base.py

import torch
import torch.nn as nn
import torch.optim as optim
import torch.nn.functional as F
import torch.nn.init as init
import torch.backends.cudnn as cudnn
from torch.utils.data.dataset import Dataset
from torch.optim.lr_scheduler import MultiStepLR

import torch_optimizer


import torch.distributed as dist
import torch.multiprocessing as mp
import torch.nn.parallel
import torch.utils.data
import torch.utils.data.distributed
from torch.utils.data.distributed import DistributedSampler


from PIL import Image

import os
import sys
import time
import types
import numpy as np
import shutil
import warnings
import random
import math

import h5py

from framework.config import get_config, get_arch, get_dataset, get_transform, get_pin_memory, DistillDataset
from framework.distill_higher import Distill
from framework.util import Summary, AverageMeter, ProgressMeter, accuracy, accuracy_ind, ImageIntervention, init_gaussian

from statistics import mean
import wandb

os.environ['MASTER_ADDR'] = 'localhost'
os.environ['MASTER_PORT'] = '5678'


def main_worker(gpu, ngpus_per_node, args):
    
    global best_acc1, best_loss1
    best_acc1 = 0
    best_loss1 = 1000
    args.gpu = gpu
    if args.gpu is not None:
        print("Use GPU: {} for training".format(args.gpu))
    if args.distributed:
        if args.dist_url == "env://" and args.rank == -1:
            args.rank = int(os.environ["RANK"])
        if args.mp_distributed:
            # For multiprocessing distributed training, rank needs to be the
            # global rank among all the processes
            args.rank = args.rank * ngpus_per_node + gpu
        
        dist.init_process_group(backend=args.dist_backend,
                                world_size=args.world_size, rank=args.rank)

    torch.manual_seed(args.rank + args.seed)
    np.random.seed(args.rank + args.seed)
    cudnn.benchmark = True
    cudnn.deterministic = True
    print("Torch Seed Specified with rank: %d"%(args.rank+args.seed))
    args.data_root = os.path.join(args.root, args.dataset)
    print("Dataset: %s"%args.dataset)
    print("Dataset Path: %s"%args.root)
    print(args)
    
    pathdir = './train_log/{}/{}_{}_Adam{}'.format(args.dataset, args.epochs, args.batch_size, int(args.lr*1000)) 
        
    # 0. Preprocess datasets
    print('==> Preparing data..')
    transform_train, transform_test = get_transform(args.dataset)
    print(transform_train, transform_test)

    train1, train2, testset, num_classes, shape, _ = get_dataset(args.dataset, args.data_root, transform_train, transform_test, zca=args.zca)
    print('Dataset: number of classes: {}'.format(num_classes))
    args.num_classes = num_classes
    if args.limit_train: train1 = torch.utils.data.Subset(train1, range(5000))
    print('Training set size: {}'.format(len(train1)))
    # Not finished! Use DDP for dataloader
    # Since difference processes are intialized with different random seeds, we do not need to use different data
    # We use the same data for all processes, without defining the sampler.
    if args.distributed:
        train_sampler = None #DistributedSampler(train1)
    #    train_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset)
    #    val_sampler = torch.utils.data.distributed.DistributedSampler(val_dataset, shuffle=False, drop_last=True)
    else:
        train_sampler = None
    val_sampler = None
    
    train_loader = torch.utils.data.DataLoader(
            train1, batch_size=args.batch_size, shuffle=(train_sampler is None),
            num_workers=args.workers, pin_memory=True, sampler=train_sampler)
    test_loader = torch.utils.data.DataLoader(
            testset, batch_size=args.batch_size, shuffle=False,
            num_workers=args.workers, pin_memory=True, sampler=val_sampler)
    
    # Initialize Distilled Dataset
    channel, height, width = shape
    print('Image size: channel {}, height {}, width {}'.format(channel, height, width))
    x_init = torch.randn(args.num_per_class * num_classes, channel*height*width)
    y_list = list(range(0, num_classes)) * args.num_per_class
    y_list.sort()
    y_init = torch.tensor(y_list)
    y_tmp = torch.zeros(num_classes)
    # if train_y with soft label, change it to one_hot.   
    if args.train_y:
        y_init = F.one_hot(y_init, num_classes)
    
    # Data augmentations for synthetic data and the real data
    syn_intervention, real_intervention, interv_prob = set_up_interventions(args)
    print('Synthetic images, not_single {}, keys {}'.format(syn_intervention.not_single, syn_intervention.keys))
    args.syn_intervention = syn_intervention
    args.real_intervention = real_intervention
    args.shape = shape
    
    # 1. Initialize Distilled Dataset Module
    print('==> Building model..')
    print('Initialized data with size, x: {}, y:{}'.format(x_init.shape, y_init.shape))
    model = Distill(x_init, y_init, args.arch, args.window, args.inner_lr, args.num_train_eval, img_pc=args.num_per_class, 
                    batch_pc=args.batch_per_class, train_y=args.train_y, 
                    channel=shape[0], num_classes=num_classes, task_sampler_nc=args.task_sampler_nc, 
                    im_size=(shape[1], shape[2]), inner_optim=args.inner_optim, 
                    syn_intervention=syn_intervention, real_intervention=real_intervention, cctype=args.cctype)
    print(model.net)
    
    # The data is intialized with random values, not from real images to remove any bias
    # We project the data to a unit ball to control the norm.
    model.data.weight.data = project(model.data.weight)
    
    if not torch.cuda.is_available():
        print('using CPU, this will be slow')
    elif args.distributed:
        if torch.cuda.is_available():
            if args.gpu is not None:
                torch.cuda.set_device(args.gpu)
                model.cuda(args.gpu)
                if not args.train_y: model.label = model.label.cuda(args.gpu)
                # we could use the same data across all gpus?
                args.batch_size = int(args.batch_size / ngpus_per_node)
                args.workers = int((args.workers + ngpus_per_node - 1) / ngpus_per_node)
                model = torch.nn.parallel.DistributedDataParallel(model, device_ids=[args.gpu], find_unused_parameters=True)
            else:
                model.cuda()
                if not args.train_y: model.label = model.label.cuda()
                # DistributedDataParallel will divide and allocate batch_size to all
                # available GPUs if device_ids are not set
                model = torch.nn.parallel.DistributedDataParallel(model, find_unused_parameters=True)
    elif args.gpu is not None and torch.cuda.is_available():
        torch.cuda.set_device(args.gpu)
        model = model.cuda(args.gpu)
        if not args.train_y: model.label = model.label.cuda(args.gpu)
        print('use direct cuda')
    else:
        # DataParallel will divide and allocate batch_size to all available GPUs
        print('use data parallel only')
        if not args.train_y: model.label = model.label.cuda()
        model = torch.nn.DataParallel(model).cuda()
    
    if torch.cuda.is_available():
        if args.gpu:
            device = torch.device('cuda:{}'.format(args.gpu))
        else:
            device = torch.device("cuda")
    elif torch.backends.mps.is_available():
        device = torch.device("mps")
    else:
        device = torch.device("cpu")
        
    criterion = nn.CrossEntropyLoss(reduction='mean').to(device)
    model.module.dd_type = args.ddtype
    continue_training = False
    
    # if ckptname is given, load the data from the file as initialization
    if args.ckptname != 'none':
        db = h5py.File(args.ckptname, 'r')
        print(db['data'].shape[0], int(x_init.shape[0]/num_classes), args.num_per_class)
        base_data = torch.tensor(db['data'][:]).cuda()
        if args.train_y:
            label_data = torch.tensor(db['label'][:]).cuda()
            model.module.label.weight.data = label_data
        continue_training = True
        
    start_test_epoch = 0 if continue_training else 1
        
    if args.train_y:
        if args.outer_optim == 'Adam':
            optimizer = optim.Adam([{'params': model.module.data.weight}, 
                                   {'params': model.module.label.weight, 'lr': args.lr/args.label_lr_scale}], 
                                   lr=args.lr, betas=(0.9, 0.999), eps=args.eps, weight_decay=args.wd)
        else:
            raise NotImplementedError()
    else:
        #optimizer = optim.SGD([model.module.data], lr=args.lr, momentum=0.5, weight_decay=0)
        optimizer = optim.Adam([model.module.data.weight], lr=args.lr, betas=(0.9, 0.999), eps=args.eps, weight_decay=args.wd)
        
        
    # how to visualize in DDP?
    best_rec = {}
    grad_acc = []
    best_loss_ind = 0
    
    distill_steps = 0
    if args.ddtype == 'curriculum' and args.cctype != 2: 
        model.module.curriculum = [args.totwindow-args.window, args.minwindow, 0, 0, 0, 0][args.cctype]
    
    if model.module.data.weight.get_device() == 0 and args.wandb:
        wandb.init(
                # set the wandb project where this run will be logged
                project="Distributed-Distillation",
                name = args.name,
                config=vars(args),
                settings=wandb.Settings(start_method="fork")) 
    
    if args.ddtype == 'standard':
        fname = './grad_save_init_IPC_'+str(args.num_per_class)+'_no_curr_unroll_'+str(args.window)+args.fname+'.h5'
    else:
        fname = './save/{}/IPC_{}_{}_curr_unroll_{}_{}_{}_{}.h5'.format(str(args.dataset), str(args.num_per_class),
                                                                                                               str(args.cctype), str(args.window), 
                                                                                                       str(args.totwindow), str(args.minwindow), args.fname)
    
    if args.load_ckpt:
        checkpoint = torch.load(fname[:-3]+'.pth')
        # Load the model and optimizer state_dicts
        base_data = checkpoint['model_state_dict']['module.data.weight'].cuda()
        model.module.data.weight.data = base_data
        if args.train_y:
            label_data = checkpoint['label_state_dict']['module.label.weight'].cuda()
            model.module.label.weight.data = label_data
        if not args.train_y: 
            optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
            args.start_epoch = checkpoint['epoch']
            distill_steps = args.start_epoch * int(50000 / args.batch_size)
        print('=== Successfully loading the data from {} ==='.format(fname[:-3]+'.pth'))
        model.module.ema_init(args.clip_coef)
        ### TO DO: add EMA loading for distilled data

    for epoch in range(args.start_epoch, args.epochs):
        # initialize the EMA
        if epoch == 0: model.module.ema_init(args.clip_coef)  
        grad_tmp, losses_avg, distill_steps = train(train_loader, model, criterion, optimizer, epoch, device, distill_steps, args)
        grad_acc.append(grad_tmp)
        print('The current update step is {}'.format(distill_steps))

        # evaluate on validation set
        if (epoch - args.start_epoch + start_test_epoch) % args.test_freq == 0:
            if model.module.data.weight.get_device() == 0: print('The current seed is {}'.format(torch.seed()))
            if model.module.data.weight.get_device() == 0: print('The current lr is: {}'.format(model.module.lr))
            if model.module.data.weight.get_device() == 0: print('Testing Results:')
            test_acc, test_loss, scores = test([test_loader, train_loader], model, criterion, args)
            if model.module.data.weight.get_device() == 0: print(test_acc)
            tmp_index = test_acc[0].index(max(test_acc[0]))
                        
            if model.module.data.weight.get_device() == 0 and args.wandb:
                wandb.log({"loss": test_loss, "epoch": int(epoch), 'distill_steps':distill_steps, "grad_norm": grad_tmp[-1],
                        "train_acc": test_acc[1][-1], "test_acc":test_acc[0][-1], "curr": model.module.curriculum})
            
        # remember best acc@1 and save checkpoint
            is_best = test_acc[0][tmp_index] > best_acc1
            if is_best:
                best_acc1 = max(test_acc[0][tmp_index], best_acc1)
                if model.module.data.weight.get_device() == 0:
                    best_rec['acc'] = test_acc[0][tmp_index]
                    best_rec['test'] = test_acc[0]
                    best_rec['train'] = test_acc[0]
                    best_rec['ind'] = tmp_index
                    best_rec['epoch'] = epoch + 1
                    best_rec['data'] = model.module.data.weight.data.cpu().clone().numpy()
                    if args.train_y:
                        best_rec['label'] = model.module.label.weight.data.cpu().clone().numpy()
            
            if model.module.data.weight.get_device() == 0:
                with h5py.File(fname, 'w') as f:
                    f.create_dataset('data', data=best_rec['data'])
                    f.create_dataset('epoch', data=best_rec['epoch'])
                    if args.train_y:
                        f.create_dataset('label', data=best_rec['label'])
                checkpoint = {
                    'epoch': epoch,
                    'model_state_dict': {'module.data.weight': model.state_dict()['module.data.weight'].cpu().clone()},
                    'optimizer_state_dict': optimizer.state_dict(),
                }
                if args.train_y:
                    checkpoint['label_state_dict'] = {'module.label.weight': model.state_dict()['module.label.weight'].cpu().clone()}
                torch.save(checkpoint, fname[:-3]+'.pth')
            # update curriculum
            
            if test_loss < best_loss1:
                best_loss1 = test_loss
                best_loss_ind = epoch
            else:
                if epoch >= best_loss_ind + 60:
                    best_loss_ind = epoch
                    if args.ddtype == 'curriculum': 
                        if args.cctype == 0:
                            if model.module.curriculum == args.minwindow: break
                        elif args.cctype == 1:
                            if model.module.curriculum == args.totwindow-args.window: break
                            model.module.curriculum += args.window
                            model.module.curriculum = min(args.totwindow-args.window, model.module.curriculum)
            print('train loss {}, epoch {}, best loss {}, best_epoch {}'.format(test_loss, epoch, best_loss1, best_loss_ind))
            

    if model.module.data.weight.get_device() == 0: 
        print('=== Final results:')
        print(best_rec)
        if args.wandb: wandb.log({"best_acc": best_rec['acc']})
    grad_acc = np.concatenate(grad_acc)
    if model.module.data.weight.get_device() == 0:
        with h5py.File(fname, 'w') as f:
            f.create_dataset('grad', data=grad_acc)
            best_rec['data'] = model.module.data.weight.data.cpu().clone().numpy()
            f.create_dataset('data', data=best_rec['data'])
    
def train(train_loader, model, criterion, optimizer, epoch, device, distill_steps, args):
    batch_time = AverageMeter('Time', ':6.3f')
    data_time = AverageMeter('Data', ':6.3f')
    losses = AverageMeter('Loss', ':.4e')
    top1 = AverageMeter('Acc@1', ':6.2f')
    progress = ProgressMeter(
        len(train_loader),
        [batch_time, data_time, losses, top1],
        prefix="Epoch: [{}]".format(epoch))

    # switch to train mode
    model.train()
    model.module.net.train()

    end = time.time()
    grad_acc = []
    if model.module.cctype == 2:
        if args.rank == 0:
            shared_curriculum = torch.tensor(random.randint(args.minwindow, args.totwindow-args.window))
        else:
            shared_curriculum = torch.tensor(0)
        shared_curriculum = shared_curriculum.to(device)
        if args.mp_distributed:
            dist.broadcast(shared_curriculum, src=0)
        
        model.module.curriculum = shared_curriculum.item()

        
    print('GPU_{}_using curriculum {} with window {}'.format(args.rank, model.module.curriculum, model.module.window))
    
    
    for train1 in enumerate(train_loader):
        
        if args.complete_random:
            if model.module.cctype == 2:
                if args.rank == 0:
                    shared_curriculum = torch.tensor(random.randint(args.minwindow, args.totwindow-args.window))
                else:
                    shared_curriculum = torch.tensor(0)
                shared_curriculum = shared_curriculum.to(device)
                if args.mp_distributed:
                    dist.broadcast(shared_curriculum, src=0)

                model.module.curriculum = shared_curriculum.item()

        
            print('GPU_{}_using curriculum {} with window {}'.format(args.rank, model.module.curriculum, model.module.window))
            
        data_time.update(time.time() - end)
        
        i, (inputs, targets) = train1
        inputs = inputs.to(device, non_blocking=True)
        targets = targets.to(device, non_blocking=True)
        output, _ = model(inputs)

        loss = criterion(output, targets)

        acc = accuracy(output, targets)
        losses.update(loss.item(), inputs.size(0))
        top1.update(acc, inputs.size(0))

        # compute gradient and do SGD step
        optimizer.zero_grad()
        loss.backward()
        for clear_cache in range(5):
            torch.cuda.empty_cache()
                      
        grad_norm = calculate_grad_norm(torch.norm(optimizer.param_groups[0]['params'][0].grad.clone().detach(), dim=1))

        grad_acc.append(grad_norm)
        # obtain the ema norm and perform gradient clipping
        clip_coef = model.module.ema_update((torch.norm(optimizer.param_groups[0]['params'][0].grad.clone().detach(), dim=1)**2).sum().item() ** 0.5)

        torch.nn.utils.clip_grad_norm_(
                    model.module.data.weight, max_norm=clip_coef * 2)
        
        
        optimizer.step()


        optimizer.zero_grad()
        model.module.net.zero_grad()
        
        if args.train_y:
             with torch.no_grad():
                model.module.label.weight.data = torch.clip(model.module.label.weight.data, min=0)

        # measure elapsed time
        batch_time.update(time.time() - end)
        end = time.time()

        distill_steps += 1
        torch.cuda.empty_cache()
        
        # print the results
        if (i+1) % args.print_freq == 0 and model.module.data.weight.get_device() == 0:
            progress.display(i+1)
    return grad_acc, losses.avg, distill_steps

# use pair_aug with train will apply a deterministic augmentation for all the data
def set_up_interventions(args):
    syn_intervention  = ImageIntervention(
                             'syn_aug',
                             args.syn_strategy,
                             phase='test',
                             not_single=args.comp_aug
                         )
    real_intervention = ImageIntervention(
                             'real_aug',
                             args.real_strategy,
                             phase='test',
                             not_single=args.comp_aug_real
                         )
    # This is a customizable prob \in [0, 1]
    intervention_prob  = 1.0

    return syn_intervention, real_intervention, intervention_prob


def calculate_grad_norm(grad_norm):
    return grad_norm[grad_norm>1e-5].mean().item()

# testing function for train_y = True
def one_gpu_test(val_loader, model, args):
    cross_entropy_loss = nn.CrossEntropyLoss(reduction='none')
    def run_validate(loader, base_progress=0):
        acc_ind = []
        with torch.no_grad():
            end = time.time()
            for i, (images, target) in enumerate(loader):
                i = base_progress + i
                
                if args.gpu is not None and torch.cuda.is_available():
                    images = images.cuda(args.gpu, non_blocking=True)
                    target = target.cuda(args.gpu, non_blocking=True)
                elif torch.cuda.is_available():
                    images = images.cuda()
                    target = target.cuda()

                output, _ = model.module.test(images)
                sample_wise_loss = cross_entropy_loss(output, target)
                acc_ind.append(sample_wise_loss)
        return torch.cat(acc_ind, 0)

    acc_ind = run_validate(val_loader)
    #print(acc_ind.shape)
    return acc_ind

# testing function for not training labels
def one_gpu_test_2(val_loader, model, args):
    def run_validate(loader, base_progress=0):
        acc_ind = []
        with torch.no_grad():
            end = time.time()
            for i, (images, target) in enumerate(loader):
                i = base_progress + i
                
                if args.gpu is not None and torch.cuda.is_available():
                    images = images.cuda(args.gpu, non_blocking=True)
                    target = target.cuda(args.gpu, non_blocking=True)
                elif torch.cuda.is_available():
                    images = images.cuda()
                    target = target.cuda()

                output, _ = model.module.test(images)
                if len(target.shape) == 2: target = target.max(1)[1]
                acc_ind.append(accuracy_ind(output, target))
        return torch.cat(acc_ind, 0)

    acc_ind = run_validate(val_loader)
    return acc_ind.to(torch.int)
                                        
# project the data to a unit ball
def project(data, pgd_coef=1):
    coef_norm = 1 / math.sqrt(data.shape[1])
    data_norm = torch.reshape(
        torch.norm(torch.flatten(data, start_dim=1, end_dim=-1), dim=-1),
        [data.shape[0], *[1] * (data.dim() - 1)])
    return data / data_norm * pgd_coef

def test(data_loaders, model, criterion, args):
    epoch_list = [300, 600, 1000, 3000]
    acc = []
    for i in range(len(data_loaders)): acc.append([0] * (len(epoch_list)))
    loss = 0
    
    for train_ind in range(args.num_train_eval):
        model.module.init_train(0, init=True)
        start_epoch = 0
        for train_time in range(len(epoch_list)):
            model.train()
            model.module.net.train()
            model.module.init_train(epoch_list[train_time] - start_epoch)
            for loader_i in range(len(data_loaders)):
                tmp_acc, tmp_loss = default_test(data_loaders[loader_i], model, criterion, args)
                acc[loader_i][train_time] += tmp_acc
            start_epoch = epoch_list[train_time]
        loss += tmp_loss
        if train_ind == 0:
            acc_ind = one_gpu_test(data_loaders[0], model, args)
        else:
            acc_ind += one_gpu_test(data_loaders[0], model, args)
    acc_ind = args.num_train_eval - acc_ind
        
    for loader_i in range(len(data_loaders)):
        acc[loader_i] = [acc_id/args.num_train_eval for acc_id in acc[loader_i]]
        if model.module.data.weight.get_device() == 0:  
            for train_time in range(len(epoch_list)):
                if model.module.data.weight.get_device() == 0:  
                    print('Training for {} epoch: {}'.format(epoch_list[train_time], acc[loader_i][train_time]))
    return acc, tmp_loss / args.num_train_eval, acc_ind

def default_test(val_loader, model, criterion, args):
    def run_validate(loader, base_progress=0):
        with torch.no_grad():
            end = time.time()
            for i, (images, target) in enumerate(loader):
                i = base_progress + i
                
                if args.gpu is not None and torch.cuda.is_available():
                    images = images.cuda(args.gpu, non_blocking=True)
                    target = target.cuda(args.gpu, non_blocking=True)
                elif torch.cuda.is_available():
                    images = images.cuda()
                    target = target.cuda()

                output, _ = model.module.test(images)
                loss = criterion(output, target)

                # measure accuracy and record loss
                if len(target.shape) == 2: target = target.max(1)[1]
                acc = accuracy(output, target, topk=(1, 5))
                losses.update(loss.item(), images.size(0))
                top1.update(acc, images.size(0))

                # measure elapsed time
                batch_time.update(time.time() - end)
                end = time.time()

                #if (i+1) % args.print_freq == 0 and model.module.data.get_device() == 0:
                #    progress.display(i + 1)

    batch_time = AverageMeter('Time', ':6.3f', Summary.NONE)
    losses = AverageMeter('Loss', ':.4e', Summary.NONE)
    top1 = AverageMeter('Acc@1', ':6.2f', Summary.AVERAGE)
    progress = ProgressMeter(
        len(val_loader) + (args.distributed and (len(val_loader.sampler) * args.world_size < len(val_loader.dataset))),
        [batch_time, losses, top1],
        prefix='Test: ')

    run_validate(val_loader)
    if args.distributed:
        top1.all_reduce()
        losses.all_reduce()

    if args.distributed and (len(val_loader.sampler) * args.world_size < len(val_loader.dataset)):
        aux_val_dataset = Subset(val_loader.dataset,
                                 range(len(val_loader.sampler) * args.world_size, len(val_loader.dataset)))
        aux_val_loader = torch.utils.data.DataLoader(
            aux_val_dataset, batch_size=args.batch_size, shuffle=False,
            num_workers=args.workers, pin_memory=True)
        run_validate(aux_val_loader, len(val_loader))
    if model.module.data.weight.get_device() == 0: progress.display_summary()

    return top1.avg, losses.avg