train.py

import argparse
import math
import random
import os

import numpy as np
import torch
from torch import nn, autograd, optim
from torch.nn import functional as F
from torch.utils import data
import torch.distributed as dist
from torchvision import transforms, utils
from torch.utils.tensorboard import SummaryWriter
from tqdm import tqdm

import model
from dataset import MultiScaleDataset, ImageDataset
from calculate_fid import calculate_fid
from distributed import get_rank, synchronize, reduce_loss_dict
from tensor_transforms import convert_to_coord_format


def data_sampler(dataset, shuffle, distributed):
    if distributed:
        return data.distributed.DistributedSampler(dataset, shuffle=shuffle)

    if shuffle:
        return data.RandomSampler(dataset)

    else:
        return data.SequentialSampler(dataset)


def requires_grad(model, flag=True):
    for p in model.parameters():
        p.requires_grad = flag


def accumulate(model1, model2, decay=0.999):
    par1 = dict(model1.named_parameters())
    par2 = dict(model2.named_parameters())

    for k in par1.keys():
        par1[k].data.mul_(decay).add_(1 - decay, par2[k].data)


def sample_data(loader):
    while True:
        for batch in loader:
            yield batch


def d_logistic_loss(real_pred, fake_pred):
    real_loss = F.softplus(-real_pred)
    fake_loss = F.softplus(fake_pred)

    return real_loss.mean() + fake_loss.mean()


def d_r1_loss(real_pred, real_img):
    grad_real, = autograd.grad(
        outputs=real_pred.sum(), inputs=real_img, create_graph=True
    )
    grad_penalty = grad_real.pow(2).view(grad_real.shape[0], -1).sum(1).mean()

    return grad_penalty


def g_nonsaturating_loss(fake_pred):
    loss = F.softplus(-fake_pred).mean()

    return loss


def make_noise(batch, latent_dim, n_noise, device):
    if n_noise == 1:
        return torch.randn(batch, latent_dim, device=device)

    noises = torch.randn(n_noise, batch, latent_dim, device=device).unbind(0)

    return noises


def mixing_noise(batch, latent_dim, prob, device):
    if prob > 0 and random.random() < prob:
        return make_noise(batch, latent_dim, 2, device)

    else:
        return [make_noise(batch, latent_dim, 1, device)]


def train(args, loader, generator, discriminator, g_optim, d_optim, g_ema, device):
    loader = sample_data(loader)

    pbar = range(args.iter)

    if get_rank() == 0:
        pbar = tqdm(pbar, initial=args.start_iter, dynamic_ncols=True, smoothing=0.01)

    mean_path_length = 0

    r1_loss = torch.tensor(0.0, device=device)
    path_loss = torch.tensor(0.0, device=device)
    path_lengths = torch.tensor(0.0, device=device)
    mean_path_length_avg = 0
    loss_dict = {}

    if args.distributed:
        g_module = generator.module
        d_module = discriminator.module

    else:
        g_module = generator
        d_module = discriminator
        
    accum = 0.5 ** (32 / (10 * 1000))

    sample_z = torch.randn(args.n_sample, args.latent, device=device)

    for idx in pbar:
        i = idx + args.start_iter

        if i > args.iter:
            print('Done!')

            break

        data = next(loader)
        key = np.random.randint(n_scales)
        real_stack = data[key].to(device)

        real_img, converted = real_stack[:, :3], real_stack[:, 3:]

        requires_grad(generator, False)
        requires_grad(discriminator, True)

        noise = mixing_noise(args.batch, args.latent, args.mixing, device)

        fake_img, _ = generator(converted, noise)
        fake = fake_img if args.img2dis else torch.cat([fake_img, converted], 1)
        fake_pred = discriminator(fake, key)

        real = real_img if args.img2dis else real_stack
        real_pred = discriminator(real, key)
        d_loss = d_logistic_loss(real_pred, fake_pred)

        loss_dict['d'] = d_loss
        loss_dict['real_score'] = real_pred.mean()
        loss_dict['fake_score'] = fake_pred.mean()

        discriminator.zero_grad()
        d_loss.backward()
        d_optim.step()

        d_regularize = i % args.d_reg_every == 0

        if d_regularize:
            real.requires_grad = True
            real_pred = discriminator(real, key)
            r1_loss = d_r1_loss(real_pred, real)

            discriminator.zero_grad()
            (args.r1 / 2 * r1_loss * args.d_reg_every + 0 * real_pred[0]).backward()

            d_optim.step()

        loss_dict['r1'] = r1_loss

        requires_grad(generator, True)
        requires_grad(discriminator, False)

        noise = mixing_noise(args.batch, args.latent, args.mixing, device)

        fake_img, _ = generator(converted, noise)
        fake = fake_img if args.img2dis else torch.cat([fake_img, converted], 1)
        fake_pred = discriminator(fake, key)
        g_loss = g_nonsaturating_loss(fake_pred)

        loss_dict['g'] = g_loss

        generator.zero_grad()
        g_loss.backward()
        g_optim.step()

        loss_dict['path'] = path_loss
        loss_dict['path_length'] = path_lengths.mean()

        accumulate(g_ema, g_module, accum)

        loss_reduced = reduce_loss_dict(loss_dict)

        d_loss_val = loss_reduced['d'].mean().item()
        g_loss_val = loss_reduced['g'].mean().item()
        r1_val = loss_reduced['r1'].mean().item()
        path_loss_val = loss_reduced['path'].mean().item()
        real_score_val = loss_reduced['real_score'].mean().item()
        fake_score_val = loss_reduced['fake_score'].mean().item()
        path_length_val = loss_reduced['path_length'].mean().item()

        if get_rank() == 0:
            pbar.set_description(
                (
                    f'd: {d_loss_val:.4f}; g: {g_loss_val:.4f}; r1: {r1_val:.4f}; '
                    f'path: {path_loss_val:.4f}; mean path: {mean_path_length_avg:.4f}'
                )
            )

            if i % 100 == 0:
                writer.add_scalar("Generator", g_loss_val, i)
                writer.add_scalar("Discriminator", d_loss_val, i)
                writer.add_scalar("R1", r1_val, i)
                writer.add_scalar("Path Length Regularization", path_loss_val, i)
                writer.add_scalar("Mean Path Length", mean_path_length, i)
                writer.add_scalar("Real Score", real_score_val, i)
                writer.add_scalar("Fake Score", fake_score_val, i)
                writer.add_scalar("Path Length", path_length_val, i)

            if i % 500 == 0:
                with torch.no_grad():
                    g_ema.eval()
                    converted_full = convert_to_coord_format(sample_z.size(0), args.size, args.size, device,
                                                             integer_values=args.coords_integer_values)
                    if args.generate_by_one:
                        converted_full = convert_to_coord_format(1, args.size, args.size, device,
                                                                 integer_values=args.coords_integer_values)
                        samples = []
                        for sz in sample_z:
                            sample, _ = g_ema(converted_full, [sz.unsqueeze(0)])
                            samples.append(sample)
                        sample = torch.cat(samples, 0)
                    else:
                        sample, _ = g_ema(converted_full, [sample_z])

                    utils.save_image(
                        sample,
                        os.path.join(path, 'outputs', args.output_dir, 'images', f'{str(i).zfill(6)}.png'),
                        nrow=int(args.n_sample ** 0.5),
                        normalize=True,
                        range=(-1, 1),
                    )

                    if i == 0:
                        utils.save_image(
                            fake_img,
                            os.path.join(
                                path,
                                f'outputs/{args.output_dir}/images/fake_patch_{str(key)}_{str(i).zfill(6)}.png'),
                            nrow=int(fake_img.size(0) ** 0.5),
                            normalize=True,
                            range=(-1, 1),
                        )

                        utils.save_image(
                            real_img,
                            os.path.join(
                                path,
                                f'outputs/{args.output_dir}/images/real_patch_{str(key)}_{str(i).zfill(6)}.png'),
                            nrow=int(real_img.size(0) ** 0.5),
                            normalize=True,
                            range=(-1, 1),
                        )

            if i % args.save_checkpoint_frequency == 0:
                torch.save(
                    {
                        'g': g_module.state_dict(),
                        'd': d_module.state_dict(),
                        'g_ema': g_ema.state_dict(),
                        'g_optim': g_optim.state_dict(),
                        'd_optim': d_optim.state_dict(),
                    },
                    os.path.join(
                        path,
                        f'outputs/{args.output_dir}/checkpoints/{str(i).zfill(6)}.pt'),
                )
                if i > 0:
                    cur_metrics = calculate_fid(g_ema, fid_dataset=fid_dataset, bs=args.fid_batch, size=args.coords_size,
                                                num_batches=args.fid_samples//args.fid_batch, latent_size=args.latent,
                                                save_dir=args.path_fid, integer_values=args.coords_integer_values)
                    writer.add_scalar("fid", cur_metrics['frechet_inception_distance'], i)
                    print(i, "fid",  cur_metrics['frechet_inception_distance'])


if __name__ == '__main__':
    device = 'cuda'

    parser = argparse.ArgumentParser()

    parser.add_argument('path', type=str)
    parser.add_argument('--output_dir', type=str)
    parser.add_argument('--out_path', type=str, default='.')

    # fid
    parser.add_argument('--fid_samples', type=int, default=50000)
    parser.add_argument('--fid_batch', type=int, default=16)

    # training
    parser.add_argument('--iter', type=int, default=1200000)
    parser.add_argument('--n_sample', type=int, default=64)
    parser.add_argument('--generate_by_one', action='store_true')
    parser.add_argument('--ckpt', type=str, default=None)
    parser.add_argument('--lr', type=float, default=0.002)
    parser.add_argument('--local_rank', type=int, default=0)
    parser.add_argument('--save_checkpoint_frequency', type=int, default=20000)

    # dataset
    parser.add_argument('--batch', type=int, default=4)
    parser.add_argument('--num_workers', type=int, default=32)
    parser.add_argument('--to_crop', action='store_true')
    parser.add_argument('--crop', type=int, default=256)
    parser.add_argument('--coords_size', type=int, default=256)

    # Generator params
    parser.add_argument('--Generator', type=str, default='ModSIREN')
    parser.add_argument('--coords_integer_values', action='store_true')
    parser.add_argument('--size', type=int, default=256)
    parser.add_argument('--fc_dim', type=int, default=512)
    parser.add_argument('--latent', type=int, default=512)
    parser.add_argument('--activation', type=str, default=None)
    parser.add_argument('--channel_multiplier', type=int, default=2)
    parser.add_argument('--mixing', type=float, default=0.)
    parser.add_argument('--g_reg_every', type=int, default=4)

    # Discriminator params
    parser.add_argument('--Discriminator', type=str, default='Discriminator')
    parser.add_argument('--d_reg_every', type=int, default=16)
    parser.add_argument('--r1', type=float, default=10)
    parser.add_argument('--img2dis',  action='store_true')
    parser.add_argument('--n_first_layers', type=int, default=0)

    args = parser.parse_args()
    path = args.out_path

    Generator = getattr(model, args.Generator)
    print('Generator', Generator)
    Discriminator = getattr(model, args.Discriminator)
    print('Discriminator', Discriminator)

    os.makedirs(os.path.join(path, 'outputs', args.output_dir, 'images'), exist_ok=True)
    os.makedirs(os.path.join(path, 'outputs', args.output_dir, 'checkpoints'), exist_ok=True)
    args.logdir = os.path.join(path, 'tensorboard', args.output_dir)
    os.makedirs(args.logdir, exist_ok=True)
    args.path_fid = os.path.join(path, 'fid', args.output_dir)
    os.makedirs(args.path_fid, exist_ok=True)

    n_gpu = int(os.environ['WORLD_SIZE']) if 'WORLD_SIZE' in os.environ else 1
    args.distributed = n_gpu > 1

    if args.distributed:
        torch.cuda.set_device(args.local_rank)
        torch.distributed.init_process_group(backend='nccl', init_method='env://')
        synchronize()

    args.n_mlp = 8
    args.dis_input_size = 3 if args.img2dis else 5
    print('img2dis', args.img2dis, 'dis_input_size', args.dis_input_size)

    args.start_iter = 0
    n_scales = int(math.log(args.size//args.crop, 2)) + 1
    print('n_scales', n_scales)

    generator = Generator(size=args.size, hidden_size=args.fc_dim, style_dim=args.latent, n_mlp=args.n_mlp,
                          activation=args.activation, channel_multiplier=args.channel_multiplier,
                          ).to(device)

    print('generator N params', sum(p.numel() for p in generator.parameters() if p.requires_grad))
    discriminator = Discriminator(
        size=args.crop, channel_multiplier=args.channel_multiplier, n_scales=n_scales, input_size=args.dis_input_size,
        n_first_layers=args.n_first_layers,
    ).to(device)
    g_ema = Generator(size=args.size, hidden_size=args.fc_dim, style_dim=args.latent, n_mlp=args.n_mlp,
                      activation=args.activation, channel_multiplier=args.channel_multiplier,
                      ).to(device)
    g_ema.eval()
    accumulate(g_ema, generator, 0)

    g_reg_ratio = args.g_reg_every / (args.g_reg_every + 1)
    d_reg_ratio = args.d_reg_every / (args.d_reg_every + 1)

    g_optim = optim.Adam(
        generator.parameters(),
        lr=args.lr * g_reg_ratio,
        betas=(0 ** g_reg_ratio, 0.99 ** g_reg_ratio),
    )
    d_optim = optim.Adam(
        discriminator.parameters(),
        lr=args.lr * d_reg_ratio,
        betas=(0 ** d_reg_ratio, 0.99 ** d_reg_ratio),
    )

    if args.ckpt is not None:
        print('load model:', args.ckpt)

        ckpt = torch.load(args.ckpt)

        try:
            ckpt_name = os.path.basename(args.ckpt)
            args.start_iter = int(os.path.splitext(ckpt_name)[0])

        except ValueError:
            pass

        generator.load_state_dict(ckpt['g'])
        discriminator.load_state_dict(ckpt['d'])
        g_ema.load_state_dict(ckpt['g_ema'])

        g_optim.load_state_dict(ckpt['g_optim'])
        d_optim.load_state_dict(ckpt['d_optim'])

        del ckpt
        torch.cuda.empty_cache()

    if args.distributed:
        generator = nn.parallel.DistributedDataParallel(
            generator,
            device_ids=[args.local_rank],
            output_device=args.local_rank,
            broadcast_buffers=False,
        )

        discriminator = nn.parallel.DistributedDataParallel(
            discriminator,
            device_ids=[args.local_rank],
            output_device=args.local_rank,
            broadcast_buffers=False,
        )

    transform = transforms.Compose(
        [
            transforms.RandomHorizontalFlip(),
            transforms.ToTensor(),
            transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5), inplace=True),
        ]
    )
    transform_fid = transforms.Compose([
                                       transforms.ToTensor(),
                                       transforms.Lambda(lambda x: x.mul_(255.).byte())])

    dataset = MultiScaleDataset(args.path, transform=transform, resolution=args.coords_size, crop_size=args.crop,
                                integer_values=args.coords_integer_values, to_crop=args.to_crop)
    fid_dataset = ImageDataset(args.path, transform=transform_fid, resolution=args.coords_size, to_crop=args.to_crop)
    fid_dataset.length = args.fid_samples
    loader = data.DataLoader(
        dataset,
        batch_size=args.batch,
        sampler=data_sampler(dataset, shuffle=True, distributed=args.distributed),
        drop_last=True,
        num_workers=args.num_workers,
        pin_memory=True,
    )

    writer = SummaryWriter(log_dir=args.logdir)

    train(args, loader, generator, discriminator, g_optim, d_optim, g_ema, device)