train.py

# pytorch 0.4.1 , cuda 9.0
import time
import torch.backends.cudnn as cudnn
import torch.optim
import torch.utils.data
import torchvision.transforms as transforms
from torch import nn
from torch.nn.utils.rnn import pack_padded_sequence
from models import DecoderWithAttention
from datasets import *
from utils import *
from nltk.translate.bleu_score import corpus_bleu

# Data parameters
# data_folder = 'final_dataset'  # folder with data files saved by create_input_files.py
data_folder = "/home/iplab/final_dataset"
data_name = 'coco_5_cap_per_img_5_min_word_freq'  # base name shared by data files

# Model parameters
emb_dim = 1024  # dimension of word embeddings
attention_dim = 1024  # dimension of attention linear layers
decoder_dim = 1024  # dimension of decoder RNN
topic_dim = 80
attribute_dim = 999
dropout = 0.5
clip = 0.25
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")  # sets device for model and PyTorch tensors


print(device)
cudnn.benchmark = True  # set to true only if inputs to model are fixed size; otherwise lot of computational overhead

# Training parameters
decoder_lr = 1e-4
start_epoch = 0
epochs = 100  # number of epochs to train for (if early stopping is not triggered)
epochs_since_improvement = 0  # keeps track of number of epochs since there's been an improvement in validation BLEU
batch_size = 16 #100
workers = 1  # for data-loading; right now, only 1 works with h5py
best_bleu4 = 0.  # BLEU-4 score right now
print_freq = 100  # print training/validation stats every __ batches
#checkpoint = None  # path to checkpoint, None if none
checkpoint ="pre_BEST_23checkpoint_coco_5_cap_per_img_5_min_word_freq.pth.tar"
# checkpoint = './pretrained_model/BEST_34checkpoint_coco_5_cap_per_img_5_min_word_freq.pth.tar'
#checkpoint = 'pre_NOT_BEST_36checkpoint_coco_5_cap_per_img_5_min_word_freq.pth.tar'
#checkpoint='./best_models/_BEST_35checkpoint_coco_5_cap_per_img_5_min_word_freq.pth.tar'
#checkpoint = "pre_NOT_BEST_35checkpoint_coco_5_cap_per_img_5_min_word_freq.pth.tar"
def main():
    """
    Training and validation.
    """
    global best_bleu4, epochs_since_improvement, checkpoint, start_epoch,data_name, word_map

    # Read word map
    word_map_file = os.path.join(data_folder, 'WORDMAP_' + data_name + '.json')
    with open(word_map_file, 'r') as j:
        word_map = json.load(j)

    # Initialize / load checkpoint
    if checkpoint is None:

        decoder = DecoderWithAttention(attention_dim=attention_dim,
                                       embed_dim=emb_dim,
                                       decoder_dim=decoder_dim,
                                       topic_dim=topic_dim,
                                       attribute_dim=attribute_dim,
                                       vocab_size=len(word_map),
                                       dropout=dropout)

        decoder_optimizer = torch.optim.Adamax(params=filter(lambda p: p.requires_grad, decoder.parameters()), lr=decoder_lr ,weight_decay=1e-8)

    else:




        checkpoint = torch.load(checkpoint)
        start_epoch = checkpoint['epoch'] + 1
        epochs_since_improvement = checkpoint['epochs_since_improvement']
        best_bleu4 = checkpoint['bleu-4']
        decoder = checkpoint['decoder']
        decoder_optimizer = checkpoint['decoder_optimizer']
       
        

        

        # for param_group in decoder_optimizer.param_groups:
        #     param_group['lr'] = param_group['lr'] * 0.001
    # Move to GPU, if available
    decoder = decoder.to(device)

    # Loss functions
    criterion_ce = nn.CrossEntropyLoss().to(device)
    criterion_dis = nn.MultiLabelMarginLoss().to(device)

    # Custom dataloaders
    train_loader = torch.utils.data.DataLoader(
        CaptionDataset(data_folder, data_name, 'TRAIN'),
        batch_size=batch_size, shuffle=True, num_workers=workers, pin_memory=True)
    val_loader = torch.utils.data.DataLoader(
        CaptionDataset(data_folder, data_name, 'VAL'),
        batch_size=batch_size, shuffle=True, num_workers=workers, pin_memory=True)

    # Epochs
    for epoch in range(start_epoch, epochs):

        # Decay learning rate if there is no improvement for 8 consecutive epochs, and terminate training after 20
        if epochs_since_improvement == 20:
            break
        # if epochs_since_improvement > 0 and epochs_since_improvement % 8 == 0:
        #     adjust_learning_rate(decoder_optimizer, 0.8)


        # for param_group in decoder_optimizer.param_groups:
        #     param_group['lr'] = 1e-5


    
        # One epoch's training
        train(train_loader=train_loader,
              decoder=decoder,
              criterion_ce = criterion_ce,
              criterion_dis=criterion_dis,
              decoder_optimizer=decoder_optimizer,
              epoch=epoch)

        # One epoch's validation
        recent_bleu4 = validate(val_loader=val_loader,
                                decoder=decoder,
                                criterion_ce=criterion_ce,
                                criterion_dis=criterion_dis)

        # Check if there was an improvement
        is_best = recent_bleu4 > best_bleu4
        best_bleu4 = max(recent_bleu4, best_bleu4)
        if not is_best:
            epochs_since_improvement += 1
            print("\nEpochs since last improvement: %d\n" % (epochs_since_improvement,))
        else:
            epochs_since_improvement = 0

        # Save checkpoint
        save_checkpoint(data_name, epoch, epochs_since_improvement, decoder,decoder_optimizer, recent_bleu4, is_best)


def train(train_loader, decoder, criterion_ce, criterion_dis, decoder_optimizer, epoch):
    """
    Performs one epoch's training.
    :param train_loader: DataLoader for training data
    :param decoder: decoder model
    :param criterion_ce: cross entropy loss layer
    :param criterion_dis : discriminative loss layer
    :param decoder_optimizer: optimizer to update decoder's weights
    :param epoch: epoch number
    """

    decoder.train()  # train mode (dropout and batchnorm is used)

    batch_time = AverageMeter()  # forward prop. + back prop. time
    data_time = AverageMeter()  # data loading time
    losses = AverageMeter()  # loss (per word decoded)
    top5accs = AverageMeter()  # top5 accuracy

    start = time.time()

    print("Decoder Parameters: ",decoder.parameters)
    #print(filter(lambda p: p.requires_grad, decoder.parameters()))
    #print(decoder.attention.weight.grad)

    # Batches
    for i, (imgs, caps, caplens, topic, attribute) in enumerate(train_loader):
        data_time.update(time.time() - start)

        # Move to GPU, if available
        imgs = imgs.to(device)
        caps = caps.to(device)
        caplens = caplens.to(device)
        topic = topic.to(device)
        attribute = attribute.to(device)

        # Forward prop.
        scores, scores_d,caps_sorted, decode_lengths, sort_ind = decoder(imgs, caps, caplens, topic, attribute)
        
        #Max-pooling across predicted words across time steps for discriminative supervision
        scores_d = scores_d.max(1)[0]

        # Since we decoded starting with <start>, the targets are all words after <start>, up to <end>
        targets = caps_sorted[:, 1:]
        targets_d = torch.zeros(scores_d.size(0),scores_d.size(1)).to(device)
        targets_d.fill_(-1)

        for length in decode_lengths:
            targets_d[:,:length-1] = targets[:,:length-1]

        # Remove timesteps that we didn't decode at, or are pads
        # pack_padded_sequence is an easy trick to do this
        scores, _ = pack_padded_sequence(scores, decode_lengths, batch_first=True)
        targets, _ = pack_padded_sequence(targets, decode_lengths, batch_first=True)

        # Calculate loss
        loss_d = criterion_dis(scores_d,targets_d.long())
        loss_g = criterion_ce(scores, targets)
        loss = loss_g + (10 * loss_d)
        
        # Back prop.
        decoder_optimizer.zero_grad()
        loss.backward()
    
        # Clip gradients when they are getting too large
        torch.nn.utils.clip_grad_norm_(filter(lambda p: p.requires_grad, decoder.parameters()), clip)
        #torch.nn.utils.clip_grad_norm_(decoder.parameters(), 0.25)

        # Update weights
        decoder_optimizer.step()

        # Keep track of metrics
        top5 = accuracy(scores, targets, 5)
        losses.update(loss.item(), sum(decode_lengths))
        top5accs.update(top5, sum(decode_lengths))
        batch_time.update(time.time() - start)

        start = time.time()

        # Print status
        if i % print_freq == 0:
            print('Epoch: [{0}][{1}/{2}]\t'
                  'Batch Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
                  'Data Load Time {data_time.val:.3f} ({data_time.avg:.3f})\t'
                  'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
                  'Top-5 Accuracy {top5.val:.3f} ({top5.avg:.3f})'.format(epoch, i, len(train_loader),
                                                                          batch_time=batch_time,
                                                                          data_time=data_time, loss=losses,
                                                                          top5=top5accs))


def validate(val_loader, decoder, criterion_ce, criterion_dis):
    """
    Performs one epoch's validation.
    :param val_loader: DataLoader for validation data.
    :param decoder: decoder model
    :param criterion_ce: cross entropy loss layer
    :param criterion_dis : discriminative loss layer
    :return: BLEU-4 score
    """
    decoder.eval()  # eval mode (no dropout or batchnorm)

    batch_time = AverageMeter()
    losses = AverageMeter()
    top5accs = AverageMeter()

    start = time.time()

    references = list()  # references (true captions) for calculating BLEU-4 score
    hypotheses = list()  # hypotheses (predictions)

    # Batches
    with torch.no_grad(): 
        for i, (imgs, caps, caplens,allcaps, topic, attribute) in enumerate(val_loader):

            # Move to device, if available
            imgs = imgs.to(device)
            caps = caps.to(device)
            caplens = caplens.to(device)
            topic = topic.to(device)
            attribute = attribute.to(device)


            scores, scores_d, caps_sorted, decode_lengths, sort_ind = decoder(imgs, caps, caplens, topic,attribute)
            
            #Max-pooling across predicted words across time steps for discriminative supervision
            scores_d = scores_d.max(1)[0]

            # Since we decoded starting with <start>, the targets are all words after <start>, up to <end>
            targets = caps_sorted[:, 1:]
            targets_d = torch.zeros(scores_d.size(0),scores_d.size(1)).to(device)
            targets_d.fill_(-1)

            for length in decode_lengths:
                targets_d[:,:length-1] = targets[:,:length-1]

            # Remove timesteps that we didn't decode at, or are pads
            # pack_padded_sequence is an easy trick to do this
            scores_copy = scores.clone()
            scores, _ = pack_padded_sequence(scores, decode_lengths, batch_first=True)
            targets, _ = pack_padded_sequence(targets, decode_lengths, batch_first=True)

            # Calculate loss
            loss_d = criterion_dis(scores_d,targets_d.long())
            loss_g = criterion_ce(scores, targets)
            loss = loss_g + (10 * loss_d)

            # Keep track of metrics
            losses.update(loss.item(), sum(decode_lengths))
            top5 = accuracy(scores, targets, 5)
            top5accs.update(top5, sum(decode_lengths))
            batch_time.update(time.time() - start)

            start = time.time()

            if i % print_freq == 0:
                print('Validation: [{0}/{1}]\t'
                      'Batch Time {batch_time.val:.3f} ({batch_time.avg:.3f})\t'
                      'Loss {loss.val:.4f} ({loss.avg:.4f})\t'
                      'Top-5 Accuracy {top5.val:.3f} ({top5.avg:.3f})\t'.format(i, len(val_loader), batch_time=batch_time,
                                                                                loss=losses, top5=top5accs))

            # Store references (true captions), and hypothesis (prediction) for each image
            # If for n images, we have n hypotheses, and references a, b, c... for each image, we need -
            # references = [[ref1a, ref1b, ref1c], [ref2a, ref2b], ...], hypotheses = [hyp1, hyp2, ...]

            # References
            allcaps = allcaps[sort_ind]  # because images were sorted in the decoder
            for j in range(allcaps.shape[0]):
                img_caps = allcaps[j].tolist()
                img_captions = list(
                    map(lambda c: [w for w in c if w not in {word_map['<start>'], word_map['<pad>']}],
                        img_caps))  # remove <start> and pads
                references.append(img_captions)

            # Hypotheses
            _, preds = torch.max(scores_copy, dim=2)
            preds = preds.tolist()
            temp_preds = list()
            for j, p in enumerate(preds):
                temp_preds.append(preds[j][:decode_lengths[j]])  # remove pads
            preds = temp_preds
            hypotheses.extend(preds)

            assert len(references) == len(hypotheses)

    # Calculate BLEU-4 scores
    bleu4 = corpus_bleu(references, hypotheses)
    bleu4 = round(bleu4,4)

    print(
        '\n * LOSS - {loss.avg:.3f}, TOP-5 ACCURACY - {top5.avg:.3f}, BLEU-4 - {bleu}\n'.format(
            loss=losses,
            top5=top5accs,
            bleu=bleu4))

    return bleu4


if __name__ == '__main__':
    main()