forward edge with ID i has reverse edge with ID E + i
"""
E = graph.num_edges() // 2
reverse_eids = torch.cat([torch.arange(E, 2 * E), torch.arange(0, E)])
return reverse_eids
sampler = NeighborSampler([int(fanout) for fanout in args.fan_out.split(",")])
train_sampler = as_edge_prediction_sampler(
sampler,
exclude="reverse_id",
reverse_eids=get_reverse_mapping(train_g),
negative_sampler=negative_sampler.Uniform(1),
)
val_sampler = as_edge_prediction_sampler(
sampler,
exclude="reverse_id",
reverse_eids=get_reverse_mapping(train_val_g),
negative_sampler=negative_sampler.Uniform(1),
)
use_uva = args.mode == "mixed"
# step 7 : collect edge IDs for forward edges in all three graphs
def get_ids(graph, split):
"""
assumption:
-----------
edge i has reverse edge (graph.num_edges() // 2 + i) and viceversa
args:
-----
graph - homogenous graph with forward and backward edges
split - str among ["train_mask", "val_mask", "test_mask"]
returns:
--------
returns forward edges' IDs for the defined split
"""
mask = graph.edata[split] # mask will be of size graph.num_edges()
num_edges = mask.shape[0]
assert(num_edges % 2 == 0)
eidx = torch.nonzero(mask[:num_edges//2], as_tuple=False).squeeze()
return eidx
# sanity check 1
train_eidx = get_ids(train_g, "train_mask")
num_train_edges = torch.sum(train_val_hg.edges[train_val_hg.canonical_etypes[0][1]].data["train_mask"]).int()
assert(train_eidx.shape[0] == num_train_edges)
# sanity check 2
val_eidx = get_ids(train_val_g, "val_mask")
num_val_edges = torch.sum(train_val_hg.edges[train_val_hg.canonical_etypes[0][1]].data["val_mask"]).int()
assert(val_eidx.shape[0] == num_val_edges)
train_dataloader = DataLoader(
train_g,
get_ids(train_g, "train_mask"),
train_sampler,
batch_size=args.batch_size,
shuffle=True,
drop_last=False,
num_workers=4,
use_uva=use_uva,
)
val_dataloader = DataLoader(
train_val_g,
get_ids(train_val_g, "val_mask"),
val_sampler,
batch_size=args.batch_size_eval,
shuffle=True,
drop_last=False,
num_workers=4,
use_uva=use_uva,
)
vocab_size = train_val_g.num_nodes()
model = Model(vocab_size, args.num_hidden_gnn, args.num_hidden_gnn, args.num_hidden_mlp, args.num_layers)
# model training
print("Training...")
train(args, train_dataloader, val_dataloader, model)
if __name__ == "__main__":
main(sys.argv[1:])