transH_pytorch.py

#!/usr/bin/env python
# -*- coding: utf-8 -*-
# @Date    : 2017-12-25 14:28:38
# @Author  : jimmy (jimmywangheng@qq.com)
# @Link    : http://sdcs.sysu.edu.cn
# @Version : $Id$

import os

import torch
import torch.autograd as autograd
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim

import numpy as np
import time
import datetime
import random

from utils import *
from data import *
from evaluation import *
import loss
import model

from hyperboard import Agent

USE_CUDA = torch.cuda.is_available()

"""
The meaning of parameters:
self.dataset: Which dataset is used to train the model? Such as 'FB15k', 'WN18', etc.
self.learning_rate: Initial learning rate (lr) of the model.
self.early_stopping_round: How many times will lr decrease? If set to 0, it remains constant.
self.L1_flag: If set to True, use L1 distance as dissimilarity; else, use L2.
self.embedding_size: The embedding size of entities and relations.
self.num_batches: How many batches to train in one epoch?
self.train_times: The maximum number of epochs for training.
self.margin: The margin set for MarginLoss.
self.filter: Whether to check a generated negative sample is false negative.
self.momentum: The momentum of the optimizer.
self.optimizer: Which optimizer to use? Such as SGD, Adam, etc.
self.loss_function: Which loss function to use? Typically, we use margin loss.
self.entity_total: The number of different entities.
self.relation_total: The number of different relations.
self.batch_size: How many instances is contained in one batch?
"""

class Config(object):
	def __init__(self):
		self.dataset = None
		self.learning_rate = 0.001
		self.early_stopping_round = 0
		self.L1_flag = True
		self.embedding_size = 100
		self.num_batches = 100
		self.train_times = 1000
		self.margin = 1.0
		self.filter = True
		self.momentum = 0.9
		self.optimizer = optim.Adam
		self.loss_function = loss.marginLoss
		self.entity_total = 0
		self.relation_total = 0
		self.batch_size = 0

if __name__ == "__main__":

	import argparse
	argparser = argparse.ArgumentParser()

	"""
	The meaning of some parameters:
	seed: Fix the random seed. Except for 0, which means no setting of random seed.
	port: The port number used by hyperboard, 
	which is a demo showing training curves in real time.
	You can refer to https://github.com/WarBean/hyperboard to know more.
	num_processes: Number of processes used to evaluate the result.
	"""

	argparser.add_argument('-d', '--dataset', type=str)
	argparser.add_argument('-l', '--learning_rate', type=float, default=0.001)
	argparser.add_argument('-es', '--early_stopping_round', type=int, default=0)
	argparser.add_argument('-L', '--L1_flag', type=int, default=1)
	argparser.add_argument('-em', '--embedding_size', type=int, default=100)
	argparser.add_argument('-nb', '--num_batches', type=int, default=100)
	argparser.add_argument('-n', '--train_times', type=int, default=1000)
	argparser.add_argument('-m', '--margin', type=float, default=1.0)
	argparser.add_argument('-f', '--filter', type=int, default=1)
	argparser.add_argument('-mo', '--momentum', type=float, default=0.9)
	argparser.add_argument('-s', '--seed', type=int, default=0)
	argparser.add_argument('-op', '--optimizer', type=int, default=1)
	argparser.add_argument('-lo', '--loss_type', type=int, default=0)
	argparser.add_argument('-p', '--port', type=int, default=5000)
	argparser.add_argument('-np', '--num_processes', type=int, default=4)

	args = argparser.parse_args()

	# Start the hyperboard agent
	agent = Agent(address='127.0.0.1', port=args.port)

	if args.seed != 0:
		torch.manual_seed(args.seed)

	trainTotal, trainList, trainDict = loadTriple('./data/' + args.dataset, 'train2id.txt')
	validTotal, validList, validDict = loadTriple('./data/' + args.dataset, 'valid2id.txt')
	tripleTotal, tripleList, tripleDict = loadTriple('./data/' + args.dataset, 'triple2id.txt')
	config = Config()
	config.dataset = args.dataset
	config.learning_rate = args.learning_rate

	config.early_stopping_round = args.early_stopping_round

	if args.L1_flag == 1:
		config.L1_flag = True
	else:
		config.L1_flag = False

	config.embedding_size = args.embedding_size
	config.num_batches = args.num_batches
	config.train_times = args.train_times
	config.margin = args.margin

	if args.filter == 1:
		config.filter = True
	else:
		config.filter = False

	config.momentum = args.momentum

	if args.optimizer == 0:
		config.optimizer = optim.SGD
	elif args.optimizer == 1:
		config.optimizer = optim.Adam
	elif args.optimizer == 2:
		config.optimizer = optim.RMSprop

	if args.loss_type == 0:
		config.loss_function = loss.marginLoss

	config.entity_total = getAnythingTotal('./data/' + config.dataset, 'entity2id.txt')
	config.relation_total = getAnythingTotal('./data/' + config.dataset, 'relation2id.txt')
	config.batch_size = trainTotal // config.num_batches

	shareHyperparameters = {'dataset': args.dataset,
		'learning_rate': args.learning_rate,
		'early_stopping_round': args.early_stopping_round,
		'L1_flag': args.L1_flag,
		'embedding_size': args.embedding_size,
		'margin': args.margin,
		'filter': args.filter,
		'momentum': args.momentum,
		'seed': args.seed,
		'optimizer': args.optimizer,
		'loss_type': args.loss_type,
		}

	trainHyperparameters = shareHyperparameters.copy()
	trainHyperparameters.update({'type': 'train_loss'})

	validHyperparameters = shareHyperparameters.copy()
	validHyperparameters.update({'type': 'valid_loss'})

	hit10Hyperparameters = shareHyperparameters.copy()
	hit10Hyperparameters.update({'type': 'hit10'})

	meanrankHyperparameters = shareHyperparameters.copy()
	meanrankHyperparameters.update({'type': 'mean_rank'})

	trainCurve = agent.register(trainHyperparameters, 'train loss', overwrite=True)
	validCurve = agent.register(validHyperparameters, 'valid loss', overwrite=True)
	hit10Curve = agent.register(hit10Hyperparameters, 'hit@10', overwrite=True)
	meanrankCurve = agent.register(meanrankHyperparameters, 'mean rank', overwrite=True)

	loss_function = config.loss_function()
	model = model.TransHModel(config)

	if USE_CUDA:
		model.cuda()
		loss_function.cuda()
		longTensor = torch.cuda.LongTensor
		floatTensor = torch.cuda.FloatTensor

	else:
		longTensor = torch.LongTensor
		floatTensor = torch.FloatTensor

	optimizer = config.optimizer(model.parameters(), lr=config.learning_rate)
	margin = autograd.Variable(floatTensor([config.margin]))

	start_time = time.time()

	filename = '_'.join(
		['l', str(args.learning_rate),
		 'es', str(args.early_stopping_round),
		 'L', str(args.L1_flag),
		 'em', str(args.embedding_size),
		 'nb', str(args.num_batches),
		 'n', str(args.train_times),
		 'm', str(args.margin),
		 'f', str(args.filter),
		 'mo', str(args.momentum),
		 's', str(args.seed),
		 'op', str(args.optimizer),
		 'lo', str(args.loss_type),]) + '_TransH.ckpt'

	trainBatchList = getBatchList(trainList, config.num_batches)

	for epoch in range(config.train_times):
		total_loss = floatTensor([0.0])
		random.shuffle(trainBatchList)
		for batchList in trainBatchList:
			if config.filter == True:
				pos_h_batch, pos_t_batch, pos_r_batch, neg_h_batch, neg_t_batch, neg_r_batch = getBatch_filter_all(batchList, 
					config.entity_total, tripleDict)
			else:
				pos_h_batch, pos_t_batch, pos_r_batch, neg_h_batch, neg_t_batch, neg_r_batch = getBatch_raw_all(batchList, 
					config.entity_total)

			batch_entity_set = set(pos_h_batch + pos_t_batch + neg_h_batch + neg_t_batch)
			batch_relation_set = set(pos_r_batch + neg_r_batch)
			batch_entity_list = list(batch_entity_set)
			batch_relation_list = list(batch_relation_set)

			pos_h_batch = autograd.Variable(longTensor(pos_h_batch))
			pos_t_batch = autograd.Variable(longTensor(pos_t_batch))
			pos_r_batch = autograd.Variable(longTensor(pos_r_batch))
			neg_h_batch = autograd.Variable(longTensor(neg_h_batch))
			neg_t_batch = autograd.Variable(longTensor(neg_t_batch))
			neg_r_batch = autograd.Variable(longTensor(neg_r_batch))

			model.zero_grad()
			pos, neg = model(pos_h_batch, pos_t_batch, pos_r_batch, neg_h_batch, neg_t_batch, neg_r_batch)

			if args.loss_type == 0:
				losses = loss_function(pos, neg, margin)
			else:
				losses = loss_function(pos, neg)
			ent_embeddings = model.ent_embeddings(torch.cat([pos_h_batch, pos_t_batch, neg_h_batch, neg_t_batch]))
			rel_embeddings = model.rel_embeddings(torch.cat([pos_r_batch, neg_r_batch]))
			norm_embeddings = model.norm_embeddings(torch.cat([pos_r_batch, neg_r_batch]))
			losses += loss.orthogonalLoss(rel_embeddings, norm_embeddings)
			losses = losses + loss.normLoss(ent_embeddings) + loss.normLoss(rel_embeddings)
			
			losses.backward()
			optimizer.step()
			total_loss += losses.data

			normalize_norm_emb = F.normalize(model.norm_embeddings.weight.data[longTensor(batch_relation_list)], p=2, dim=1)
			model.norm_embeddings.weight.data[longTensor(batch_relation_list)] = normalize_norm_emb

		agent.append(trainCurve, epoch, total_loss[0])

		if epoch % 10 == 0:
			now_time = time.time()
			print(now_time - start_time)
			print("Train total loss: %d %f" % (epoch, total_loss[0]))

		if epoch % 10 == 0:
			if config.filter == True:
				pos_h_batch, pos_t_batch, pos_r_batch, neg_h_batch, neg_t_batch, neg_r_batch = getBatch_filter_random(validList, 
					config.batch_size, config.entity_total, tripleDict)
			else:
				pos_h_batch, pos_t_batch, pos_r_batch, neg_h_batch, neg_t_batch, neg_r_batch = getBatch_raw_random(validList, 
					config.batch_size, config.entity_total)
			pos_h_batch = autograd.Variable(longTensor(pos_h_batch))
			pos_t_batch = autograd.Variable(longTensor(pos_t_batch))
			pos_r_batch = autograd.Variable(longTensor(pos_r_batch))
			neg_h_batch = autograd.Variable(longTensor(neg_h_batch))
			neg_t_batch = autograd.Variable(longTensor(neg_t_batch))
			neg_r_batch = autograd.Variable(longTensor(neg_r_batch))

			pos, neg = model(pos_h_batch, pos_t_batch, pos_r_batch, neg_h_batch, neg_t_batch, neg_r_batch)

			if args.loss_type == 0:
				losses = loss_function(pos, neg, margin)
			else:
				losses = loss_function(pos, neg)
			ent_embeddings = model.ent_embeddings(torch.cat([pos_h_batch, pos_t_batch, neg_h_batch, neg_t_batch]))
			rel_embeddings = model.rel_embeddings(torch.cat([pos_r_batch, neg_r_batch]))
			norm_embeddings = model.norm_embeddings(torch.cat([pos_r_batch, neg_r_batch]))
			losses += loss.orthogonalLoss(rel_embeddings, norm_embeddings)
			losses = losses + loss.normLoss(ent_embeddings) + loss.normLoss(rel_embeddings)

			print("Valid batch loss: %d %f" % (epoch, losses.data[0]))
			agent.append(validCurve, epoch, losses.data[0])

		if config.early_stopping_round > 0:
			if epoch == 0:
				ent_embeddings = model.ent_embeddings.weight.data
				rel_embeddings = model.rel_embeddings.weight.data
				norm_embeddings = model.norm_embeddings.weight.data
				L1_flag = model.L1_flag
				filter = model.filter	
				hit10, best_meanrank = evaluation_transH(validList, tripleDict, ent_embeddings, rel_embeddings, 
					norm_embeddings, L1_flag, filter, k=config.batch_size, num_processes=args.num_processes)
				agent.append(hit10Curve, epoch, hit10)
				agent.append(meanrankCurve, epoch, best_meanrank)
				torch.save(model, os.path.join('./model/' + args.dataset, filename))	
				best_epoch = 0
				meanrank_not_decrease_time = 0
				lr_decrease_time = 0
				#if USE_CUDA:
					#model.cuda()

			# Evaluate on validation set for every 5 epochs
			elif epoch % 5 == 0:
				ent_embeddings = model.ent_embeddings.weight.data
				rel_embeddings = model.rel_embeddings.weight.data
				norm_embeddings = model.norm_embeddings.weight.data
				L1_flag = model.L1_flag
				filter = model.filter
				hit10, now_meanrank = evaluation_transH(validList, tripleDict, ent_embeddings, rel_embeddings, 
					norm_embeddings, L1_flag, filter, k=config.batch_size, num_processes=args.num_processes)
				agent.append(hit10Curve, epoch, hit10)
				agent.append(meanrankCurve, epoch, now_meanrank)
				if now_meanrank < best_meanrank:
					meanrank_not_decrease_time = 0
					best_meanrank = now_meanrank
					torch.save(model, os.path.join('./model/' + args.dataset, filename))
				else:
					meanrank_not_decrease_time += 1
					# If the result hasn't improved for consecutive 5 evaluations, decrease learning rate
					if meanrank_not_decrease_time == 5:
						lr_decrease_time += 1
						if lr_decrease_time == config.early_stopping_round:
							break
						else:
							optimizer.param_groups[0]['lr'] *= 0.5
							meanrank_not_decrease_time = 0
				#if USE_CUDA:
					#model.cuda()

		elif (epoch + 1) % 10 == 0 or epoch == 0:
			torch.save(model, os.path.join('./model/' + args.dataset, filename))


	testTotal, testList, testDict = loadTriple('./data/' + args.dataset, 'test2id.txt')
	oneToOneTotal, oneToOneList, oneToOneDict = loadTriple('./data/' + args.dataset, 'one_to_one.txt')
	oneToManyTotal, oneToManyList, oneToManyDict = loadTriple('./data/' + args.dataset, 'one_to_many.txt')
	manyToOneTotal, manyToOneList, manyToOneDict = loadTriple('./data/' + args.dataset, 'many_to_one.txt')
	manyToManyTotal, manyToManyList, manyToManyDict = loadTriple('./data/' + args.dataset, 'many_to_many.txt')
	
	ent_embeddings = model.ent_embeddings.weight.data
	rel_embeddings = model.rel_embeddings.weight.data
	norm_embeddings = model.norm_embeddings.weight.data
	L1_flag = model.L1_flag
	filter = model.filter

	hit10Test, meanrankTest = evaluation_transH(testList, tripleDict, ent_embeddings, rel_embeddings, norm_embeddings, L1_flag, filter, head=0)

	hit10OneToOneHead, meanrankOneToOneHead = evaluation_transH(oneToOneList, tripleDict, ent_embeddings, rel_embeddings, norm_embeddings, L1_flag, filter, head=1)
	hit10OneToManyHead, meanrankOneToManyHead = evaluation_transH(oneToManyList, tripleDict, ent_embeddings, rel_embeddings, norm_embeddings, L1_flag, filter, head=1)
	hit10ManyToOneHead, meanrankManyToOneHead = evaluation_transH(manyToOneList, tripleDict, ent_embeddings, rel_embeddings, norm_embeddings, L1_flag, filter, head=1)
	hit10ManyToManyHead, meanrankManyToManyHead = evaluation_transH(manyToManyList, tripleDict, ent_embeddings, rel_embeddings, norm_embeddings, L1_flag, filter, head=1)

	hit10OneToOneTail, meanrankOneToOneTail = evaluation_transH(oneToOneList, tripleDict, ent_embeddings, rel_embeddings, norm_embeddings, L1_flag, filter, head=2)
	hit10OneToManyTail, meanrankOneToManyTail = evaluation_transH(oneToManyList, tripleDict, ent_embeddings, rel_embeddings, norm_embeddings, L1_flag, filter, head=2)
	hit10ManyToOneTail, meanrankManyToOneTail = evaluation_transH(manyToOneList, tripleDict, ent_embeddings, rel_embeddings, norm_embeddings, L1_flag, filter, head=2)
	hit10ManyToManyTail, meanrankManyToManyTail = evaluation_transH(manyToManyList, tripleDict, ent_embeddings, rel_embeddings, norm_embeddings, L1_flag, filter, head=2)	

	writeList = [filename, 
		'testSet', '%.6f' % hit10Test, '%.6f' % meanrankTest, 
		'one_to_one_head', '%.6f' % hit10OneToOneHead, '%.6f' % meanrankOneToOneHead, 
		'one_to_many_head', '%.6f' % hit10OneToManyHead, '%.6f' % meanrankOneToManyHead, 
		'many_to_one_head', '%.6f' % hit10ManyToOneHead, '%.6f' % meanrankManyToOneHead, 
		'many_to_many_head', '%.6f' % hit10ManyToManyHead, '%.6f' % meanrankManyToManyHead,
		'one_to_one_tail', '%.6f' % hit10OneToOneTail, '%.6f' % meanrankOneToOneTail, 
		'one_to_many_tail', '%.6f' % hit10OneToManyTail, '%.6f' % meanrankOneToManyTail, 
		'many_to_one_tail', '%.6f' % hit10ManyToOneTail, '%.6f' % meanrankManyToOneTail, 
		'many_to_many_tail', '%.6f' % hit10ManyToManyTail, '%.6f' % meanrankManyToManyTail,]

	# Write the result into file
	with open(os.path.join('./result/', args.dataset + '.txt'), 'a') as fw:
		fw.write('\t'.join(writeList) + '\n')