"""
Train a model on TACRED.
"""

import os
import sys
from datetime import datetime
import time
import numpy as np
import random
import argparse
from shutil import copyfile
import torch
import torch.nn as nn
import torch.optim as optim
from torch.autograd import Variable

from data.loader import DataLoader
from model.trainer import GCNTrainer
from utils import torch_utils, scorer, constant, helper
from utils.vocab import Vocab

parser = argparse.ArgumentParser()
parser.add_argument('--data_dir', type=str, default='dataset/tacred')
parser.add_argument('--vocab_dir', type=str, default='dataset/vocab')
parser.add_argument('--emb_dim', type=int, default=300, help='Word embedding dimension.')
parser.add_argument('--ner_dim', type=int, default=30, help='NER embedding dimension.')
parser.add_argument('--pos_dim', type=int, default=30, help='POS embedding dimension.')
parser.add_argument('--hidden_dim', type=int, default=300, help='RNN hidden state size.')
parser.add_argument('--num_layers', type=int, default=2, help='Num of RNN layers.')
parser.add_argument('--input_dropout', type=float, default=0.5, help='Input dropout rate.')
parser.add_argument('--gcn_dropout', type=float, default=0.5, help='GCN layer dropout rate.')
parser.add_argument('--word_dropout', type=float, default=0.04, help='The rate at which randomly set a word to UNK.')
parser.add_argument('--topn', type=int, default=1e10, help='Only finetune top N word embeddings.')
parser.add_argument('--lower', dest='lower', action='store_true', help='Lowercase all words.')
parser.add_argument('--no-lower', dest='lower', action='store_false')
parser.set_defaults(lower=False)

parser.add_argument('--heads', type=int, default=3, help='Num of heads in multi-head attention.')
parser.add_argument('--sublayer_first', type=int, default=2, help='Num of the first sublayers in dcgcn block.')
parser.add_argument('--sublayer_second', type=int, default=4, help='Num of the second sublayers in dcgcn block.')

parser.add_argument('--pooling', choices=['max', 'avg', 'sum'], default='max', help='Pooling function type. Default max.')
parser.add_argument('--pooling_l2', type=float, default=0.002, help='L2-penalty for all pooling output.')
parser.add_argument('--mlp_layers', type=int, default=1, help='Number of output mlp layers.')
parser.add_argument('--no_adj', dest='no_adj', action='store_true', help="Zero out adjacency matrix for ablation.")

parser.add_argument('--no-rnn', dest='rnn', action='store_false', help='Do not use RNN layer.')
parser.add_argument('--rnn_hidden', type=int, default=300, help='RNN hidden state size.')
parser.add_argument('--rnn_layers', type=int, default=1, help='Number of RNN layers.')
parser.add_argument('--rnn_dropout', type=float, default=0.5, help='RNN dropout rate.')

parser.add_argument('--lr', type=float, default=0.7, help='Applies to sgd and adagrad.')
parser.add_argument('--lr_decay', type=float, default=0.9, help='Learning rate decay rate.')
parser.add_argument('--decay_epoch', type=int, default=5, help='Decay learning rate after this epoch.')
parser.add_argument('--optim', choices=['sgd', 'adagrad', 'adam', 'adamax'], default='sgd', help='Optimizer: sgd, adagrad, adam or adamax.')
parser.add_argument('--num_epoch', type=int, default=100, help='Number of total training epochs.')
parser.add_argument('--batch_size', type=int, default=50, help='Training batch size.')
parser.add_argument('--max_grad_norm', type=float, default=5.0, help='Gradient clipping.')
parser.add_argument('--log_step', type=int, default=20, help='Print log every k steps.')
parser.add_argument('--log', type=str, default='logs.txt', help='Write training log to file.')
parser.add_argument('--save_epoch', type=int, default=100, help='Save model checkpoints every k epochs.')
parser.add_argument('--save_dir', type=str, default='./saved_models', help='Root dir for saving models.')
parser.add_argument('--id', type=str, default='00', help='Model ID under which to save models.')
parser.add_argument('--info', type=str, default='', help='Optional info for the experiment.')

parser.add_argument('--seed', type=int, default=0)
parser.add_argument('--cuda', type=bool, default=torch.cuda.is_available())
parser.add_argument('--cpu', action='store_true', help='Ignore CUDA.')

parser.add_argument('--load', dest='load', action='store_true', help='Load pretrained model.')
parser.add_argument('--model_file', type=str, help='Filename of the pretrained model.')

args = parser.parse_args()

torch.manual_seed(args.seed)
np.random.seed(args.seed)
random.seed(1234)
if args.cpu:
    args.cuda = False
elif args.cuda:
    torch.cuda.manual_seed(args.seed)
init_time = time.time()

# make opt
opt = vars(args)
label2id = constant.LABEL_TO_ID
opt['num_class'] = len(label2id)

# load vocab
vocab_file = opt['vocab_dir'] + '/vocab.pkl'
vocab = Vocab(vocab_file, load=True)
opt['vocab_size'] = vocab.size
emb_file = opt['vocab_dir'] + '/embedding.npy'
emb_matrix = np.load(emb_file)
assert emb_matrix.shape[0] == vocab.size
assert emb_matrix.shape[1] == opt['emb_dim']

# load data
print("Loading data from {} with batch size {}...".format(opt['data_dir'], opt['batch_size']))
train_batch = DataLoader(opt['data_dir'] + '/train.json', opt['batch_size'], opt, vocab, evaluation=False)
dev_batch = DataLoader(opt['data_dir'] + '/dev.json', opt['batch_size'], opt, vocab, evaluation=True)

model_id = opt['id'] if len(opt['id']) > 1 else '0' + opt['id']
model_save_dir = opt['save_dir'] + '/' + model_id
opt['model_save_dir'] = model_save_dir
helper.ensure_dir(model_save_dir, verbose=True)

# save config
helper.save_config(opt, model_save_dir + '/config.json', verbose=True)
vocab.save(model_save_dir + '/vocab.pkl')
file_logger = helper.FileLogger(model_save_dir + '/' + opt['log'], header="# epoch\ttrain_loss\tdev_loss\tdev_score\tbest_dev_score")

# print model info
helper.print_config(opt)

# model
if not opt['load']:
    trainer = GCNTrainer(opt, emb_matrix=emb_matrix)
else:
    # load pretrained model
    model_file = opt['model_file']
    print("Loading model from {}".format(model_file))
    model_opt = torch_utils.load_config(model_file)
    model_opt['optim'] = opt['optim']
    trainer = GCNTrainer(model_opt)
    trainer.load(model_file)

id2label = dict([(v,k) for k,v in label2id.items()])
dev_score_history = []
current_lr = opt['lr']

global_step = 0
global_start_time = time.time()
format_str = '{}: step {}/{} (epoch {}/{}), loss = {:.6f} ({:.3f} sec/batch), lr: {:.6f}'
max_steps = len(train_batch) * opt['num_epoch']

# start training
for epoch in range(1, opt['num_epoch']+1):
    train_loss = 0
    for i, batch in enumerate(train_batch):
        start_time = time.time()
        global_step += 1
        loss = trainer.update(batch)
        train_loss += loss
        if global_step % opt['log_step'] == 0:
            duration = time.time() - start_time
            print(format_str.format(datetime.now(), global_step, max_steps, epoch,\
                    opt['num_epoch'], loss, duration, current_lr))

    # eval on dev
    print("Evaluating on dev set...")
    predictions = []
    dev_loss = 0
    for i, batch in enumerate(dev_batch):
        preds, _, loss = trainer.predict(batch)
        predictions += preds
        dev_loss += loss
    predictions = [id2label[p] for p in predictions]
    train_loss = train_loss / train_batch.num_examples * opt['batch_size'] # avg loss per batch
    dev_loss = dev_loss / dev_batch.num_examples * opt['batch_size']

    dev_p, dev_r, dev_f1 = scorer.score(dev_batch.gold(), predictions)
    print("epoch {}: train_loss = {:.6f}, dev_loss = {:.6f}, dev_f1 = {:.4f}".format(epoch,\
        train_loss, dev_loss, dev_f1))
    dev_score = dev_f1
    file_logger.log("{}\t{:.6f}\t{:.6f}\t{:.4f}\t{:.4f}".format(epoch, train_loss, dev_loss, dev_score, max([dev_score] + dev_score_history)))

    # save
    model_file = model_save_dir + '/checkpoint_epoch_{}.pt'.format(epoch)
    trainer.save(model_file, epoch)
    if epoch == 1 or dev_score > max(dev_score_history):
        copyfile(model_file, model_save_dir + '/best_model.pt')
        print("new best model saved.")
        file_logger.log("new best model saved at epoch {}: {:.2f}\t{:.2f}\t{:.2f}"\
            .format(epoch, dev_p*100, dev_r*100, dev_score*100))
    if epoch % opt['save_epoch'] != 0:
        os.remove(model_file)

    # lr schedule
    if len(dev_score_history) > opt['decay_epoch'] and dev_score <= dev_score_history[-1] and \
            opt['optim'] in ['sgd', 'adagrad', 'adadelta']:
        current_lr *= opt['lr_decay']
        trainer.update_lr(current_lr)

    dev_score_history += [dev_score]
    print("")

print("Training ended with {} epochs.".format(epoch))