train.py

import argparse
import math
import random
import os
import torch
import lpips

import numpy as np

from torch import nn, autograd, optim
from torch.nn import functional as F
from torch.utils import data
from torchvision import transforms, utils
from tqdm import tqdm
from copy import deepcopy

from PIL import Image

try:
    import wandb

except ImportError:
    wandb = None

from lifelong_model import LifelongGenerator as Generator
from lifelong_model import Extra
from lifelong_model import LifelongPatchDiscriminator as Discriminator  # , Projection_head
from dataset import MultiResolutionDataset
from distributed import (
    get_rank,
    synchronize,
    reduce_loss_dict,
    reduce_sum,
    get_world_size,
)
from non_leaking import augment

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):
    if model.lfs:
        for name, param in model.named_parameters():
            if name.find('weight_modulation') >= 0:
                param.requires_grad = flag
            else:
                param.requires_grad = False
    else:
        for name, param in model.named_parameters():
            param.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)


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).reshape(
        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 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]
    )
    grad, = autograd.grad(
        outputs=(fake_img * noise).sum(), inputs=latents, create_graph=True,
    )
    path_lengths = torch.sqrt(grad.pow(2).sum(2).mean(1))

    path_mean = mean_path_length + decay * \
        (path_lengths.mean() - mean_path_length)

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

    return path_penalty, path_mean.detach(), path_lengths


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 get_subspace(args, init_z, vis_flag=False, size=None):
    std = args.subspace_std
    if size is None:
        bs = args.batch if not vis_flag else args.n_sample
    else:
        bs = size
    ind = np.random.randint(0, init_z.size(0), size=bs)
    z = init_z[ind]  # should give a tensor of size [batch_size, 512]
    for i in range(z.size(0)):
        for j in range(z.size(1)):
            z[i][j].data.normal_(z[i][j], std)
    return z

def extract_task_specific_parameters(model):
    state_dict = model.state_dict()
    task_specific_parameters = {}
    for key in state_dict.keys():
        if 'weight_modulation' in key:
            task_specific_parameters[key] = state_dict[key]
    return task_specific_parameters

def train(args, loader, generator, discriminator, extra, g_optim, d_optim, e_optim, device, current_task):
    loader = sample_data(loader)

    imsave_path = os.path.join('samples', args.exp)
    model_path = os.path.join('checkpoints', args.exp)

    if not os.path.exists(imsave_path):
        os.makedirs(imsave_path, exist_ok=True)
    if not os.path.exists(model_path):
        os.makedirs(model_path, exist_ok=True)

    # this defines the anchor points, and when sampling noise close to these, we impose image-level adversarial loss (Eq. 4 in the paper)
    init_z = torch.randn(args.n_train, args.latent, device=device)
    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 = {}

    accum = 0.5 ** (32 / (10 * 1000))
    ada_augment = torch.tensor([0.0, 0.0], device=device)
    ada_aug_p = args.augment_p if args.augment_p > 0 else 0.0
    ada_aug_step = args.ada_target / args.ada_length
    r_t_stat = 0

    # this defines which level feature of the discriminator is used to implement the patch-level adversarial loss: could be anything between [0, args.highp] 
    lowp, highp = 0, args.highp

    # the following defines the constant noise used for generating images at different stages of training
    if os.path.exists('sample_z.pt'):
        sample_z = torch.load('sample_z.pt')
    else:
        sample_z = torch.randn(args.n_sample, args.latent, device=device)

    for idx in pbar:
        i = idx + args.start_iter
        which = i % args.subspace_freq # defines whether we sample from anchor region in this iteration or other
        if i > args.iter:
            print("Done!")
            break

        real_img = next(loader)
        real_img = real_img.to(device)
            
        requires_grad(generator, False)
        requires_grad(discriminator, True)
        requires_grad(extra, True)
        
        if which > 0:
            # sample normally, apply patch-level adversarial loss
            noise = mixing_noise(args.batch, args.latent, args.mixing, device)
        else:
            # sample from anchors, apply image-level adversarial loss
            noise = [get_subspace(args, init_z.clone())]

        
        fake_img, _ = generator(noise)

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

        fake_pred, _ = discriminator(
            fake_img, extra=extra, flag=which, p_ind=np.random.randint(lowp, highp))
        real_pred, _ = discriminator(
            real_img, extra=extra, flag=which, p_ind=np.random.randint(lowp, highp))

        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()
        extra.zero_grad()
        d_loss.backward()
        d_optim.step()
        e_optim.step()
        
        if args.augment and args.augment_p == 0:
            ada_augment += torch.tensor(
                (torch.sign(real_pred).sum().item(), real_pred.shape[0]), device=device
            )
            ada_augment = reduce_sum(ada_augment)

            if ada_augment[1] > 255:
                pred_signs, n_pred = ada_augment.tolist()

                r_t_stat = pred_signs / n_pred

                if r_t_stat > args.ada_target:
                    sign = 1

                else:
                    sign = -1

                ada_aug_p += sign * ada_aug_step * n_pred
                ada_aug_p = min(1, max(0, ada_aug_p))
                ada_augment.mul_(0)

        d_regularize = i % args.d_reg_every == 0

        if d_regularize:
            real_img.requires_grad = True
            real_pred, _ = discriminator(
                real_img, extra=extra, flag=which, p_ind=np.random.randint(lowp, highp))
            real_pred = real_pred.view(real_img.size(0), -1)
            real_pred = real_pred.mean(dim=1).unsqueeze(1)

            r1_loss = d_r1_loss(real_pred, real_img)

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

            d_optim.step()
            e_optim.step()
            
        loss_dict["r1"] = r1_loss
        
        requires_grad(generator, True)
        requires_grad(discriminator, False)
        requires_grad(extra, False)
        
        if which > 0:
            noise = mixing_noise(args.batch, args.latent, args.mixing, device)
        else:
            noise = [get_subspace(args, init_z.clone())]

        fake_img, _ = generator(noise)

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

        fake_pred, _ = discriminator(
            fake_img, extra=extra, flag=which, p_ind=np.random.randint(lowp, highp))
        
        g_loss = g_nonsaturating_loss(fake_pred)
        
        if args.cluster_wise_mode_seeking:
            if 'lpips_fn' not in locals():
                lpips_fn = lpips.LPIPS(net='vgg').cuda()
                lpips_fn.eval()
            cms_loss = 0
            cluster_count = 0

            z = torch.randn(args.batch * 4, 512, device='cuda')
            w = generator.style(z).unsqueeze(1).repeat(1, 14, 1)
            fake_img, feats = generator([w], input_is_latent=True, return_feats=True)
            lpips_dists = torch.zeros(args.batch * 4, args.batch)

            with torch.no_grad():
                fake_img_reshaped = fake_img.repeat_interleave(args.batch, dim=0)
                real_img_reshaped = real_img.repeat(args.batch * 4, 1, 1, 1)
                lpips_dists = lpips_fn(fake_img_reshaped, real_img_reshaped).view(args.batch * 4, args.batch)
                indices = torch.argmin(lpips_dists, dim=1) # [16]

            for pos in range(real_img.shape[0]):
                if len(torch.where(indices == pos)[0]) >= 2:
                    cluster_count += 1

            for pos1 in range(real_img.shape[0]):
                cluster_length = len(torch.where(indices == pos1)[0])
                if cluster_length < 2:
                    continue
                else:
                    for pos2 in range(cluster_length):
                        for pos3 in range(pos2 + 1, cluster_length):
                            cms_loss = cms_loss + (F.l1_loss(w[pos2], w[pos3]) / F.l1_loss(z[pos2], z[pos3]))
            
            # features
            for pos1 in range(real_img.shape[0]):
                cluster_length = len(torch.where(indices == pos1)[0])
                if cluster_length < 2:
                    continue
                else:
                    for pos2 in range(cluster_length):
                        for pos3 in range(pos2 + 1, cluster_length):
                            temp_loss = 0
                            for l in range(len(feats)):
                                temp_loss = temp_loss + (F.l1_loss(feats[l][pos2], feats[l][pos3]) / F.l1_loss(w[pos2], w[pos3]))
                            cms_loss = cms_loss + temp_loss / len(feats)

            # images
            for pos1 in range(real_img.shape[0]):
                cluster_length = len(torch.where(indices == pos1)[0])
                if cluster_length < 2:
                    continue
                else:
                    for pos2 in range(cluster_length):
                        for pos3 in range(pos2 + 1, cluster_length):
                            cms_loss = cms_loss + (F.l1_loss(fake_img[pos2], fake_img[pos3]) / F.l1_loss(w[pos2], w[pos3]))

            cms_loss = cms_loss / cluster_count
            eps = 1e-5
            cms_loss = 1 / (cms_loss + eps)
            lambda_g = 1

            cms_loss = lambda_g * cms_loss
            g_loss = g_loss + cms_loss

        else:
            cms_loss = None

        loss_dict["g"] = g_loss
        if cms_loss is not None:
            loss_dict["cms_loss"] = cms_loss

        generator.zero_grad()
        g_loss.backward()
        g_optim.step()
        g_regularize = i % args.g_reg_every == 0

        # to save up space
        del g_loss, d_loss, fake_img, fake_pred, real_img, real_pred
            
        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)

            path_loss, mean_path_length, path_lengths = g_path_regularize(
                fake_img, latents, mean_path_length
            )

            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()
            )

       
        loss_dict["path"] = path_loss
        loss_dict["path_length"] = path_lengths.mean()
        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 cms_loss is not None:
            cms_loss_val = loss_reduced["cms_loss"].mean().item()
        else:
            cms_loss_val = 0

        if get_rank() == 0:
            pbar.set_description(
                (
                    f"Task: {current_task}; "
                    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}; cms_loss: {cms_loss_val:.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,
                        "New Loss G": cms_loss_val,
                    }
                )
            if i % args.img_freq == 0:
                with torch.set_grad_enabled(False):
                    generator.eval()
                    sample, _ = generator([sample_z.data])
                    utils.save_image(
                        sample,
                        f"%s/{current_task}_{str(i).zfill(6)}.png" % (imsave_path),
                        nrow=int(args.n_sample ** 0.5),
                        normalize=True,
                        value_range=(-1, 1),
                    )
                    del sample
                    generator.train()
            
            if (i % args.save_freq == 0) and (i > 0):
                torch.save(
                    {
                        "g_ema": extract_task_specific_parameters(generator),
                        # uncomment the following lines only if you wish to resume training after saving. Otherwise, saving just the generator is sufficient for evaluations

                        #"g": generator.state_dict(),
                        #"g_s": g_source.state_dict(),
                        #"d": discrimintaor.state_dict(),
                        #"g_optim": g_optim.state_dict(),
                        #"d_optim": d_optim.state_dict(),
                    },
                    f"%s/{current_task}_{str(i).zfill(6)}.pt" % (model_path),
                )


if __name__ == "__main__":
    device = "cuda"
    parser = argparse.ArgumentParser()

    parser.add_argument("--data_path", type=str, required=True)
    parser.add_argument("--iter", type=int, default=3002)
    parser.add_argument("--save_freq", type=int, default=100)
    parser.add_argument("--img_freq", type=int, default=100)
    parser.add_argument("--highp", type=int, default=1)
    parser.add_argument("--subspace_freq", type=int, default=4)
    parser.add_argument("--batch", type=int, default=4)
    parser.add_argument("--n_sample", type=int, default=25)
    parser.add_argument("--size", type=int, default=256)
    parser.add_argument("--r1", type=float, default=10)
    parser.add_argument("--path_regularize", type=float, default=2)
    parser.add_argument("--path_batch_shrink", type=int, default=2)
    parser.add_argument("--d_reg_every", type=int, default=16)
    parser.add_argument("--g_reg_every", type=int, default=4)
    parser.add_argument("--mixing", type=float, default=0.9)
    parser.add_argument("--subspace_std", type=float, default=0.1)
    parser.add_argument("--ckpt", type=str, default=None)
    parser.add_argument("--exp", type=str, default=None, required=True)
    parser.add_argument("--lr", type=float, default=0.002)
    parser.add_argument("--channel_multiplier", type=int, default=2)
    parser.add_argument("--wandb", action="store_true")
    parser.add_argument("--augment", dest='augment', action='store_true')
    parser.add_argument("--no-augment", dest='augment', action='store_false')
    parser.add_argument("--augment_p", type=float, default=0.0)
    parser.add_argument("--ada_target", type=float, default=0.6)
    parser.add_argument("--ada_length", type=int, default=500 * 1000)
    parser.add_argument("--n_train", type=int, default=10)
    parser.add_argument('--truncation', type=float, default=1)
    parser.add_argument('--seed', type=int, default=1)
    parser.add_argument('--rank', type=int, default=1)
    parser.add_argument('--left_use_add', action='store_true')
    parser.add_argument('--left_use_act', action='store_true')
    parser.add_argument('--cluster_wise_mode_seeking', action='store_true')

    args = parser.parse_args()
    print(args)
    if args.data_path[-1] == '/':
        args.data_path = args.data_path[:-1] # processed_data/Sketches/10shot/1
    args.task = args.data_path.split('/')[-1]
    torch.manual_seed(args.seed)
    random.seed(args.seed)

    n_gpu = 4
    args.distributed = n_gpu > 1

    args.latent = 512
    args.n_mlp = 8
    args.start_iter = 0

    generator = Generator(
        args.size, args.latent, args.n_mlp, channel_multiplier=args.channel_multiplier, lifelong=True, rank=args.rank, left_use_act=args.left_use_act, left_use_add=args.left_use_add
    ).to(device)
    
    discriminator = Discriminator(
        args.size, channel_multiplier=args.channel_multiplier
    ).to(device)
    
    extra = Extra().to(device)

    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),
    )

    e_optim = optim.Adam(
        extra.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, map_location=lambda storage, loc: storage)
        ckpt_source = torch.load(args.ckpt, map_location=lambda storage, loc: storage)

        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"], strict=False)

        # d_source = nn.parallel.DataParallel(d_source)
        # discriminator = nn.parallel.DataParallel(discriminator)
        discriminator.load_state_dict(ckpt["d"], strict=False)

        # if 'g_optim' in ckpt.keys():
        #     g_optim.load_state_dict(ckpt["g_optim"])
        # if 'd_optim' in ckpt.keys():
        #     d_optim.load_state_dict(ckpt["d_optim"])
                
    requires_grad(generator, True)

    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.data_path, transform, args.size)
    loader = data.DataLoader(dataset, batch_size=args.batch, sampler=data_sampler(dataset, shuffle=True, distributed=False), drop_last=True, num_workers=4)
    
    if get_rank() == 0 and wandb is not None and args.wandb:
        wandb.init(project=args.exp, name=args.task)

    s = 0
    entire_s = 0
    for name, param in generator.named_parameters():
        if param.requires_grad:
            print(name)
            s += torch.numel(param)
        if 'weight_modulation' not in name:
            entire_s += torch.numel(param)
    print('(Generator) Trainable parameters:', s)
    print('(Generator) Entire parameters:', entire_s)
    print("=" * 100)
    s = 0
    entire_s = 0
    for name, param in discriminator.named_parameters():
        if param.requires_grad:
            print(name)
            s += torch.numel(param)
        if 'weight_modulation' not in name:
            entire_s += torch.numel(param)
    print('(Discriminator) Trainable parameters:', s)
    print('(Discriminator) Entire parameters:', entire_s)
    print("=" * 100)
    
    train(args, loader, generator, discriminator, extra, g_optim, d_optim, e_optim, device, args.task)