training/training_loop.py

# Copyright (c) 2021, NVIDIA CORPORATION.  All rights reserved.
#
# NVIDIA CORPORATION and its licensors retain all intellectual property
# and proprietary rights in and to this software, related documentation
# and any modifications thereto.  Any use, reproduction, disclosure or
# distribution of this software and related documentation without an express
# license agreement from NVIDIA CORPORATION is strictly prohibited.

import os
import time
import copy
import json
import dill as pickle
import psutil
import PIL.Image
import numpy as np
import torch
import dnnlib
from torch_utils import misc
from torch_utils import training_stats
from torch_utils.ops import conv2d_gradfix
from torch_utils.ops import grid_sample_gradfix
from torchvision.utils import save_image
import math
import legacy
from metrics import metric_main
import torch.nn.functional as F
np.set_printoptions(formatter={'float': '{:0.2f}'.format})
from collections import Counter
#----------------------------------------------------------------------------

class SparsestVector:
    def __init__(self):
        self.sparsest_vector = None

    def add(self, vector):
        """Add a vector, only keeping it if it is sparser than the current stored one."""
        if self.sparsest_vector is None:
            self.sparsest_vector = vector
        else:
            current_nonzero = torch.count_nonzero(self.sparsest_vector).item()
            new_nonzero = torch.count_nonzero(vector).item()

            # Keep the new vector only if it's sparser (fewer non-zero elements)
            if new_nonzero < current_nonzero:
                self.sparsest_vector = vector

    def check(self):
        """Returns the sparsest vector currently stored."""
        return self.sparsest_vector


def setup_snapshot_image_grid(training_set, random_seed=0):
    rnd = np.random.RandomState(random_seed)
    gw = int(np.clip(768*2 // training_set.image_shape[2], 7, 32))
    gh = int(np.clip(432*2 // training_set.image_shape[1], 4, 32))

    # No labels => show random subset of training samples.
    if not training_set.has_labels:
        all_indices = list(range(len(training_set)))
        rnd.shuffle(all_indices)
        grid_indices = [all_indices[i % len(all_indices)] for i in range(gw * gh)]
        label_groups = []
    else:
        # Group training samples by label.
        label_groups = dict() # label => [idx, ...]
        for idx in range(len(training_set)):
            label = tuple(training_set.get_details(idx).raw_label.flat[::-1])
            if label not in label_groups:
                label_groups[label] = []
            label_groups[label].append(idx)
        if training_set.image_shape[1] < 256:
            gw *= 2
        gh *= len(label_groups)
        #gw = min(gw, 16)
        # Reorder.
        label_order = sorted(label_groups.keys())
        for label in label_order:
            rnd.shuffle(label_groups[label])

        # Organize into grid.
        grid_indices = []
        for y in range(len(label_groups)):
            label = label_order[y % len(label_order)]
            indices = label_groups[label]
            grid_indices += [indices[x % len(indices)] for x in range(gw)]
            label_groups[label] = [indices[(i + gw) % len(indices)] for i in range(len(indices))]

    # Load data.
    images, labels = zip(*[training_set[i] for i in grid_indices])
    return (gw, len(label_groups)), np.stack(images), np.stack(labels), len(label_groups)

#----------------------------------------------------------------------------

def save_image_grid(img, fname, drange, grid_size):
    lo, hi = drange
    img = np.asarray(img, dtype=np.float32)
    img = (img - lo) * (255 / (hi - lo))
    img = np.rint(img).clip(0, 255).astype(np.uint8)

    gw, gh = grid_size
    _N, C, H, W = img.shape
    img = img.reshape(gh, gw, C, H, W)
    img = img.transpose(0, 3, 1, 4, 2)
    img = img.reshape(gh * H, gw * W, C)

    assert C in [1, 3]
    if C == 1:
        PIL.Image.fromarray(img[:, :, 0], 'L').save(fname)
    if C == 3:
        PIL.Image.fromarray(img, 'RGB').save(fname)


class VectorHistoryChecker:
    def __init__(self, b, d, m):
        self.b = b
        self.d = d
        self.m = m
        self.history = torch.ones(b, d, m)*1e99  # Initialize history with zeros
        self.current_index = 0

    def update_history(self, new_version):
        """Update history with the new version of the vector."""
        self.history[:, :, self.current_index] = new_version.cpu()
        self.current_index = (self.current_index + 1) % self.m

    def check_history(self, input_version):
        """Check if the input version matches all m history versions for each row."""
        consistency = torch.ones(self.b, dtype=torch.bool)  # Initialize as True for all rows
        for i in range(self.m):
            # Check row-wise equality across the history
            consistency &= torch.all(self.history[:, :, i] == input_version.cpu(), dim=1)
        return consistency

    def get_history(self):
        """Get the current history."""
        return self.history

class ColumnHistoryChecker:
    def __init__(self, b, d, m):
        self.b = b
        self.d = d
        self.m = m
        self.history = torch.ones(b, d, m)*1e99  # Initialize history with zeros
        self.current_index = 0

    def update_history(self, new_version):
        """Update history with the new version of the vector."""
        self.history[:, :, self.current_index] = new_version.cpu()
        self.current_index = (self.current_index + 1) % self.m

    def check_history(self, input_version):
        """Check if the input version matches all m history versions for each row."""
        consistency = torch.ones(self.d, dtype=torch.bool)  # Initialize as True for all rows
        for i in range(self.m):
            # Check column-wise equality across the history
            consistency &= torch.all(self.history[:, :, i] == input_version.cpu(), dim=0)
        return consistency

    def get_history(self):
        """Get the current history."""
        return self.history
#----------------------------------------------------------------------------

def training_loop(
    run_dir                 = '.',      # Output directory.
    training_set_kwargs     = {},       # Options for training set.
    data_loader_kwargs      = {},       # Options for torch.utils.data.DataLoader.
    G_kwargs                = {},       # Options for generator network.
    D_kwargs                = {},       # Options for discriminator network.
    G_opt_kwargs            = {},       # Options for generator optimizer.
    D_opt_kwargs            = {},       # Options for discriminator optimizer.
    augment_kwargs          = None,     # Options for augmentation pipeline. None = disable.
    loss_kwargs             = {},       # Options for loss function.
    metrics                 = [],       # Metrics to evaluate during training.
    random_seed             = 0,        # Global random seed.
    num_gpus                = 1,        # Number of GPUs participating in the training.
    rank                    = 0,        # Rank of the current process in [0, num_gpus[.
    batch_size              = 4,        # Total batch size for one training iteration. Can be larger than batch_gpu * num_gpus.
    batch_gpu               = 4,        # Number of samples processed at a time by one GPU.
    ema_kimg                = 10,       # Half-life of the exponential moving average (EMA) of generator weights.
    ema_rampup              = None,     # EMA ramp-up coefficient.
    G_reg_interval          = 4,        # How often to perform regularization for G? None = disable lazy regularization.
    D_reg_interval          = 16,       # How often to perform regularization for D? None = disable lazy regularization.
    augment_p               = 0,        # Initial value of augmentation probability.
    ada_target              = None,     # ADA target value. None = fixed p.
    ada_interval            = 4,        # How often to perform ADA adjustment?
    ada_kimg                = 500,      # ADA adjustment speed, measured in how many kimg it takes for p to increase/decrease by one unit.
    total_kimg              = 25000,    # Total length of the training, measured in thousands of real images.
    kimg_per_tick           = 4,        # Progress snapshot interval.
    image_snapshot_ticks    = 50,       # How often to save image snapshots? None = disable.
    network_snapshot_ticks  = 50,       # How often to save network snapshots? None = disable.
    resume_pkl              = None,     # Network pickle to resume training from.
    cudnn_benchmark         = True,     # Enable torch.backends.cudnn.benchmark?
    allow_tf32              = False,    # Enable torch.backends.cuda.matmul.allow_tf32 and torch.backends.cudnn.allow_tf32?
    abort_fn                = None,     # Callback function for determining whether to abort training. Must return consistent results across ranks.
    progress_fn             = None,     # Callback function for updating training progress. Called for all ranks.
    lambda_sparse           = None,
    lambda_entropy          = None,
    lambda_ortho            = None,
    lambda_colvar            = None,
    lambda_rowvar            = None,
    lambda_equal          = None,
    lambda_epsilon = None,
    lambda_path=None,
    g_iter=None,
    temperature=1,
):
    # Initialize.
    start_time = time.time()
    device = torch.device('cuda', rank)
    np.random.seed(random_seed * num_gpus + rank)
    torch.manual_seed(random_seed * num_gpus + rank)
    torch.backends.cudnn.benchmark = cudnn_benchmark    # Improves training speed.
    torch.backends.cuda.matmul.allow_tf32 = allow_tf32  # Allow PyTorch to internally use tf32 for matmul
    torch.backends.cudnn.allow_tf32 = allow_tf32        # Allow PyTorch to internally use tf32 for convolutions
    conv2d_gradfix.enabled = True                       # Improves training speed.
    grid_sample_gradfix.enabled = True                  # Avoids errors with the augmentation pipe.

    # Load training set.
    if rank == 0:
        print('Loading training set...')
    training_set = dnnlib.util.construct_class_by_name(**training_set_kwargs) # subclass of training.dataset.Dataset
    training_set_sampler = misc.InfiniteSampler(dataset=training_set, rank=rank, num_replicas=num_gpus, seed=random_seed)
    training_set_iterator = iter(torch.utils.data.DataLoader(dataset=training_set, sampler=training_set_sampler, batch_size=batch_size//num_gpus, **data_loader_kwargs))
    if rank == 0:
        print()
        print('Num images: ', len(training_set))
        print('Image shape:', training_set.image_shape)
        print('Label shape:', training_set.label_shape)
        print()

    # Construct networks.
    if rank == 0:
        print('Constructing networks...')
    common_kwargs = dict(c_dim=training_set.label_dim, img_resolution=training_set.resolution, img_channels=training_set.num_channels)
    G = dnnlib.util.construct_class_by_name(**G_kwargs, **common_kwargs).train().requires_grad_(False).to(device) # subclass of torch.nn.Module
    D = dnnlib.util.construct_class_by_name(**D_kwargs, **common_kwargs).train().requires_grad_(False).to(device) # subclass of torch.nn.Module
    G_ema = copy.deepcopy(G).eval()

    M_kwargs = dnnlib.EasyDict(class_name='training.networks.ConceptMaskNetwork', c_dim=training_set.label_dim, i_dim=G_kwargs.mapping_kwargs.i_dim)
    M = dnnlib.util.construct_class_by_name(**M_kwargs).train().requires_grad_(False).to(device) # subclass of torch.nn.Module
    M_ema = copy.deepcopy(M).eval()

    # Resume from existing pickle.
    if (resume_pkl is not None) and (rank == 0):
        print(f'Resuming from "{resume_pkl}"')
        with dnnlib.util.open_url(resume_pkl) as f:
            resume_data = legacy.load_network_pkl(f)
        for name, module in [('G', G), ('D', D), ('G_ema', G_ema), ('M', M), ('M_ema', M_ema)]:
            misc.copy_params_and_buffers(resume_data[name], module, require_all=False)

    # Print network summary tables.
    if rank == 0:
        z = torch.empty([batch_gpu, G.z_dim], device=device)
        c = torch.empty([batch_gpu, G.c_dim], device=device)
        m = torch.empty([batch_gpu, G_kwargs.mapping_kwargs.i_dim], device=device)
        img = misc.print_module_summary(G, [z, m])
        misc.print_module_summary(D, [img, c])

    # Setup augmentation.
    if rank == 0:
        print('Setting up augmentation...')
    augment_pipe = None
    ada_stats = None
    if (augment_kwargs is not None) and (augment_p > 0 or ada_target is not None):
        augment_pipe = dnnlib.util.construct_class_by_name(**augment_kwargs).train().requires_grad_(False).to(device) # subclass of torch.nn.Module
        augment_pipe.p.copy_(torch.as_tensor(augment_p))
        if ada_target is not None:
            ada_stats = training_stats.Collector(regex='Loss/signs/real')

    # Distribute across GPUs.
    if rank == 0:
        print(f'Distributing across {num_gpus} GPUs...')
    ddp_modules = dict()
    for name, module in [('G_mapping', G.mapping), ('G_synthesis', G.synthesis), ('D', D), (None, G_ema), ('augment_pipe', augment_pipe),
                         ('M', M), (None, M_ema)
                         ]:
        if (num_gpus > 1) and (module is not None) and len(list(module.parameters())) != 0:
            module.requires_grad_(True)
            module = torch.nn.parallel.DistributedDataParallel(module, device_ids=[device], broadcast_buffers=False)
            module.requires_grad_(False)
        if name is not None:
            ddp_modules[name] = module

    # Setup training phases.
    if rank == 0:
        print('Setting up training phases...')
    loss = dnnlib.util.construct_class_by_name(device=device, **ddp_modules, **loss_kwargs) # subclass of training.loss.Loss
    phases = []
    for name, module, opt_kwargs, reg_interval in [('G', G, G_opt_kwargs, G_reg_interval), ('D', D, D_opt_kwargs, D_reg_interval)]:
        if reg_interval is None:
            opt = dnnlib.util.construct_class_by_name(params=module.parameters(), **opt_kwargs) # subclass of torch.optim.Optimizer
            phases += [dnnlib.EasyDict(name=name+'both', module=module, opt=opt, interval=1)]
        else: # Lazy regularization.
            mb_ratio = reg_interval / (reg_interval + 1)
            opt_kwargs = dnnlib.EasyDict(opt_kwargs)
            opt_kwargs.lr = opt_kwargs.lr * mb_ratio
            opt_kwargs.betas = [beta ** mb_ratio for beta in opt_kwargs.betas]
            opt = dnnlib.util.construct_class_by_name(module.parameters(), **opt_kwargs) # subclass of torch.optim.Optimizer
            phases += [dnnlib.EasyDict(name=name+'main', module=module, opt=opt, interval=1)]
            if name == 'G' and g_iter>0:
                phases += ([dnnlib.EasyDict(name=name + 'main', module=module, opt=opt, interval=1)] * g_iter)
            phases += [dnnlib.EasyDict(name=name+'reg', module=module, opt=opt, interval=reg_interval)]


    for name, module, opt_kwargs, reg_interval in [('M', M, G_opt_kwargs, G_reg_interval)]:
        mb_ratio = reg_interval / (reg_interval + 1)
        opt_kwargs = dnnlib.EasyDict(opt_kwargs)
        opt_kwargs.lr = opt_kwargs.lr * mb_ratio
        opt_kwargs.betas = [beta ** mb_ratio for beta in opt_kwargs.betas]
        #M_opt = dnnlib.util.construct_class_by_name(module.parameters(), **opt_kwargs) # subclass of torch.optim.Optimizer
        #M_opt = torch.optim.SGD(module.parameters(), lr=0.01, momentum=0.9)
        print(opt_kwargs.betas, ' >>>>>>>> opt kwargs ssss')
        M_opt = torch.optim.AdamW(module.parameters(), lr=opt_kwargs.lr, betas=(0.9, 0.999), eps=opt_kwargs.eps,
                                  weight_decay=0.01, amsgrad=False)


    for phase in phases:
        phase.start_event = None
        phase.end_event = None
        if rank == 0:
            phase.start_event = torch.cuda.Event(enable_timing=True)
            phase.end_event = torch.cuda.Event(enable_timing=True)

    # Export sample images.
    grid_size = None
    grid_z = None
    grid_c = None
    if rank == 0:
        print('Exporting sample images...')
        grid_size, images, labels, num_domains = setup_snapshot_image_grid(training_set=training_set)
        save_image_grid(images, os.path.join(run_dir, 'reals.jpg'), drange=[0,255], grid_size=grid_size)
        if labels.shape[1] > 0:
            grid_z = []
            for i in range(grid_size[1]//num_domains):
                random_z = (torch.randn(grid_size[0], G.z_dim, device=device))
                for j in range(num_domains):
                    grid_z.append(random_z)
            grid_z = torch.cat(grid_z, 0).split(batch_gpu)
        else:
            grid_z = torch.randn([labels.shape[0], G.z_dim], device=device).split(batch_gpu)
        grid_c = torch.from_numpy(labels).to(device)
        grid_c = grid_c.split(batch_gpu)
        images = torch.cat([G_ema(z=z, c=M_ema(c), noise_mode='const').cpu() for z, c in zip(grid_z, grid_c)]).numpy()
        save_image_grid(images, os.path.join(run_dir, 'fakes_init.jpg'), drange=[-1,1], grid_size=grid_size)

    # Initialize logs.
    if rank == 0:
        print('Initializing logs...')
    stats_collector = training_stats.Collector(regex='.*')
    stats_metrics = dict()
    stats_jsonl = None
    stats_tfevents = None
    if rank == 0:
        stats_jsonl = open(os.path.join(run_dir, 'stats.jsonl'), 'wt')
        try:
            import torch.utils.tensorboard as tensorboard
            stats_tfevents = tensorboard.SummaryWriter(run_dir)
        except ImportError as err:
            print('Skipping tfevents export:', err)

    # Train.
    if rank == 0:
        print(f'Training for {total_kimg} kimg...')
        print()
    cur_nimg = 0
    cur_tick = 0
    tick_start_nimg = cur_nimg
    tick_start_time = time.time()
    maintenance_time = tick_start_time - start_time
    init_temperature = 1.0
    min_temperature = 0.5
    batch_idx = 0

    if progress_fn is not None:
        progress_fn(0, total_kimg)


    names = ['Red 0', 'Red 1', 'Green 0', 'Green 1', 'Green 2', 'Green 3', 'Green 4', 'Green 5', 'Green 6', 'Green 7',
             'Green 8', 'Green 9', 'Red 2', 'Blue 0', 'Blue 1', 'Blue 2', 'Blue 3', 'Blue 4', 'Blue 5', 'Blue 6', 'Blue 7', 'Blue 8', 'Blue 9',
             'Red 3', 'Red 4', 'Red 5', 'Red 6', 'Red 7', 'Red 8', 'Red 9'
             ]
    if G.mapping.c_dim == 30:
        names = [
        'Blue 0', 'Blue 1', 'Blue 2', 'Blue 3', 'Blue 4', 'Blue 5', 'Blue 6', 'Blue 7', 'Blue 8', 'Blue 9',
        'Green 0', 'Green 1', 'Green 2', 'Green 3', 'Green 4', 'Green 5', 'Green 6', 'Green 7', 'Green 8', 'Green 9',
            'Red 0', 'Red 1', 'Red 2','Red 3', 'Red 4', 'Red 5', 'Red 6', 'Red 7', 'Red 8', 'Red 9'
             ]
    elif G.mapping.c_dim == 8:
        names = [
            'Bald NoSmile Male', 'Bald Smile Male', 'Black NoSmile Female', 'Black NoSmile Male', 'Black Smile Female', 'Black Smile Male',
            'Blond NoSmile Female', 'Blond Smile Female'

        ]

    #names = ['Green Apple', 'Green Banana', 'Green Pear', 'Red Apple', 'Red Pear', 'Red Strawberry', 'Yellow Banana', 'Yellow Pineapple', 'Yellow StarFruit']
    #names = ['Green Apple', 'Green Banana', 'Green Pear', 'Red Apple', 'Red Pear', 'Red Strawberry', 'Yellow Banana', 'Yellow Pineapple', 'Yellow StarFruit']
    #names = ['Yellow 1', 'Purple 1', 'Red 1', 'Yellow 2', 'White 1', 'White 2', 'Red 2', 'Purple 2']
    version_history_checker = VectorHistoryChecker(G.mapping.c_dim, G.mapping.i_dim, 3)
    column_history_cheker = ColumnHistoryChecker(G.mapping.c_dim, G.mapping.i_dim, 3)
    binary_mask_checker = SparsestVector()
    use_best_binary = 10
    while True:
        ready = False
        cur_kimg = cur_nimg / 1000.0
        should_restart = (cur_tick % 40 ==0)
        if cur_tick<=5:
            cur_lambda_rowvar = lambda_rowvar
            cur_lambda_colvar = 0
            cur_lambda_sparse = lambda_sparse
            cur_entropy_thr = 0.6
            cur_lambda_equal = 0
            cur_lambda_entropy = lambda_entropy
        else:
            cur_lambda_rowvar = 0
            cur_lambda_colvar = lambda_colvar
            cur_lambda_sparse = lambda_sparse
            cur_entropy_thr = 0.9
            cur_lambda_equal = lambda_equal
            cur_lambda_entropy = lambda_entropy

        cur_lambda_ortho = lambda_ortho
        cur_temperature = 1.

        # Fetch training data.
        with torch.autograd.profiler.record_function('data_fetch'):
            phase_real_img, phase_real_c = next(training_set_iterator)
            phase_real_img = (phase_real_img.to(device).to(torch.float32) / 127.5 - 1).split(batch_gpu)
            phase_real_c = phase_real_c.to(device).split(batch_gpu)
            all_gen_z = torch.randn([len(phases) * batch_size, G.z_dim], device=device)
            all_gen_z = [phase_gen_z.split(batch_gpu) for phase_gen_z in all_gen_z.split(batch_size)]
            all_gen_c = [training_set.get_label(np.random.randint(len(training_set))) for _ in range(len(phases) * batch_size)]
            """
            all_gen_c = []
            for ta in tmp_all_gen_c:
                all_gen_c.append(F.one_hot(torch.randint(0, 30, (1,)), num_classes=30).float().to(device).squeeze().cpu().numpy())
            tmp_all_gen_c = torch.from_numpy(np.stack(tmp_all_gen_c)).to(device)
            print(all_gen_c.size(), ' >>>>>>>>>>>>>>>>> all genc ',  tmp_all_gen_c.size(), ' >>>>>>>>>>>>>>>>> tmp all genc ')
            """
            all_gen_c = torch.from_numpy(np.stack(all_gen_c)).pin_memory().to(device)
            all_gen_c = [phase_gen_c.split(batch_gpu) for phase_gen_c in all_gen_c.split(batch_size)]


        loss_dict = {}
        # Execute training phases.

        gmain_count = 0
        for phase, phase_gen_z, phase_gen_c in zip(phases, all_gen_z, all_gen_c):
            if batch_idx % phase.interval != 0:
                continue

            if phase.name == 'Gmain':
                gmain_count += 1

            only1G = ((cur_tick>use_best_binary) and (gmain_count>1) and (phase.name == 'Gmain'))
            if only1G:
                continue

            # Initialize gradient accumulation.
            if phase.start_event is not None:
                phase.start_event.record(torch.cuda.current_stream(device))
            phase.opt.zero_grad(set_to_none=True)
            phase.module.requires_grad_(True)
            M_opt.zero_grad(set_to_none=True)
            if phase.name == 'Gmain':
                M.requires_grad_(True)

            # Accumulate gradients over multiple rounds.
            for round_idx, (real_img, real_c, gen_z, gen_c) in enumerate(zip(phase_real_img, phase_real_c, phase_gen_z, phase_gen_c)):
                sync = (round_idx == batch_size // (batch_gpu * num_gpus) - 1)
                gain = phase.interval
                tmp_loss_dict = loss.accumulate_gradients(phase=phase.name, real_img=real_img, real_c=real_c, gen_z=gen_z, gen_c=gen_c, sync=sync, gain=gain,
                                          lambda_sparse=cur_lambda_sparse, lambda_entropy=cur_lambda_entropy, lambda_ortho=cur_lambda_ortho, lambda_path=lambda_path,
                                                          lambda_epsilon=lambda_epsilon, lambda_colvar=cur_lambda_colvar, lambda_rowvar=cur_lambda_rowvar,
                                                          lambda_equal=cur_lambda_equal, temperature=cur_temperature, entropy_thr=cur_entropy_thr,
                                                          )
                loss_dict.update(tmp_loss_dict)


            # Update weights.
            phase.module.requires_grad_(False)
            M.requires_grad_(False)
            with torch.autograd.profiler.record_function(phase.name + '_opt'):
                for param in phase.module.parameters():
                    if param.grad is not None:
                        misc.nan_to_num(param.grad, nan=0, posinf=1e5, neginf=-1e5, out=param.grad)
                phase.opt.step()

                for param in M.parameters():
                    if param.grad is not None:
                        misc.nan_to_num(param.grad, nan=0, posinf=1e5, neginf=-1e5, out=param.grad)
                M_opt.step()

            if phase.end_event is not None:
                phase.end_event.record(torch.cuda.current_stream(device))

        # Update G_ema.
        with torch.autograd.profiler.record_function('Gema'):
            ema_nimg = ema_kimg * 1000
            if ema_rampup is not None:
                ema_nimg = min(ema_nimg, cur_nimg * ema_rampup)
            ema_beta = 0.5 ** (batch_size / max(ema_nimg, 1e-8))
            for p_ema, p in zip(G_ema.parameters(), G.parameters()):
                p_ema.copy_(p.lerp(p_ema, ema_beta))
            for b_ema, b in zip(G_ema.buffers(), G.buffers()):
                b_ema.copy_(b)

            #ema_beta = 0.9
            for p_ema, p in zip(M_ema.parameters(), M.parameters()):
                p_ema.copy_(p.lerp(p_ema, ema_beta))
            for b_ema, b in zip(M_ema.buffers(), M.buffers()):
                b_ema.copy_(b)

        # Update state.
        cur_nimg += batch_size
        batch_idx += 1

        # Execute ADA heuristic.
        if (ada_stats is not None) and (batch_idx % ada_interval == 0):
            ada_stats.update()
            adjust = np.sign(ada_stats['Loss/signs/real'] - ada_target) * (batch_size * ada_interval) / (ada_kimg * 1000)
            augment_pipe.p.copy_((augment_pipe.p + adjust).max(misc.constant(0, device=device)))

        # Perform maintenance tasks once per tick.
        done = (cur_nimg >= total_kimg * 1000)
        if (not done) and (cur_tick != 0) and (cur_nimg < tick_start_nimg + kimg_per_tick * 1000):
            continue

        # Print status line, accumulating the same information in stats_collector.
        tick_end_time = time.time()
        fields = []
        fields += [f"tick {training_stats.report0('Progress/tick', cur_tick):<5d}"]
        fields += [f"kimg {training_stats.report0('Progress/kimg', cur_nimg / 1e3):<8.1f}"]
        fields += [f"time {dnnlib.util.format_time(training_stats.report0('Timing/total_sec', tick_end_time - start_time)):<12s}"]
        fields += [f"sec/tick {training_stats.report0('Timing/sec_per_tick', tick_end_time - tick_start_time):<7.1f}"]
        fields += [f"sec/kimg {training_stats.report0('Timing/sec_per_kimg', (tick_end_time - tick_start_time) / (cur_nimg - tick_start_nimg) * 1e3):<7.2f}"]
        #fields += [f"maintenance {training_stats.report0('Timing/maintenance_sec', maintenance_time):<6.1f}"]
        #fields += [f"cpumem {training_stats.report0('Resources/cpu_mem_gb', psutil.Process(os.getpid()).memory_info().rss / 2**30):<6.2f}"]
        #fields += [f"gpumem {training_stats.report0('Resources/peak_gpu_mem_gb', torch.cuda.max_memory_allocated(device) / 2**30):<6.2f}"]
        fields += [f"sparse {loss_dict['loss_sparse']:.3f}"]
        fields += [f"entropy {loss_dict['loss_entropy']:.3f}"]
        fields += [f"path {loss_dict['loss_path']:.3f}"]
        fields += [f"equal {loss_dict['loss_equal']:.3f}"]
        fields += [f"rowvar {loss_dict['loss_rowvar']:.3f}"]
        fields += [f"colvar {loss_dict['loss_colvar']:.3f}"]
        fields += [f"lambda_sparse {cur_lambda_sparse:.3f}"]
        fields += [f"lambda_entropy {cur_lambda_entropy:.3f}"]
        fields += [f"lambda_rowvar {cur_lambda_rowvar:.3f}"]
        fields += [f"lambda_colvar {cur_lambda_colvar:.3f}"]
        fields += [f"lambda_path {lambda_path:.3f}"]
        fields += [f"lambda_equal {lambda_equal:.3f}"]
        fields += [f"thr {cur_entropy_thr:.3f}"]
        torch.cuda.reset_peak_memory_stats()
        #fields += [f"augment {training_stats.report0('Progress/augment', float(augment_pipe.p.cpu()) if augment_pipe is not None else 0):.3f}"]
        training_stats.report0('Timing/total_hours', (tick_end_time - start_time) / (60 * 60))
        training_stats.report0('Timing/total_days', (tick_end_time - start_time) / (24 * 60 * 60))
        if rank == 0:
            print(' '.join(fields))

        # Check for abort.
        if (not done) and (abort_fn is not None) and abort_fn():
            done = True
            if rank == 0:
                print()
                print('Aborting...')


        # Save image snapshot.
        if (rank == 0) and (image_snapshot_ticks is not None) and (done or cur_tick % image_snapshot_ticks == 0):
            wss = torch.cat([G_ema.mapping(z,M_ema(c)) for z,c in zip(grid_z, grid_c)])
            images = torch.cat([G_ema(z=z, c=M_ema(c), noise_mode='const').cpu() for z, c in zip(grid_z, grid_c)])

            def normalize_2nd_moment(x, dim=1, eps=1e-8):
                return x * (x.square().mean(dim=dim, keepdim=True) + eps).rsqrt()

            cs = []
            for c in grid_c:
                cs.append(c.argmax(dim=1))
            cs = torch.cat(cs, 0).view(G.mapping.c_dim, -1)
            tmp_imgs = images.reshape(G.mapping.c_dim, -1, images.shape[1], images.shape[2], images.shape[3])
            images = images.numpy()
            wss = wss.reshape(G.mapping.c_dim, -1, wss.shape[1], wss.shape[2])
            print(cs.size(), tmp_imgs.shape, wss.shape, ' >>>>>cs size tmp_imgs size <<<<<<<<')
            save_image_grid(images, os.path.join(run_dir, f'fakes{cur_nimg//1000:06d}.jpg'), drange=[-1,1], grid_size=grid_size)
            try:
                print(G_ema.mapping.importance0, G_ema.mapping.importance1)
            except:
                pass
            all_masks = []
            with torch.no_grad():
                cin = torch.arange(G.mapping.c_dim, device=device)
                cin = F.one_hot(cin, num_classes=G.mapping.c_dim).float()

                all_logit = M(cin)
                all_soft_mask = ((all_logit))
                all_hard_mask = (all_soft_mask > 0.5).float()

                for i in range(G.mapping.c_dim):
                    print('%40s' % names[i], '    ', all_soft_mask[i].cpu().numpy())
                for i in range(G.mapping.c_dim):
                    print('%40s' % names[i], '    ', all_hard_mask[i].cpu().numpy().astype(np.uint8))

                all_logit = M_ema(cin)
                all_soft_mask = ((all_logit))
                all_hard_mask = (all_soft_mask > 0.5).float()

                for i in range(G.mapping.c_dim):
                    print('%40s' % names[i], '    ', all_soft_mask[i].cpu().numpy())
                for i in range(G.mapping.c_dim):
                    print('%40s' % names[i], '    ', all_hard_mask[i].cpu().numpy().astype(np.uint8))


                dscores = []
                dhard_masks = all_hard_mask.clone()
                dsoft_masks = all_soft_mask.clone()
                for i in range(G.mapping.c_dim):
                    cur_imgs = tmp_imgs[i].to(device)
                    cur_c = F.one_hot(torch.tensor([i]*cur_imgs.size(0), device=device), num_classes=G.mapping.c_dim).float().to(device)
                    d_out = D(cur_imgs, cur_c)
                    d_out = F.softplus(d_out)
                    print('%40s mean: %.2f min: %.2f max: %.2f' % (names[i], d_out.mean().item(), d_out.min().item(), d_out.max().item()))
                    dscores.append(d_out.min().item())

                #eval_mask = M(cin, eval=True)
                #for i in range(G.mapping.c_dim):
                #    print('%10s' % names[i], '    ', eval_mask[i].cpu().numpy().astype(np.uint8))

                def normalize_2nd_moment(x, dim=1, eps=1e-8):
                    return x * (x.square().mean(dim=dim, keepdim=True) + eps).rsqrt()

                def get_onehot(y):
                    shape = y.size()
                    _, ind = y.max(dim=-1)
                    y_hard = torch.zeros_like(y).view(-1, shape[-1])
                    y_hard.scatter_(1, ind.view(-1, 1), 1)
                    y_hard = y_hard.view(*shape)
                    return y_hard


                def no_same_rows(x):
                    has = False
                    for i in range(len(x)):
                        for j in range(i+1, len(x)):
                            if torch.allclose(x[i], x[j]):
                                has = True
                    return not has

                def has_enough_concepts(x):
                    has = True
                    for i in range(len(x)):
                        if torch.sum(x[i])<=1:
                            has = False
                    return has


                if no_same_rows(dhard_masks) and has_enough_concepts(dhard_masks):
                    print('')
                    print('>>>>>>>>>>>>> This version can be used <<<<<<<<<<<<<<')
                    print('')
                    ready = True
                    binary_mask_checker.add(dhard_masks)

                try:
                    best_mask = binary_mask_checker.check()
                    for i in range(G.mapping.c_dim):
                        print('%40s' % names[i], '    ', best_mask[i].cpu().numpy().astype(np.uint8), '  best')
                except:
                    pass



                masks = all_soft_mask
                hard_masks = all_hard_mask
                for i in range(G.mapping.i_dim):
                    cur_i_imgs = []
                    sorted_index = np.argsort(masks[:, i].cpu().numpy(), axis=0)[::-1]
                    for j in sorted_index:
                        if hard_masks[j, i] == 1:
                            cur_i_imgs.append(tmp_imgs[j])
                    if len(cur_i_imgs) > 0:
                        cur_i_imgs = torch.cat(cur_i_imgs, 0)
                        save_image(cur_i_imgs, os.path.join(run_dir, f'concept_{cur_nimg // 1000:06d}_{i}.jpg'),
                                       nrow=grid_size[0], normalize=True, range=(-1, 1))

            if True:
                for i in range(G.mapping.c_dim):
                    if False:
                        M.param_net.data[i] += -1e9*(dsoft_masks[i]<0.05)
                        M_ema.param_net.data[i] += -1e9*(dsoft_masks[i]<0.05)
                        M.use_param[i] = (dsoft_masks[i]<0.05).float()
                        M_ema.use_param[i] = (dsoft_masks[i]<0.05).float()
                        #print(dscores[i], names[i], ' >>>>>>. what fuck ', M.use_param.view(-1), M.param_net[i])


                #topk = torch.topk(torch.tensor(dscores), k=5)[1]
                consistency = version_history_checker.check_history(dhard_masks)
                version_history_checker.update_history(dhard_masks)

                for i in range(G.mapping.c_dim):
                    all_sum = torch.sum(dhard_masks, dim=1)
                    target = torch.mode(all_sum)[0]
                    cur_sum = all_sum[i]

                    set_thr = 1.0
                    cond1 = (dscores[i]>=set_thr)
                    crit = (cur_sum>1 and cur_sum<=target)
                    #cond2 = (dscores[i]>=0.6 and cur_sum>1 and cur_sum<=target and (i in list(topk.cpu())))
                    cond3 = consistency[i]


                    should_use=True
                    for j in range(G.mapping.c_dim):
                        if dscores[j]> dscores[i] and torch.sum(torch.abs(dhard_masks[i]-dhard_masks[j]))==0 and j!=i:
                            should_use =  False

                    if (cond1) and should_use and crit:
                        #M.param_net.data[i] = 1e9*dhard_masks[i]
                        #M.param_net.data[i] += -1e9*(1-dhard_masks[i])
                        M.target_value[i] = dhard_masks[i]
                        M.use_param[i] = torch.ones_like(M.use_param[i])

                        #M_ema.param_net.data[i] = 1e9*dhard_masks[i]
                        #M_ema.param_net.data[i] += -1e9*(1-dhard_masks[i])
                        M_ema.target_value[i] = dhard_masks[i]
                        M_ema.use_param[i] = torch.ones_like(M.use_param[i])
                        print('>>>>>> replace classss ', names[i], '   ', dscores[i], '    ', M.target_value[i], ' << consistency ', consistency[i])



                column_consistency = column_history_cheker.check_history(dhard_masks)
                column_history_cheker.update_history(dhard_masks)

                for j in range(G.mapping.i_dim):
                    cur_soft = dsoft_masks[:,j]
                    cur_hard = dhard_masks[:,j]
                    act = cur_soft[cur_hard==1]
                    deact = cur_soft[cur_hard==0]
                    cur_sum = torch.sum(cur_hard)
                    if (act.mean()>0.9 and act.min()>0.6 and cur_sum>1 and cur_tick==5):
                        #M.param_net.data[:,j] = cur_hard*19
                        #M.param_net.data[:,j] += -1e19*(1-cur_hard)
                        M.use_param[:,j] = torch.ones_like(M.use_param[:,j])
                        M.target_value[:,j] = cur_hard

                        #M_ema.param_net.data[:,j] = cur_hard
                        #M_ema.param_net.data[:,j] += -1e19*(1-cur_hard)
                        M_ema.target_value[:,j] = cur_hard
                        M_ema.use_param[:,j] = torch.ones_like(M.use_param[:,j])
                        print('>>>>> replace columns ', j, '   ', M.target_value[:,j].view(-1), '    ', column_consistency[j])

                if cur_tick == use_best_binary:
                    best_mask = binary_mask_checker.check()
                    if best_mask is not None:
                        M.use_param = torch.ones_like(M.use_param)
                        M.target_value = best_mask
                        M_ema.use_param = torch.ones_like(M.use_param)
                        M_ema.target_value = best_mask


        if (cur_tick % 5 ==0 and cur_tick>0) or cur_tick == use_best_binary:
            for param in M.parameters():
                torch.distributed.broadcast(param.data, 0)
            torch.distributed.broadcast(M.use_param, 0)
            torch.distributed.broadcast(M_ema.use_param, 0)
            torch.distributed.broadcast(M.target_value, 0)
            torch.distributed.broadcast(M_ema.target_value, 0)
            for param in M_ema.parameters():
                torch.distributed.broadcast(param.data, 0)

            torch.distributed.barrier()

            #print(M.use_param, ' >>>>>>> m M use_oaramssss bripdcatss    ')

        # Save network snapshot.
        snapshot_pkl = None
        snapshot_data = None
        if (network_snapshot_ticks is not None) and (done or cur_tick % network_snapshot_ticks == 0) and cur_tick>0:
            snapshot_data = dict(training_set_kwargs=dict(training_set_kwargs))
            for name, module in [('G', G), ('D', D), ('G_ema', G_ema), ('augment_pipe', augment_pipe), ('M', M), ('M_ema', M_ema)]:
                if module is not None:
                    if num_gpus > 1:
                        misc.check_ddp_consistency(module, ignore_regex=r'.*\.w_avg')
                    module = copy.deepcopy(module).eval().requires_grad_(False).cpu()
                snapshot_data[name] = module
                del module # conserve memory
            snapshot_pkl = os.path.join(run_dir, f'network-snapshot-{cur_nimg//1000:06d}.pkl')
            if rank == 0:
                #pass
                with open(snapshot_pkl, 'wb') as f:
                    pickle.dump(snapshot_data, f)

        # Evaluate metrics.
        if (snapshot_data is not None) and (len(metrics) > 0):
            if rank == 0:
                print('Evaluating metrics...')
            for metric in metrics:
                result_dict = metric_main.calc_metric(metric=metric, G=snapshot_data['G_ema'], M=snapshot_data['M_ema'],
                    dataset_kwargs=training_set_kwargs, num_gpus=num_gpus, rank=rank, device=device)
                if rank == 0:
                    metric_main.report_metric(result_dict, run_dir=run_dir, snapshot_pkl=snapshot_pkl)
                stats_metrics.update(result_dict.results)
        del snapshot_data # conserve memory

        # Collect statistics.
        for phase in phases:
            value = []
            if (phase.start_event is not None) and (phase.end_event is not None):
                phase.end_event.synchronize()
                value = phase.start_event.elapsed_time(phase.end_event)
            training_stats.report0('Timing/' + phase.name, value)
        stats_collector.update()
        stats_dict = stats_collector.as_dict()

        # Update logs.
        timestamp = time.time()
        if stats_jsonl is not None:
            fields = dict(stats_dict, timestamp=timestamp)
            stats_jsonl.write(json.dumps(fields) + '\n')
            stats_jsonl.flush()
        if stats_tfevents is not None:
            global_step = int(cur_nimg / 1e3)
            walltime = timestamp - start_time
            for name, value in stats_dict.items():
                stats_tfevents.add_scalar(name, value.mean, global_step=global_step, walltime=walltime)
            for name, value in stats_metrics.items():
                stats_tfevents.add_scalar(f'Metrics/{name}', value, global_step=global_step, walltime=walltime)
            stats_tfevents.flush()
        if progress_fn is not None:
            progress_fn(cur_nimg // 1000, total_kimg)

        # Update state.
        if False and cur_tick%5==0:
            for paramgroup in M_opt.param_groups:
                paramgroup['lr'] = paramgroup['lr'] * 0.1
            print('>>>>>>>LR decay <<<<<<< %.7f' % paramgroup['lr'])
        cur_tick += 1
        tick_start_nimg = cur_nimg
        tick_start_time = time.time()
        maintenance_time = tick_start_time - tick_end_time
        if done:
            break

    # Done.
    if rank == 0:
        print()
        print('Exiting...')

#----------------------------------------------------------------------------