train_gan.py

import argparse
import math
import random
import os
from gan_model import Generator, Discriminator
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 tqdm import tqdm

try:
    import wandb

except ImportError:
    wandb = None


from utils.gan_dataset import MultiResolutionDataset
from distributed import (
    get_rank,
    synchronize,
    reduce_loss_dict,
    reduce_sum,
    get_world_size,
)
from op import conv2d_gradfix
from utils.non_leaking import augment, AdaptiveAugment


# 返回 DataLoader 的采样idx
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_(par2[k].data, alpha=1 - decay)

# a=sample_data(loader), 每次调用next(a)迭代产生1个batch, 回到loader时重新进行DataLoader()
def sample_data(loader):
    while True:
        for batch in loader:
            yield batch

# 判别器损失函数，real_pred越大，fake_pred越小，loss越小
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()

# 判别器输出对输入的梯度惩罚
def d_r1_loss(real_pred, real_img):
    with conv2d_gradfix.no_weight_gradients():
        grad_real, = autograd.grad(
            outputs=real_pred.sum(), inputs=real_img, create_graph=True
        )
    grad_penalty = grad_real.pow(2).reshape(grad_real.shape[0], -1).sum(1).mean()

    return grad_penalty

# 生成器损失函数,fake_pred越大，损失越小
def g_nonsaturating_loss(fake_pred):
    loss = F.softplus(-fake_pred).mean()

    return loss

# 生成器 Path Length 正则项
def g_path_regularize(fake_img, latents, mean_path_length, decay=0.01):
    noise = torch.randn_like(fake_img) / math.sqrt(
        fake_img.shape[2] * fake_img.shape[3]
    )  # 防止分辨率对求和的影响，因此除以h*w
    grad, = autograd.grad(
        outputs=(fake_img * noise).sum(), inputs=latents, create_graph=True
    )  # latents = batch_size * 风格隐向量个数 * 512
    path_lengths = torch.sqrt(grad.pow(2).sum(2).mean(1))  # batch_size * path_length

    path_mean = mean_path_length + decay * (path_lengths.mean() - mean_path_length)  # 滑动平均计算平均path长度

    path_penalty = (path_lengths - path_mean).pow(2).mean()

    return path_penalty, path_mean.detach(), path_lengths

# n_noise==1, 返回 tensor=batch*512 噪声；n_noise==2, 返回 (tensor1,tensor2)
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

# 按照概率 prob 生成混合噪声
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)]

# 根据参数名对梯度清零
def set_grad_none(model, targets):
    for n, p in model.named_parameters():
        if n in targets:
            p.grad = None


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

    d_loss_val = 0
    r1_loss = torch.tensor(0.0, device=device)
    g_loss_val = 0
    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))
    ada_aug_p = args.augment_p if args.augment_p > 0 else 0.0
    r_t_stat = 0

    if args.augment and args.augment_p == 0:
        ada_augment = AdaptiveAugment(args.ada_target, args.ada_length, 8, device)

    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

        real_img = next(loader)
        real_img = real_img.to(device)  # load 真实图像

        requires_grad(generator, False)
        requires_grad(discriminator, True)  # 训练判别器

        noise = mixing_noise(args.batch, args.latent, args.mixing, device)  # 生成噪声，P=args.mixing为混合噪声，否则不混合
        fake_img, _ = generator(noise)  # 根据 z 生成假图像

        if args.augment:
            real_img_aug, _ = augment(real_img, ada_aug_p)
            fake_img, _ = augment(fake_img, ada_aug_p)

        else:
            real_img_aug = real_img

        fake_pred = discriminator(fake_img)
        real_pred = discriminator(real_img_aug)
        d_loss = d_logistic_loss(real_pred, fake_pred)  # 判别器基础损失

        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()

        if args.augment and args.augment_p == 0:
            ada_aug_p = ada_augment.tune(real_pred)
            r_t_stat = ada_augment.r_t_stat

        d_regularize = i % args.d_reg_every == 0   # i % args.d_reg_every == 0，d_regularize=True，非lazy正则启动

        if d_regularize:
            real_img.requires_grad = True

            if args.augment:
                real_img_aug, _ = augment(real_img, ada_aug_p)

            else:
                real_img_aug = real_img

            real_pred = discriminator(real_img_aug)
            r1_loss = d_r1_loss(real_pred, real_img)

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

            d_optim.step()

        loss_dict["r1"] = r1_loss

        # 开始训练生成器
        requires_grad(generator, True)
        requires_grad(discriminator, False)

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

        if args.augment:
            fake_img, _ = augment(fake_img, ada_aug_p)

        fake_pred = discriminator(fake_img)
        g_loss = g_nonsaturating_loss(fake_pred)

        loss_dict["g"] = g_loss

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

        g_regularize = i % args.g_reg_every == 0  # 生成器lazy正则

        if g_regularize:
            path_batch_size = max(1, args.batch // args.path_batch_shrink)
            noise = mixing_noise(path_batch_size, args.latent, args.mixing, device)
            fake_img, latents = generator(noise, return_latents=True)  # 返回w空间向量

            path_loss, mean_path_length, path_lengths = g_path_regularize(
                fake_img, latents, mean_path_length
            )  # 返回路径损失，指数滑动平均长度，本次路径长度

            generator.zero_grad()
            weighted_path_loss = args.path_regularize * args.g_reg_every * path_loss

            if args.path_batch_shrink:
                weighted_path_loss += 0 * fake_img[0, 0, 0, 0]

            weighted_path_loss.backward()

            g_optim.step()

            mean_path_length_avg = (
                    reduce_sum(mean_path_length).item() / get_world_size()
            )  # 求各进程的平均路径长度

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

        accumulate(g_ema, g_module, accum) # 更新 g_ema

        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}; "
                    f"augment: {ada_aug_p:.4f}"
                )
            )

            if wandb and args.wandb:
                wandb.log(
                    {
                        "Generator": g_loss_val,
                        "Discriminator": d_loss_val,
                        "Augment": ada_aug_p,
                        "Rt": r_t_stat,
                        "R1": r1_val,
                        "Path Length Regularization": path_loss_val,
                        "Mean Path Length": mean_path_length,
                        "Real Score": real_score_val,
                        "Fake Score": fake_score_val,
                        "Path Length": path_length_val,
                    }
                )

            # 生成图像
            if i % 5000 == 0:
                with torch.no_grad():
                    g_ema.eval()
                    sample, _ = g_ema([sample_z])
                    utils.save_image(
                        sample,
                        f"sample/{str(i).zfill(6)}.png",
                        nrow=int(args.n_sample ** 0.5),
                        normalize=True,
                        range=(-1, 1),
                    )

            # 按照字典存储checkpoint
            if i % 10000 == 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(),
                        "args": args,
                        "ada_aug_p": ada_aug_p,
                    },
                    f"checkpoint/{str(i).zfill(6)}.pt",
                )


if __name__ == "__main__":
    device = "cuda"

    parser = argparse.ArgumentParser(description="StyleGAN2 trainer")

    parser.add_argument("path", type=str, help="path to the lmdb dataset")
    parser.add_argument('--arch', type=str, default='stylegan2', help='model architectures (stylegan2 | swagan)')
    parser.add_argument(
        "--iter", type=int, default=800000, help="total training iterations"
    )
    parser.add_argument(
        "--batch", type=int, default=16, help="batch sizes for each gpus"
    )
    parser.add_argument(
        "--n_sample",
        type=int,
        default=64,
        help="number of the samples generated during training",
    )
    parser.add_argument(
        "--size", type=int, default=256, help="image sizes for the model"
    )
    parser.add_argument(
        "--r1", type=float, default=10, help="weight of the r1 regularization"
    )
    parser.add_argument(
        "--path_regularize",
        type=float,
        default=2,
        help="weight of the path length regularization",
    )
    parser.add_argument(
        "--path_batch_shrink",
        type=int,
        default=2,
        help="batch size reducing factor for the path length regularization (reduce memory consumption)",
    )
    parser.add_argument(
        "--d_reg_every",
        type=int,
        default=16,
        help="interval of the applying r1 regularization",
    )
    parser.add_argument(
        "--g_reg_every",
        type=int,
        default=4,
        help="interval of the applying path length regularization",
    )
    parser.add_argument(
        "--mixing", type=float, default=0.9, help="probability of latent code mixing"
    )
    parser.add_argument(
        "--ckpt",
        type=str,
        default=None,
        help="path to the checkpoints to resume training",
    )
    parser.add_argument("--lr", type=float, default=0.002, help="learning rate")
    parser.add_argument(
        "--channel_multiplier",
        type=int,
        default=2,
        help="channel multiplier factor for the model. config-f = 2, else = 1",
    )
    parser.add_argument(
        "--wandb", action="store_true", help="use weights and biases logging"
    )  # 默认 False，运行时有加--wandb，则变为 True
    parser.add_argument(
        "--local_rank", type=int, default=0, help="local rank for distributed training"
    )
    parser.add_argument(
        "--augment", action="store_true", help="apply non leaking augmentation"
    )  # 默认 False
    parser.add_argument(
        "--augment_p",
        type=float,
        default=0,
        help="probability of applying augmentation. 0 = use adaptive augmentation",
    )
    parser.add_argument(
        "--ada_target",
        type=float,
        default=0.6,
        help="target augmentation probability for adaptive augmentation",
    )
    parser.add_argument(
        "--ada_length",
        type=int,
        default=500 * 1000,
        help="target duraing to reach augmentation probability for adaptive augmentation",
    )
    parser.add_argument(
        "--ada_every",
        type=int,
        default=256,
        help="probability update interval of the adaptive augmentation",
    )

    args = parser.parse_args()

    # os.environ["WORLD_SIZE"] 由 torch.distributed.launch.py 产生
    n_gpu = int(os.environ["WORLD_SIZE"]) if "WORLD_SIZE" in os.environ else 1
    args.distributed = n_gpu > 1

    # local_rank 同样由 torch.distributed.launch 传递而来，只需要设置位置参数来接受
    if args.distributed:
        torch.cuda.set_device(args.local_rank)
        torch.distributed.init_process_group(backend="nccl", init_method="env://")
        synchronize()

    args.latent = 512   # 隐向量维度（z和w都是512）
    args.n_mlp = 8      # 8个全连接层

    args.start_iter = 0


    # 生成器(图片大小，隐向量维度，全连接层个数，large_net=*2)
    generator = Generator(
        args.size, args.latent, args.n_mlp, channel_multiplier=args.channel_multiplier
    ).to(device)
    discriminator = Discriminator(
        args.size, channel_multiplier=args.channel_multiplier
    ).to(device)
    g_ema = Generator(
        args.size, args.latent, args.n_mlp, channel_multiplier=args.channel_multiplier
    ).to(device)
    g_ema.eval()
    accumulate(g_ema, generator, 0)  # 指数平均g_ema

    # lazy比例 如每隔4次进行1次正则，则lazy比为4/5
    g_reg_ratio = args.g_reg_every / (args.g_reg_every + 1)
    d_reg_ratio = args.d_reg_every / (args.d_reg_every + 1)

    # 优化器，beta1=0，Adam退化成近似RMSProp算法
    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),
    )

    # 若存在checkpoint，则从中加载模型
    if args.ckpt is not None:
        print("load model:", args.ckpt)

        ckpt = torch.load(args.ckpt, map_location=lambda storage, loc: storage)  # 加载checkpoint到cpu

        try:
            ckpt_name = os.path.basename(args.ckpt)  # 返回 args.ckpt 的最后文件名
            args.start_iter = int(os.path.splitext(ckpt_name)[0])  # 分离文件名和扩展名，此处返回文件名（数字）

        except ValueError:
            pass

        # 加载checkpoint
        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"])

    if args.distributed:
        generator = nn.parallel.DistributedDataParallel(
            generator,
            device_ids=[args.local_rank],
            output_device=args.local_rank,
            broadcast_buffers=False,
        )  # broadcast_buffers=False 不同步batch_norm的统计量：均值和方差

        discriminator = nn.parallel.DistributedDataParallel(
            discriminator,
            device_ids=[args.local_rank],
            output_device=args.local_rank,
            broadcast_buffers=False,
        )
        print(f'Enable Distributed Training, Processing Number: {n_gpu}')

    transform = transforms.Compose(
        [
            transforms.RandomHorizontalFlip(),
            transforms.ToTensor(),
            transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5), inplace=True),
        ]
    )

    dataset = MultiResolutionDataset(args.path, transform, args.size)
    loader = data.DataLoader(
        dataset,
        batch_size=args.batch,
        sampler=data_sampler(dataset, shuffle=True, distributed=args.distributed),
        drop_last=True,
    )

    if get_rank() == 0 and wandb is not None and args.wandb:
        wandb.init(project="stylegan 2")

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