train.py

"""Train Real NVP on MNIST.

Train script adapted from: https://github.com/kuangliu/pytorch-cifar/
"""


import argparse
import os
import torch
import torch.optim as optim
import torch.backends.cudnn as cudnn
import torch.utils.data as data
import torchvision
import torchvision.transforms as transforms
from torch.utils.tensorboard import SummaryWriter
import util
import sys # for my favorite debugging method: print(..) ; sys.exit()

from models import RealNVP, RealNVPLoss
from tqdm import tqdm


IN_CHANNELS = 1 # Standard: IN_CHANNELS = 3 (for CIFAR10)
MID_CHANNELS = 64 # Standard
NUM_BLOCKS = 3 # Standard: NUM_BLOCKS = 8 // # Corresponds to NR of resnet blocks in scale and translation networks
NUM_SCALES = 1 # Standard: NUM_SCALES=2 ; # with NUM_SCALES = 1 we just have 4 coupling layers with checkerboard masking and w/o channel wise masking
NUM_EPOCHS = 25
NUM_SAMPLES_TRAIN = 200 # number samples per epoch in train time. There are 60k images in MNIST
NUM_SAMPLES_TEST = 200 # number samples per epoch in test time evaluation # 10k test samples in total
# BATCH_SIZE = 64 # Standard: BATCH_SIZE = 64
BATCH_SIZE = 50 # Standard: BATCH_SIZE = 64
MODEL_PATH = 'model_checkpoints/model_test.pth.tar'
RESOLUTION = [28, 28]

def main(args):

    device = 'cuda' if torch.cuda.is_available() and len(args.gpu_ids) > 0 else 'cpu'
    start_epoch = 0

    # Note: No normalization applied, since RealNVP expects inputs in (0, 1).
    transform_train = transforms.Compose([
        # transforms.RandomHorizontalFlip(),
        transforms.Resize(RESOLUTION),
        transforms.ToTensor()    
    ])

    transform_test = transforms.Compose([
        transforms.ToTensor()
    ])

    # if we run the script via floyd, the data is mounted in the folder /floyd/input in the cloud
    if args.floyd:
        data_set_path = '/floyd/input'
    else:
        data_set_path = 'data'
        
    # The original code used CIFAR10
    # trainset = torchvision.datasets.CIFAR10(root='data', train=True, download=True, transform=transform_train)
    trainset = torchvision.datasets.MNIST(root=data_set_path, train=True, download=True, transform=transform_train)
    trainloader = data.DataLoader(trainset, batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers)

    # The original code used CIFAR10
    # testset = torchvision.datasets.CIFAR10(root='data', train=False, download=True, transform=transform_test)
    testset = torchvision.datasets.MNIST(root=data_set_path, train=False, download=True, transform=transform_test)
    testloader = data.DataLoader(testset, batch_size=args.batch_size, shuffle=False, num_workers=args.num_workers)
    # 10k samples in testloader

    print('Nr of samples in train set',len(trainloader.dataset))
    print('Nr of samples in test set',len(testloader.dataset))

    # Model
    print('Building model..')
    net = RealNVP(num_scales=NUM_SCALES, in_channels=IN_CHANNELS, mid_channels=MID_CHANNELS, num_blocks=NUM_BLOCKS)
    print(net)
    print(net.modules)
    import sys
    sys.exit()
    
    

    net = net.to(device)
    
    # NOTE: DataParallel was usally within the if == device clause. A model trained with DataParallel has module.* as naming convention for the modules.
    # e.g. flows.in_couplings.0.st_net.in_norm.bias --> becomes module.flows.in_couplings.0.st_net.in_norm.bias
    # This causes problems when model is trained on cuda but should be evaluated via CPU. 
    # Thus, we also add DataParallel if we are in CPU mode to obtain the same naming convetion.
    # TODO: Have to check if that hurts performance somehow
    net = torch.nn.DataParallel(net, args.gpu_ids)

    if device == 'cuda':
        # net = torch.nn.DataParallel(net, args.gpu_ids)
        cudnn.benchmark = args.benchmark

    if args.resume:
        # Load checkpoint.
        print('Resuming from checkpoint at ckpts/best.pth.tar...')
        assert os.path.isdir('ckpts'), 'Error: no checkpoint directory found!'
        checkpoint = torch.load('ckpts/best.pth.tar', map_location= torch.device(device))
        
        net.load_state_dict(checkpoint['net'])
        
        global best_loss
        best_loss = checkpoint['test_loss']
        start_epoch = checkpoint['epoch']

    

    loss_fn = RealNVPLoss()
    param_groups = util.get_param_groups(net, args.weight_decay, norm_suffix='weight_g')
    optimizer = optim.Adam(param_groups, lr=args.lr)

    writer = SummaryWriter()

    with tqdm(total = args.num_epochs, initial = start_epoch) as pbar:

        for epoch in range(start_epoch, start_epoch + args.num_epochs):
            pbar.set_description(f'Epoch {epoch}')

            loss, bpd = train(epoch, net, trainloader, device, optimizer, loss_fn, args.max_grad_norm)
            writer.add_scalar('train/loss', loss, epoch)
            writer.add_scalar('train/bpd', bpd, epoch)

            loss, bpd = test(epoch, net, testloader, device, loss_fn, args.num_samples)
            writer.add_scalar('test/loss', loss, epoch)
            writer.add_scalar('test/bpd', bpd, epoch)


def train(epoch, net, trainloader, device, optimizer, loss_fn, max_grad_norm):
    
    net.train()
    loss_meter = util.AverageMeter()

    with tqdm(total=NUM_SAMPLES_TRAIN) as progress_bar:

        progress_bar.set_description('Train')

        for idx, (x, _) in enumerate(trainloader):

            if (idx+1)*BATCH_SIZE >= NUM_SAMPLES_TRAIN: 
                break
            
            x = x.to(device)

            optimizer.zero_grad()
            z, sldj = net(x, reverse=False)
            
            loss = loss_fn(z, sldj)
            loss_meter.update(loss.item(), x.size(0))
            loss.backward()
            util.clip_grad_norm(optimizer, max_grad_norm)
            optimizer.step()

            progress_bar.set_postfix(loss=loss_meter.avg,
                                     bpd=util.bits_per_dim(x, loss_meter.avg))
            progress_bar.update(x.size(0))
        
            # print('bpd',util.bits_per_dim(x, loss_meter.avg))
            # print('loss',loss_meter.avg)

        # # save model after each epoch
        # torch.save(net.state_dict(), MODEL_PATH)

    return loss_meter.avg, util.bits_per_dim(x, loss_meter.avg)


def sample(net, batch_size, device):
    """Sample from RealNVP model.

    Args:
        net (torch.nn.DataParallel): The RealNVP model wrapped in DataParallel.
        batch_size (int): Number of samples to generate.
        device (torch.device): Device to use.
    """
    # for CIFAR10
    #z = torch.randn((batch_size, 3, 32, 32), dtype=torch.float32, device=device)
    # for MNIST
    z = torch.randn((batch_size, 1, 28, 28), dtype=torch.float32, device=device)
    x, _ = net(z, reverse=True)
    x = torch.sigmoid(x)

    return x


def test(epoch, net, testloader, device, loss_fn, num_samples):
    
    global best_loss
    net.eval()
    loss_meter = util.AverageMeter()

    with torch.no_grad():
        
        with tqdm(total=NUM_SAMPLES_TEST) as progress_bar:
            progress_bar.set_description('Test')
            
            for idx, (x, _) in enumerate(testloader):

                if (idx+1)*BATCH_SIZE >= NUM_SAMPLES_TEST: 
                    break

                x = x.to(device)
                z, sldj = net(x, reverse=False)         
                loss = loss_fn(z, sldj)
                loss_meter.update(loss.item(), x.size(0))
                progress_bar.set_postfix(loss=loss_meter.avg,
                                         bpd=util.bits_per_dim(x, loss_meter.avg))
                progress_bar.update(x.size(0))

    # Save checkpoint
    if loss_meter.avg < best_loss:
        print('Saving...')
        state = {
            'net': net.state_dict(),
            'test_loss': loss_meter.avg,
            'epoch': epoch,
        }
        os.makedirs('ckpts', exist_ok=True)
        torch.save(state, 'ckpts/best.pth.tar')
        best_loss = loss_meter.avg

    # Save samples and data
    images = sample(net, num_samples, device)
    os.makedirs('samples', exist_ok=True)
    images_concat = torchvision.utils.make_grid(images, nrow=int(num_samples ** 0.5), padding=2, pad_value=255)
    torchvision.utils.save_image(images_concat, 'samples/epoch_{}.png'.format(epoch))

    return loss_meter.avg, util.bits_per_dim(x, loss_meter.avg)    

if __name__ == '__main__':
    parser = argparse.ArgumentParser(description='RealNVP on MNIST')

    parser.add_argument('--batch_size', default=BATCH_SIZE, type=int, help='Batch size')
    parser.add_argument('--benchmark', action='store_true', help='Turn on CUDNN benchmarking')
    parser.add_argument('--gpu_ids', default='[0]', type=eval, help='IDs of GPUs to use')
    parser.add_argument('--lr', default=1e-3, type=float, help='Learning rate')
    parser.add_argument('--max_grad_norm', type=float, default=100., help='Max gradient norm for clipping')
    parser.add_argument('--num_epochs', default=NUM_EPOCHS, type=int, help='Number of epochs to train')
    parser.add_argument('--num_samples', default=64, type=int, help='Number of samples at test time')
    parser.add_argument('--num_workers', default=8, type=int, help='Number of data loader threads')
    parser.add_argument('--resume', '-r', action='store_true', help='Resume from checkpoint')
    parser.add_argument('--weight_decay', default=5e-5, type=float,
                        help='L2 regularization (only applied to the weight norm scale factors)')
    parser.add_argument('--floyd', action='store_true', help='Other data set path if we use floyd')
                                            
    best_loss = 100000

    main(parser.parse_args())