train.py

# Copyright 2021 Alibaba Group Holding Limited. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# =============================================================================
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

import argparse
import datetime
import os
import sys
import time
import torch

import graphlearn as gl
import graphlearn.python.nn.pytorch as thg
import numpy as np
import torch.distributed as dist
import torch.nn.functional as F

from torch_geometric.nn import GCNConv
from torch_geometric.data import Data
from tqdm import tqdm

from gcn import GCN

os.environ["ODPS_CONFIG_FILE_PATH"] = "odps_config.ini"
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

def init_env(args):
  world_size = 1
  rank = 0
  if args.ddp:
    if torch.cuda.is_available():
      dist.init_process_group('nccl')
    else:
      dist.init_process_group('gloo')
    world_size = dist.get_world_size()
    rank = dist.get_rank()
  print('world_size:', world_size, 'rank:', rank)
  args.use_mp = args.client_num > 0
  args.world_size = world_size
  args.rank = rank


def load_graph(args):
  dataset_folder = args.dataset_folder
  node_type = 'item'
  edge_type = 'relation'
  # shoud be split when distributed training.
  node_path = dataset_folder + "node_table"
  edge_path = dataset_folder + "edge_table"
  train_path = dataset_folder + "train_table"
  val_path = dataset_folder + "val_table"
  test_path = dataset_folder + "test_table"


  g = gl.Graph()                                                           \
        .node(node_path, node_type=node_type,
              decoder=gl.Decoder(labeled=True,
                                 attr_types=["float"] * args.features_num,
                                 attr_delimiter=":"))                      \
        .edge(edge_path,
              edge_type=(node_type, node_type, edge_type),
              decoder=gl.Decoder(weighted=True), directed=False)           \
        .node(train_path, node_type=node_type,
              decoder=gl.Decoder(weighted=True), mask=gl.Mask.TRAIN)       \
        .node(val_path, node_type=node_type,
              decoder=gl.Decoder(weighted=True), mask=gl.Mask.VAL)         \
        .node(test_path, node_type=node_type,
              decoder=gl.Decoder(weighted=True), mask=gl.Mask.TEST)
  return g

def query(g, args, mask=gl.Mask.TRAIN):
  if mask == gl.Mask.TRAIN:
    seed = g.V('item', mask=mask).batch(args.train_batch_size).shuffle(traverse=True).alias('src')
  else:
    seed = g.V('item', mask=mask).batch(args.test_batch_size).alias('src')
  seed.outV('relation').sample(args.nbrs_num).by('full').alias('src_hop1')
  return seed.values()

def induce_func(data_dict):
  """induce the src and 1-hop neighhbor to a list of pyG `Data`."""
  src = data_dict['src']
  nbr = data_dict['src_hop1']
  subgraphs = []
  offset = 0
  for i in range (src.ids.size):
    float_attrs = np.expand_dims(src.float_attrs[i], axis=0)
    labels = np.expand_dims(src.labels[i], axis=0)
    row, col = [], []
    begin, end = offset, offset + nbr.offsets[i]
    float_attrs = np.concatenate((float_attrs, nbr.float_attrs[begin:end]), axis=0)
    labels = np.concatenate((labels, nbr.labels[begin:end]), axis=0)
    for j in range(nbr.offsets[i]):
      row.append(0)
      col.append(j+1)
      col.append(0)
      row.append(j+1)
    offset += nbr.offsets[i]
    edge_index = np.stack([np.array(row), np.array(col)], axis=0)
    subgraph = Data(torch.from_numpy(float_attrs),
                    torch.from_numpy(edge_index).to(torch.long),
                    y=torch.from_numpy(labels).to(torch.long))
    subgraphs.append(subgraph)
  return subgraphs

def train(model, loader, optimizer, args):
  model.train()
  for i, data in tqdm(enumerate(loader)):
    optimizer.zero_grad()
    data = data.to(device)
    x = model(data)
    x = F.log_softmax(x, dim=1)
    loss = F.nll_loss(x, data.y)
    print('loss: ', loss.item())
    loss.backward()
    optimizer.step()

@torch.no_grad()
def test(model, loader, args):
  model.eval()
  y_pred, y_true = [], []
  for i, data in tqdm(enumerate(loader)):
    data = data.to(device)
    y_pred.append(F.log_softmax(model(data), dim=1).cpu().max(1)[1])
    y_true.append(data.y.view(-1).cpu().to(torch.float))
  test_pred, test_true = torch.cat(y_pred), torch.cat(y_true)
  acc = test_pred.eq(test_true).sum().item() / test_true.size()[0]
  return acc

def run(args):
  gl.set_tape_capacity(1)
  g = load_graph(args)
  gl.set_tracker_mode(0)
  if args.use_mp:
    thg.set_client_num(args.client_num)
    thg.launch_server(g)
  else:
    g.init(task_index=args.rank, hosts=thg.get_cluster_spec())

  # train loader
  train_query = query(g, args, mask=gl.Mask.TRAIN)
  if args.use_mp:
    train_dataset = thg.Dataset(train_query, window=10, induce_func=induce_func, graph=g)
    graph_counts = thg.get_counts()
    while True:
      if not graph_counts:
        time.sleep(1)
      else:
        item_count_per_server = graph_counts[gl.get_mask_type('item', gl.Mask.TRAIN)]
        break
  else:
    train_dataset = thg.Dataset(train_query, window=10, induce_func=induce_func)
    item_count_per_server = g.server_get_stats()[gl.get_mask_type('item', gl.Mask.TRAIN)]

  print('item node count per server: ', item_count_per_server)
  length_per_worker =  min(item_count_per_server) // args.train_batch_size
  print('length_per_worker being set to: ' + str(length_per_worker))
  train_loader = thg.PyGDataLoader(train_dataset, multi_process=args.use_mp, length=length_per_worker)

  # test loader
  test_query = query(g, args, mask=gl.Mask.TEST)
  if args.use_mp:
    test_dataset = thg.Dataset(test_query, window=10, induce_func=induce_func, graph=g)
  else:
    test_dataset = thg.Dataset(test_query, window=10, induce_func=induce_func)
  test_loader = thg.PyGDataLoader(test_dataset, multi_process=args.use_mp)

  # define model
  model = GCN(input_dim=args.features_num,
              hidden_dim=args.hidden_dim,
              output_dim=args.class_num,
              depth=args.depth,
              drop_rate=args.drop_rate).to(device)
  if dist.is_initialized():
    model = torch.nn.parallel.DistributedDataParallel(model)
  optimizer = torch.optim.Adam(model.parameters(), lr=args.learning_rate)

  # train and test
  for epoch in range(0, args.epoch):
    train(model, train_loader, optimizer, args)
    test_acc = test(model, test_loader, args)
    log = 'Epoch: {:03d}, Test: {:.4f}'
    print(log.format(epoch, test_acc))

  if not args.use_mp:
    g.close()


if __name__ == "__main__":
  cur_path = sys.path[0]
  argparser = argparse.ArgumentParser("Train GCN.")
  argparser.add_argument('--dataset_folder', type=str,
                         default=os.path.join(cur_path, '../../data/cora/'),
                         help="Dataset Folder, list files are node_table, edge_table, "
                              "train_table, val_table and test_table")
  argparser.add_argument('--class_num', type=int, default=7)
  argparser.add_argument('--features_num', type=int, default=1433)
  argparser.add_argument('--train_batch_size', type=int, default=140)
  argparser.add_argument('--val_batch_size', type=int, default=300)
  argparser.add_argument('--test_batch_size', type=int, default=1000)
  argparser.add_argument('--hidden_dim', type=int, default=128)
  argparser.add_argument('--depth', type=int, default=2)
  argparser.add_argument('--nbrs_num', type=int, default=100)
  argparser.add_argument('--drop_rate', type=float, default=0.0)
  argparser.add_argument('--learning_rate', type=float, default=0.01)
  argparser.add_argument('--epoch', type=int, default=60)
  argparser.add_argument('--client_num', type=int, default=0,
                         help="The number of graphlearn client on each pytorch worker, "
                              "which is used as `num_workers` of pytorh dataloader.")
  argparser.add_argument('--ddp', action="store_true", help="Whether use pytorch ddp.")
  args = argparser.parse_args()

  init_env(args)

  run(args)