xautodl/exps/search-transformable.py

#####################################################
# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2019.01 #
#######################################################################
# Network Pruning via Transformable Architecture Search, NeurIPS 2019 #
#######################################################################
import sys, time, torch, random, argparse
from PIL     import ImageFile
from os      import path as osp
ImageFile.LOAD_TRUNCATED_IMAGES = True
import numpy as np
from copy    import deepcopy
from pathlib import Path

lib_dir = (Path(__file__).parent / '..' / 'lib').resolve()
if str(lib_dir) not in sys.path: sys.path.insert(0, str(lib_dir))
from config_utils import load_config, configure2str, obtain_search_args as obtain_args
from procedures   import prepare_seed, prepare_logger, save_checkpoint, copy_checkpoint
from procedures   import get_optim_scheduler, get_procedures
from datasets     import get_datasets, SearchDataset
from models       import obtain_search_model, obtain_model, change_key
from utils        import get_model_infos
from log_utils    import AverageMeter, time_string, convert_secs2time


def main(args):
  assert torch.cuda.is_available(), 'CUDA is not available.'
  torch.backends.cudnn.enabled   = True
  torch.backends.cudnn.benchmark = True
  #torch.backends.cudnn.deterministic = True
  torch.set_num_threads( args.workers )
  
  prepare_seed(args.rand_seed)
  logger = prepare_logger(args)
  
  # prepare dataset
  train_data, valid_data, xshape, class_num = get_datasets(args.dataset, args.data_path, args.cutout_length)
  #train_loader = torch.utils.data.DataLoader(train_data, batch_size=args.batch_size, shuffle=True , num_workers=args.workers, pin_memory=True)
  valid_loader = torch.utils.data.DataLoader(valid_data, batch_size=args.batch_size, shuffle=False, num_workers=args.workers, pin_memory=True)

  split_file_path = Path(args.split_path)
  assert split_file_path.exists(), '{:} does not exist'.format(split_file_path)
  split_info      = torch.load(split_file_path)

  train_split, valid_split = split_info['train'], split_info['valid']
  assert len( set(train_split).intersection( set(valid_split) ) ) == 0, 'There should be 0 element that belongs to both train and valid'
  assert len(train_split) + len(valid_split) == len(train_data), '{:} + {:} vs {:}'.format(len(train_split), len(valid_split), len(train_data))
  search_dataset  = SearchDataset(args.dataset, train_data, train_split, valid_split)
  
  search_train_loader = torch.utils.data.DataLoader(train_data, batch_size=args.batch_size,
                      sampler=torch.utils.data.sampler.SubsetRandomSampler(train_split), pin_memory=True, num_workers=args.workers)
  search_valid_loader = torch.utils.data.DataLoader(train_data, batch_size=args.batch_size,
                      sampler=torch.utils.data.sampler.SubsetRandomSampler(valid_split), pin_memory=True, num_workers=args.workers)
  search_loader       = torch.utils.data.DataLoader(search_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.workers, pin_memory=True, sampler=None)
  # get configures
  if args.ablation_num_select is None or args.ablation_num_select <= 0:
    model_config = load_config(args.model_config, {'class_num': class_num, 'search_mode': 'shape'}, logger)
  else:
    model_config = load_config(args.model_config, {'class_num': class_num, 'search_mode': 'ablation', 'num_random_select': args.ablation_num_select}, logger)

  # obtain the model
  search_model = obtain_search_model(model_config)
  MAX_FLOP, param  = get_model_infos(search_model, xshape)
  optim_config = load_config(args.optim_config, {'class_num': class_num, 'FLOP': MAX_FLOP}, logger)
  logger.log('Model Information : {:}'.format(search_model.get_message()))
  logger.log('MAX_FLOP = {:} M'.format(MAX_FLOP))
  logger.log('Params   = {:} M'.format(param))
  logger.log('train_data : {:}'.format(train_data))
  logger.log('search-data: {:}'.format(search_dataset))
  logger.log('search_train_loader : {:} samples'.format( len(train_split) ))
  logger.log('search_valid_loader : {:} samples'.format( len(valid_split) ))
  base_optimizer, scheduler, criterion = get_optim_scheduler(search_model.base_parameters(), optim_config)
  arch_optimizer = torch.optim.Adam(search_model.arch_parameters(optim_config.arch_LR), lr=optim_config.arch_LR, betas=(0.5, 0.999), weight_decay=optim_config.arch_decay)
  logger.log('base-optimizer : {:}'.format(base_optimizer))
  logger.log('arch-optimizer : {:}'.format(arch_optimizer))
  logger.log('scheduler      : {:}'.format(scheduler))
  logger.log('criterion      : {:}'.format(criterion))
  
  last_info, model_base_path, model_best_path = logger.path('info'), logger.path('model'), logger.path('best')
  network, criterion = torch.nn.DataParallel(search_model).cuda(), criterion.cuda()

  # load checkpoint
  if last_info.exists() or (args.resume is not None and osp.isfile(args.resume)): # automatically resume from previous checkpoint
    if args.resume is not None and osp.isfile(args.resume):
      resume_path = Path(args.resume)
    elif last_info.exists():
      resume_path = last_info
    else: raise ValueError('Something is wrong.')
    logger.log("=> loading checkpoint of the last-info '{:}' start".format(resume_path))
    checkpoint  = torch.load(resume_path)
    if 'last_checkpoint' in checkpoint:
      last_checkpoint_path = checkpoint['last_checkpoint']
      if not last_checkpoint_path.exists():
        logger.log('Does not find {:}, try another path'.format(last_checkpoint_path))
        last_checkpoint_path = resume_path.parent / last_checkpoint_path.parent.name / last_checkpoint_path.name
      assert last_checkpoint_path.exists(), 'can not find the checkpoint from {:}'.format(last_checkpoint_path)
      checkpoint = torch.load( last_checkpoint_path )
    start_epoch = checkpoint['epoch'] + 1
    #for key, value in checkpoint['search_model'].items():
    #  print('K {:} = Shape={:}'.format(key, value.shape))
    search_model.load_state_dict( checkpoint['search_model'] )
    scheduler.load_state_dict ( checkpoint['scheduler'] )
    base_optimizer.load_state_dict ( checkpoint['base_optimizer'] )
    arch_optimizer.load_state_dict ( checkpoint['arch_optimizer'] )
    valid_accuracies = checkpoint['valid_accuracies']
    arch_genotypes   = checkpoint['arch_genotypes']
    discrepancies    = checkpoint['discrepancies']
    max_bytes        = checkpoint['max_bytes']
    logger.log("=> loading checkpoint of the last-info '{:}' start with {:}-th epoch.".format(resume_path, start_epoch))
  else:
    logger.log("=> do not find the last-info file : {:} or resume : {:}".format(last_info, args.resume))
    start_epoch, valid_accuracies, arch_genotypes, discrepancies, max_bytes = 0, {'best': -1}, {}, {}, {}

  # main procedure
  train_func, valid_func = get_procedures(args.procedure)
  total_epoch = optim_config.epochs + optim_config.warmup
  start_time, epoch_time = time.time(), AverageMeter()
  for epoch in range(start_epoch, total_epoch):
    scheduler.update(epoch, 0.0)
    search_model.set_tau(args.gumbel_tau_max, args.gumbel_tau_min, epoch*1.0/total_epoch)
    need_time = 'Time Left: {:}'.format( convert_secs2time(epoch_time.avg * (total_epoch-epoch), True) )
    epoch_str = 'epoch={:03d}/{:03d}'.format(epoch, total_epoch)
    LRs       = scheduler.get_lr()
    find_best = False
   
    logger.log('\n***{:s}*** start {:s} {:s}, LR=[{:.6f} ~ {:.6f}], scheduler={:}, tau={:}, FLOP={:.2f}'.format(time_string(), epoch_str, need_time, min(LRs), max(LRs), scheduler, search_model.tau, MAX_FLOP))

    # train for one epoch
    train_base_loss, train_arch_loss, train_acc1, train_acc5 = train_func(search_loader, network, criterion, scheduler, base_optimizer, arch_optimizer, optim_config, \
                                                                                {'epoch-str'  : epoch_str,        'FLOP-exp': MAX_FLOP * args.FLOP_ratio,
                                                                                 'FLOP-weight': args.FLOP_weight, 'FLOP-tolerant': MAX_FLOP * args.FLOP_tolerant}, args.print_freq, logger)
    # log the results
    logger.log('***{:s}*** TRAIN [{:}] base-loss = {:.6f}, arch-loss = {:.6f}, accuracy-1 = {:.2f}, accuracy-5 = {:.2f}'.format(time_string(), epoch_str, train_base_loss, train_arch_loss, train_acc1, train_acc5))
    cur_FLOP, genotype = search_model.get_flop('genotype', model_config._asdict(), None)
    arch_genotypes[epoch]  = genotype
    arch_genotypes['last'] = genotype
    logger.log('[{:}] genotype : {:}'.format(epoch_str, genotype))
    # save the configuration
    configure2str(genotype, str( logger.path('log') / 'seed-{:}-temp.config'.format(args.rand_seed) ))
    arch_info, discrepancy = search_model.get_arch_info()
    logger.log(arch_info)
    discrepancies[epoch]   = discrepancy
    logger.log('[{:}] FLOP : {:.2f} MB, ratio : {:.4f}, Expected-ratio : {:.4f}, Discrepancy : {:.3f}'.format(epoch_str, cur_FLOP, cur_FLOP/MAX_FLOP, args.FLOP_ratio, np.mean(discrepancy)))

    #if cur_FLOP/MAX_FLOP > args.FLOP_ratio:
    #  init_flop_weight = init_flop_weight * args.FLOP_decay
    #else:
    #  init_flop_weight = init_flop_weight / args.FLOP_decay
    
    # evaluate the performance
    if (epoch % args.eval_frequency == 0) or (epoch + 1 == total_epoch):
      logger.log('-'*150)
      valid_loss, valid_acc1, valid_acc5 = valid_func(search_valid_loader, network, criterion, epoch_str, args.print_freq_eval, logger)
      valid_accuracies[epoch] = valid_acc1
      logger.log('***{:s}*** VALID [{:}] loss = {:.6f}, accuracy@1 = {:.2f}, accuracy@5 = {:.2f} | Best-Valid-Acc@1={:.2f}, Error@1={:.2f}'.format(time_string(), epoch_str, valid_loss, valid_acc1, valid_acc5, valid_accuracies['best'], 100-valid_accuracies['best']))
      if valid_acc1 > valid_accuracies['best']:
        valid_accuracies['best'] = valid_acc1
        arch_genotypes['best']   = genotype
        find_best                = True
        logger.log('Currently, the best validation accuracy found at {:03d}-epoch :: acc@1={:.2f}, acc@5={:.2f}, error@1={:.2f}, error@5={:.2f}, save into {:}.'.format(epoch, valid_acc1, valid_acc5, 100-valid_acc1, 100-valid_acc5, model_best_path))
      # log the GPU memory usage
      #num_bytes = torch.cuda.max_memory_allocated( next(network.parameters()).device ) * 1.0
      num_bytes = torch.cuda.max_memory_cached( next(network.parameters()).device ) * 1.0
      logger.log('[GPU-Memory-Usage on {:} is {:} bytes, {:.2f} KB, {:.2f} MB, {:.2f} GB.]'.format(next(network.parameters()).device, int(num_bytes), num_bytes / 1e3, num_bytes / 1e6, num_bytes / 1e9))
      max_bytes[epoch] = num_bytes

    # save checkpoint
    save_path = save_checkpoint({
          'epoch'        : epoch,
          'args'         : deepcopy(args),
          'max_bytes'    : deepcopy(max_bytes),
          'valid_accuracies': deepcopy(valid_accuracies),
          'model-config' : model_config._asdict(),
          'optim-config' : optim_config._asdict(),
          'search_model' : search_model.state_dict(),
          'scheduler'    : scheduler.state_dict(),
          'base_optimizer': base_optimizer.state_dict(),
          'arch_optimizer': arch_optimizer.state_dict(),
          'arch_genotypes': arch_genotypes,
          'discrepancies' : discrepancies,
          }, model_base_path, logger)
    if find_best: copy_checkpoint(model_base_path, model_best_path, logger)
    last_info = save_checkpoint({
          'epoch': epoch,
          'args' : deepcopy(args),
          'last_checkpoint': save_path,
          }, logger.path('info'), logger)

    # measure elapsed time
    epoch_time.update(time.time() - start_time)
    start_time = time.time()
    

  logger.log('')
  logger.log('-'*100)
  last_config_path = logger.path('log') / 'seed-{:}-last.config'.format(args.rand_seed)
  configure2str(arch_genotypes['last'], str(last_config_path))
  logger.log('save the last config int {:} :\n{:}'.format(last_config_path, arch_genotypes['last']))

  best_arch, valid_acc = arch_genotypes['best'], valid_accuracies['best']
  for key, config in arch_genotypes.items():
    if key == 'last': continue
    FLOP_ratio = config['estimated_FLOP'] / MAX_FLOP
    if abs(FLOP_ratio - args.FLOP_ratio) <= args.FLOP_tolerant:
      if valid_acc <= valid_accuracies[key]:
        best_arch, valid_acc = config, valid_accuracies[key]
  print('Best-Arch : {:}\nRatio={:}, Valid-ACC={:}'.format(best_arch, best_arch['estimated_FLOP'] / MAX_FLOP, valid_acc))
  best_config_path = logger.path('log') / 'seed-{:}-best.config'.format(args.rand_seed)
  configure2str(best_arch, str(best_config_path))
  logger.log('save the last config int {:} :\n{:}'.format(best_config_path, best_arch))
  logger.log('\n' + '-'*200)
  logger.log('Finish training/validation in {:} with Max-GPU-Memory of {:.2f} GB, and save final checkpoint into {:}'.format(convert_secs2time(epoch_time.sum, True), max(v for k, v in max_bytes.items()) / 1e9, logger.path('info')))
  logger.close()


if __name__ == '__main__':
  args = obtain_args()
  main(args)