train.py

import argparse
import torch
import torch.nn as nn
import torch.optim as optim
from torch.nn.utils import clip_grad_norm_
import numpy as np
import os
import sys
import time
import json
import uuid
import logging
import pickle
from datetime import datetime

from dataloader import DataLoader
from model import CaptionModel, CrossEntropyCriterion, RewardCriterion

import utils
import opts

import sys
sys.path.append("cider")
from pyciderevalcap.cider.cider import Cider
from pyciderevalcap.ciderD.ciderD import CiderD

sys.path.append('coco-caption')
from pycocoevalcap.bleu.bleu import Bleu
from pycocoevalcap.meteor.meteor import Meteor
from pycocoevalcap.rouge.rouge import Rouge

logger = logging.getLogger(__name__)


def language_eval(predictions, cocofmt_file, opt):
    logger.info('>>> Language evaluating ...')
    tmp_checkpoint_json = os.path.join(
        opt.model_file + str(uuid.uuid4()) + '.json')
    json.dump(predictions, open(tmp_checkpoint_json, 'w'))
    lang_stats = utils.language_eval(cocofmt_file, tmp_checkpoint_json)
    os.remove(tmp_checkpoint_json)
    return lang_stats


def train(model,
          criterion,
          optimizer,
          train_loader,
          val_loader,
          opt,
          rl_criterion=None):

    infos = {
        'iter': 0,
        'epoch': 0,
        'start_epoch': 0,
        'best_score': float('-inf'),
        'best_iter': 0,
        'best_epoch': opt.max_epochs
    }

    checkpoint_checked = False
    rl_training = False
    seq_per_img = train_loader.get_seq_per_img()
    infos_history = {}

    if os.path.exists(opt.start_from):
        if os.path.isdir(opt.start_from):
            # loading the same model file at a different experiment dir
            start_from_file = os.path.join(opt.start_from,
                                           os.path.basename(opt.model_file))
        else:
            start_from_file = opt.start_from
        logger.info('Loading state from: %s', start_from_file)
        checkpoint = torch.load(start_from_file)
        model.load_state_dict(checkpoint['model'])
        infos = checkpoint['infos']
        infos['start_epoch'] = infos['epoch']
        checkpoint_checked = True  # this epoch is already checked
    else:
        logger.info('No checkpoint found! Training from the scratch')

    if not os.path.exists(opt.model_file):
        # copy start_from model to new directory
        logger.info('>>> No model file found. Write a base checkpoint.')

        torch.save({
            'model': model.state_dict(),
            'infos': infos,
            'opt': opt
        }, opt.model_file)

        logger.info('Wrote checkpoint to: %s', opt.model_file)

    if opt.use_rl == 1 and opt.use_rl_after == 0:
        opt.use_rl_after = infos['epoch']
        opt.use_cst_after = infos['epoch']
        train_loader.set_current_epoch(infos['epoch'])

    while True:
        t_start = time.time()
        model.train()
        data = train_loader.get_batch()
        feats = [feat.cuda() for feat in data['feats']]
        labels = data['labels'].cuda()
        masks = data['masks'].cuda()

        # implement scheduled sampling
        opt.ss_prob = 0
        if opt.use_ss == 1 and infos['epoch'] >= opt.use_ss_after:
            annealing_prob = opt.ss_k / \
                (opt.ss_k + np.exp((infos['epoch'] - opt.use_ss_after) / opt.ss_k))
            opt.ss_prob = min(1 - annealing_prob, opt.ss_max_prob)
            model.set_ss_prob(opt.ss_prob)

        if opt.use_rl == 1 and infos['epoch'] >= opt.use_rl_after and not rl_training:
            logger.info('Using RL objective...')
            rl_training = True
            bcmr_scorer = {
                'Bleu_4': Bleu(),
                'CIDEr': CiderD(df=opt.train_cached_tokens),
                'METEOR': Meteor(),
                'ROUGE_L': Rouge()
            }[opt.eval_metric]

            #logger.info('loading gt refs: %s', train_loader.cocofmt_file)
            #gt_refs = utils.load_gt_refs(train_loader.cocofmt_file)

        mixer_from = opt.mixer_from
        if opt.use_mixer == 1 and rl_training:
            #annealing_mixer = opt.ss_k / \
            #    (opt.ss_k + np.exp((infos['epoch'] - opt.use_rl_after) / opt.ss_k))
            #annealing_mixer = int(round(annealing_mixer * opt.seq_length))

            # -1 for annealing
            if opt.mixer_from == -1:
                annealing_mixer = opt.seq_length - int(
                    np.ceil((infos['epoch'] - opt.use_rl_after + 1) /
                            opt.mixer_descrease_every))
                mixer_from = max(1, annealing_mixer)

            model.set_mixer_from(mixer_from)

        scb_captions = opt.scb_captions
        if opt.use_cst == 1 and rl_training:
            # if opt.use_cst == 1 and opt.ss_k == 0,
            # then do not using annealing, but the fixed scb_captions provided
            #annealing_robust = opt.ss_k / \
            #    (opt.ss_k + np.exp((infos['epoch'] - opt.use_rl_after) / opt.ss_k))
            #annealing_robust = int(round((1 - annealing_robust) * seq_per_img))

            # do not use robust before fully mixed
            # if opt.use_mixer == 1 and mixer_from > 1:
            #    opt.use_cst_after = infos['epoch']

            # if opt.scb_captions is -1, then use the annealing value,
            # otherwise, use the set value
            if opt.scb_captions == -1:
                annealing_robust = int(
                    np.ceil((infos['epoch'] - opt.use_cst_after + 1) /
                            opt.cst_increase_every))
                scb_captions = min(annealing_robust, seq_per_img - 1)

        optimizer.zero_grad()
        model.set_seq_per_img(seq_per_img)

        if rl_training:
            # sampling from model distribution
            # model_res, logprobs = model.sample(
            #    feats, {'sample_max': 0, 'expand_feat': opt.expand_feat, 'temperature': 1})

            # using mixer
            pred, model_res, logprobs = model(feats, labels)

            if opt.use_cst == 0:
                # greedy decoding baseline in SCST paper
                greedy_baseline, _ = model.sample(feats, {
                    'sample_max': 1,
                    'expand_feat': opt.expand_feat
                })

            if opt.use_cst == 1:
                bcmrscores = data['bcmrscores']
                reward, m_score, g_score = utils.get_cst_reward(
                    model_res.cpu().numpy(),
                    data['gts'],
                    bcmr_scorer,
                    bcmrscores=bcmrscores,
                    expand_feat=opt.expand_feat,
                    seq_per_img=train_loader.get_seq_per_img(),
                    scb_captions=scb_captions,
                    scb_baseline=opt.scb_baseline,
                    use_eos=opt.use_eos,
                    use_mixer=opt.use_mixer)
            else:
                # use greedy baseline by default, compute self-critical reward
                reward, m_score, g_score = utils.get_self_critical_reward(
                    model_res.cpu().numpy(),
                    greedy_baseline.cpu().numpy(),
                    data['gts'],
                    bcmr_scorer,
                    expand_feat=opt.expand_feat,
                    seq_per_img=train_loader.get_seq_per_img(),
                    use_eos=opt.use_eos)

            loss = rl_criterion(
                model_res,
                logprobs,
                torch.from_numpy(reward).float().cuda(),
            )

        else:
            pred = model(feats, labels)[0]
            loss = criterion(pred, labels[:, 1:], masks[:, 1:])

        loss.backward()
        clip_grad_norm_(model.parameters(), opt.grad_clip)
        optimizer.step()
        infos['TrainLoss'] = loss.item()
        infos['mixer_from'] = mixer_from
        infos['scb_captions'] = scb_captions

        if infos['iter'] % opt.print_log_interval == 0:
            elapsed_time = time.time() - t_start

            log_info = [('Epoch', infos['epoch']), ('Iter', infos['iter']),
                        ('Loss', infos['TrainLoss'])]

            if rl_training:
                log_info += [('Reward',
                              np.mean(reward[:, 0])), ('{} (m)'.format(
                                  opt.eval_metric), m_score), ('{} (b)'.format(
                                      opt.eval_metric), g_score)]

            if opt.use_ss == 1:
                log_info += [('ss_prob', opt.ss_prob)]

            if opt.use_mixer == 1:
                log_info += [('mixer_from', mixer_from)]

            if opt.use_cst == 1:
                log_info += [('scb_captions', scb_captions)]

            log_info += [('Time', elapsed_time)]
            logger.info('%s', '\t'.join(
                ['{}: {}'.format(k, v) for (k, v) in log_info]))

        infos['iter'] += 1

        if infos['epoch'] < train_loader.get_current_epoch():
            infos['epoch'] = train_loader.get_current_epoch()
            checkpoint_checked = False
            learning_rate = utils.adjust_learning_rate(
                opt, optimizer, infos['epoch'] - infos['start_epoch'])
            logger.info('===> Learning rate: %f: ', learning_rate)

        if (infos['epoch'] >= opt.save_checkpoint_from and
                infos['epoch'] % opt.save_checkpoint_every == 0 and
                not checkpoint_checked):
            # evaluate the validation performance
            results = validate(model, criterion, val_loader, opt)
            logger.info('Validation output: %s',
                        json.dumps(results['scores'], indent=4, sort_keys=True))
            infos.update(results['scores'])

            check_model(model, opt, infos, infos_history)
            checkpoint_checked = True

        if (infos['epoch'] >= opt.max_epochs or
                infos['epoch'] - infos['best_epoch'] > opt.max_patience):
            logger.info('>>> Terminating...')
            break

    return infos


def validate(model, criterion, loader, opt):

    model.eval()
    loader.reset()

    num_videos = loader.get_num_videos()
    batch_size = loader.get_batch_size()
    num_iters = int(np.ceil(num_videos / batch_size))
    last_batch_size = num_videos % batch_size
    seq_per_img = loader.get_seq_per_img()
    model.set_seq_per_img(seq_per_img)

    loss_sum = 0
    logger.info('#num_iters: %d, batch_size: %d, seg_per_image: %d', num_iters,
                batch_size, seq_per_img)
    predictions = []
    gt_avglogps = []
    test_avglogps = []
    for ii in range(num_iters):
        data = loader.get_batch()
        feats = data['feats']
        if loader.has_label:
            labels = data['labels']
            masks = data['masks']

        if ii == (num_iters - 1) and last_batch_size > 0:
            feats = [f[:last_batch_size] for f in feats]
            if loader.has_label:
                labels = labels[:last_batch_size *
                                seq_per_img]  # labels shape is DxN
                masks = masks[:last_batch_size * seq_per_img]

        feats = [feat.cuda() for feat in feats]
        if loader.has_label:
            labels = labels.cuda()
            masks = masks.cuda()

        if loader.has_label:
            with torch.no_grad():
                pred, gt_seq, gt_logseq = model(feats, labels)
            gt_seq = gt_seq.cpu().numpy()
            gt_logseq = gt_logseq.cpu().numpy()
            if opt.output_logp == 1:
                gt_avglogp = utils.compute_avglogp(gt_seq, gt_logseq)
                gt_avglogps.extend(gt_avglogp)


            loss = criterion(pred, labels[:, 1:], masks[:, 1:])
            loss_sum += loss.item()

        with torch.no_grad():
            seq, logseq = model.sample(feats, {'beam_size': opt.beam_size})
        seq = seq.cpu().numpy()
        logseq = logseq.cpu().numpy()
        sents = utils.decode_sequence(opt.vocab, seq)
        if opt.output_logp == 1:
            test_avglogp = utils.compute_avglogp(seq, logseq)
            test_avglogps.extend(test_avglogp)

        for jj, sent in enumerate(sents):
            if opt.output_logp == 1:
                entry = {
                    'image_id': data['ids'][jj],
                    'caption': sent,
                    'avglogp': test_avglogp[jj]
                }
            else:
                entry = {'image_id': data['ids'][jj], 'caption': sent}
            predictions.append(entry)
            logger.debug('[%d] video %s: %s' % (jj, entry['image_id'],
                                                entry['caption']))

    loss = round(loss_sum / num_iters, 3)
    results = {}
    lang_stats = {}

    if opt.language_eval == 1 and loader.has_label:
        lang_stats = language_eval(predictions, loader.cocofmt_file, opt)

    results['predictions'] = predictions
    results['scores'] = {'Loss': -loss}
    results['scores'].update(lang_stats)

    if opt.output_logp == 1:
        avglogp = sum(test_avglogps) / len(test_avglogps)
        results['scores'].update({'avglogp': avglogp})

        gt_avglogps = np.array(gt_avglogps).reshape(-1, seq_per_img)
        assert num_videos == gt_avglogps.shape[0]

        gt_avglogps_file = opt.model_file.replace('.pth', '_gt_avglogps.pkl', 1)
        pickle.dump(
            gt_avglogps,
            open(gt_avglogps_file, 'wb'),
            protocol=pickle.HIGHEST_PROTOCOL)

        logger.info('Wrote GT logp to: %s', gt_avglogps_file)

    return results


def test(model, criterion, loader, opt):
    results = validate(model, criterion, loader, opt)
    logger.info('Test output: %s', json.dumps(results['scores'], indent=4))

    json.dump(results, open(opt.result_file, 'w'))
    logger.info('Wrote output caption to: %s ', opt.result_file)


def check_model(model, opt, infos, infos_history):
    if opt.eval_metric == 'MSRVTT':
        current_score = infos['Bleu_4'] + \
            infos['METEOR'] + infos['ROUGE_L'] + infos['CIDEr']
    else:
        current_score = infos[opt.eval_metric]

    # write the full model checkpoint as well if we did better than ever
    if current_score >= infos['best_score']:
        infos['best_score'] = current_score
        infos['best_iter'] = infos['iter']
        infos['best_epoch'] = infos['epoch']

        logger.info('>>> Found new best [%s] score: %f, at iter: %d, epoch %d',
                    opt.eval_metric, current_score, infos['iter'],
                    infos['epoch'])

        torch.save({
            'model': model.state_dict(),
            'infos': infos,
            'opt': opt
        }, opt.model_file)
        logger.info('Wrote checkpoint to: %s', opt.model_file)

    else:
        logger.info('>>> Current best [%s] score: %f, at iter %d, epoch %d',
                    opt.eval_metric, infos['best_score'], infos['best_iter'],
                    infos['best_epoch'])

    infos_history[infos['epoch']] = infos.copy()
    with open(opt.history_file, 'w') as of:
        json.dump(infos_history, of)
    logger.info('Updated history to: %s', opt.history_file)


if __name__ == '__main__':

    opt = opts.parse_opts()

    logging.basicConfig(
        level=getattr(logging, opt.loglevel.upper()),
        format='%(asctime)s:%(levelname)s: %(message)s')

    logger.info('Input arguments: %s',
                json.dumps(vars(opt), sort_keys=True, indent=4))

    # Set the random seed manually for reproducibility.
    np.random.seed(opt.seed)
    torch.manual_seed(opt.seed)
    torch.cuda.manual_seed(opt.seed)

    train_opt = {
        'label_h5': opt.train_label_h5,
        'batch_size': opt.batch_size,
        'feat_h5': opt.train_feat_h5,
        'cocofmt_file': opt.train_cocofmt_file,
        'bcmrscores_pkl': opt.train_bcmrscores_pkl,
        'eval_metric': opt.eval_metric,
        'seq_per_img': opt.train_seq_per_img,
        'num_chunks': opt.num_chunks,
        'mode': 'train'
    }

    val_opt = {
        'label_h5': opt.val_label_h5,
        'batch_size': opt.test_batch_size,
        'feat_h5': opt.val_feat_h5,
        'cocofmt_file': opt.val_cocofmt_file,
        'seq_per_img': opt.test_seq_per_img,
        'num_chunks': opt.num_chunks,
        'mode': 'test'
    }

    test_opt = {
        'label_h5': opt.test_label_h5,
        'batch_size': opt.test_batch_size,
        'feat_h5': opt.test_feat_h5,
        'cocofmt_file': opt.test_cocofmt_file,
        'seq_per_img': opt.test_seq_per_img,
        'num_chunks': opt.num_chunks,
        'mode': 'test'
    }

    train_loader = DataLoader(train_opt)
    val_loader = DataLoader(val_opt)
    test_loader = DataLoader(test_opt)

    opt.vocab = train_loader.get_vocab()
    opt.vocab_size = train_loader.get_vocab_size()
    opt.seq_length = train_loader.get_seq_length()
    opt.feat_dims = train_loader.get_feat_dims()
    opt.history_file = opt.model_file.replace('.pth', '_history.json', 1)

    logger.info('Building model...')
    model = CaptionModel(opt)

    xe_criterion = CrossEntropyCriterion()
    rl_criterion = RewardCriterion()

    model.cuda()
    xe_criterion.cuda()
    rl_criterion.cuda()

    logger.info('Start training...')
    start = datetime.now()

    optimizer = optim.Adam(model.parameters(), lr=opt.learning_rate)
    infos = train(
        model,
        xe_criterion,
        optimizer,
        train_loader,
        val_loader,
        opt,
        rl_criterion=rl_criterion)
    logger.info('Best val %s score: %f. Best iter: %d. Best epoch: %d',
                opt.eval_metric, infos['best_score'], infos['best_iter'],
                infos['best_epoch'])

    logger.info('Training time: %s', datetime.now() - start)

    if opt.result_file:
        logger.info('Start testing...')
        start = datetime.now()

        logger.info('Loading model: %s', opt.model_file)
        checkpoint = torch.load(opt.model_file)
        model.load_state_dict(checkpoint['model'])

        test(model, xe_criterion, test_loader, opt)
        logger.info('Testing time: %s', datetime.now() - start)
    train_loader.close()
    val_loader.close()
    test_loader.close()