iicd/xautodl
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Update TAS abd FBV2 for NAS-Bench

Browse Source
This commit is contained in:
D-X-Y
2020-07-24 12:56:34 +00:00
parent b9fbe5577c
commit 4a2292a863
8 changed files with 491 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
exps/algos-v2/search-cell.py
View File

@@ -338,8 +338,7 @@ def main(xargs):
else:
extra_info = {'class_num': class_num, 'xshape': xshape, 'epochs': xargs.overwite_epochs}
config = load_config(xargs.config_path, extra_info, logger)
search_loader, train_loader, valid_loader = get_nas_search_loaders(train_data, valid_data, xargs.dataset, 'configs/nas-benchmark/', \
(config.batch_size, config.test_batch_size), xargs.workers)
search_loader, train_loader, valid_loader = get_nas_search_loaders(train_data, valid_data, xargs.dataset, 'configs/nas-benchmark/', (config.batch_size, config.test_batch_size), xargs.workers)
logger.log('||||||| {:10s} ||||||| Search-Loader-Num={:}, Valid-Loader-Num={:}, batch size={:}'.format(xargs.dataset, len(search_loader), len(valid_loader), config.batch_size))
logger.log('||||||| {:10s} ||||||| Config={:}'.format(xargs.dataset, config))

334
exps/algos-v2/search-size.py Normal file
View File

@@ -0,0 +1,334 @@
##################################################
# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2020 #
######################################################################################
# python ./exps/algos-v2/search-size.py --dataset cifar10 --data_path $TORCH_HOME/cifar.python --algo tas --rand_seed 777
# python ./exps/algos-v2/search-size.py --dataset cifar100 --data_path $TORCH_HOME/cifar.python --algo tas --rand_seed 777
# python ./exps/algos-v2/search-size.py --dataset ImageNet16-120 --data_path $TORCH_HOME/cifar.python/ImageNet16 --algo tas --rand_seed 777
####
# python ./exps/algos-v2/search-size.py --dataset cifar10 --data_path $TORCH_HOME/cifar.python --algo fbv2 --rand_seed 777
# python ./exps/algos-v2/search-size.py --dataset cifar100 --data_path $TORCH_HOME/cifar.python --algo fbv2 --rand_seed 777
# python ./exps/algos-v2/search-size.py --dataset ImageNet16-120 --data_path $TORCH_HOME/cifar.python/ImageNet16 --algo fbv2 --rand_seed 777
######################################################################################
import os, sys, time, random, argparse
import numpy as np
from copy import deepcopy
import torch
import torch.nn as nn
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, dict2config, configure2str
from datasets import get_datasets, get_nas_search_loaders
from procedures import prepare_seed, prepare_logger, save_checkpoint, copy_checkpoint, get_optim_scheduler
from utils import count_parameters_in_MB, obtain_accuracy
from log_utils import AverageMeter, time_string, convert_secs2time
from models import get_cell_based_tiny_net, get_search_spaces
from nas_201_api import NASBench301API as API
def search_func(xloader, network, criterion, scheduler, w_optimizer, a_optimizer, epoch_str, print_freq, logger):
data_time, batch_time = AverageMeter(), AverageMeter()
base_losses, base_top1, base_top5 = AverageMeter(), AverageMeter(), AverageMeter()
arch_losses, arch_top1, arch_top5 = AverageMeter(), AverageMeter(), AverageMeter()
end = time.time()
network.train()
for step, (base_inputs, base_targets, arch_inputs, arch_targets) in enumerate(xloader):
scheduler.update(None, 1.0 * step / len(xloader))
base_inputs = base_inputs.cuda(non_blocking=True)
arch_inputs = arch_inputs.cuda(non_blocking=True)
base_targets = base_targets.cuda(non_blocking=True)
arch_targets = arch_targets.cuda(non_blocking=True)
# measure data loading time
data_time.update(time.time() - end)
# Update the weights
network.zero_grad()
_, logits = network(base_inputs)
base_loss = criterion(logits, base_targets)
base_loss.backward()
w_optimizer.step()
# record
base_prec1, base_prec5 = obtain_accuracy(logits.data, base_targets.data, topk=(1, 5))
base_losses.update(base_loss.item(), base_inputs.size(0))
base_top1.update (base_prec1.item(), base_inputs.size(0))
base_top5.update (base_prec5.item(), base_inputs.size(0))
# update the architecture-weight
network.zero_grad()
_, logits = network(arch_inputs)
arch_loss = criterion(logits, arch_targets)
arch_loss.backward()
a_optimizer.step()
# record
arch_prec1, arch_prec5 = obtain_accuracy(logits.data, arch_targets.data, topk=(1, 5))
arch_losses.update(arch_loss.item(), arch_inputs.size(0))
arch_top1.update (arch_prec1.item(), arch_inputs.size(0))
arch_top5.update (arch_prec5.item(), arch_inputs.size(0))
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
if step % print_freq == 0 or step + 1 == len(xloader):
Sstr = '*SEARCH* ' + time_string() + ' [{:}][{:03d}/{:03d}]'.format(epoch_str, step, len(xloader))
Tstr = 'Time {batch_time.val:.2f} ({batch_time.avg:.2f}) Data {data_time.val:.2f} ({data_time.avg:.2f})'.format(batch_time=batch_time, data_time=data_time)
Wstr = 'Base [Loss {loss.val:.3f} ({loss.avg:.3f}) Prec@1 {top1.val:.2f} ({top1.avg:.2f}) Prec@5 {top5.val:.2f} ({top5.avg:.2f})]'.format(loss=base_losses, top1=base_top1, top5=base_top5)
Astr = 'Arch [Loss {loss.val:.3f} ({loss.avg:.3f}) Prec@1 {top1.val:.2f} ({top1.avg:.2f}) Prec@5 {top5.val:.2f} ({top5.avg:.2f})]'.format(loss=arch_losses, top1=arch_top1, top5=arch_top5)
logger.log(Sstr + ' ' + Tstr + ' ' + Wstr + ' ' + Astr)
return base_losses.avg, base_top1.avg, base_top5.avg, arch_losses.avg, arch_top1.avg, arch_top5.avg
def train_controller(xloader, network, criterion, optimizer, prev_baseline, epoch_str, print_freq, logger):
# config. (containing some necessary arg)
# baseline: The baseline score (i.e. average val_acc) from the previous epoch
data_time, batch_time = AverageMeter(), AverageMeter()
GradnormMeter, LossMeter, ValAccMeter, EntropyMeter, BaselineMeter, RewardMeter, xend = AverageMeter(), AverageMeter(), AverageMeter(), AverageMeter(), AverageMeter(), AverageMeter(), time.time()
controller_num_aggregate = 20
controller_train_steps = 50
controller_bl_dec = 0.99
controller_entropy_weight = 0.0001
network.eval()
network.controller.train()
network.controller.zero_grad()
loader_iter = iter(xloader)
for step in range(controller_train_steps * controller_num_aggregate):
try:
inputs, targets = next(loader_iter)
except:
loader_iter = iter(xloader)
inputs, targets = next(loader_iter)
inputs = inputs.cuda(non_blocking=True)
targets = targets.cuda(non_blocking=True)
# measure data loading time
data_time.update(time.time() - xend)
log_prob, entropy, sampled_arch = network.controller()
with torch.no_grad():
network.set_cal_mode('dynamic', sampled_arch)
_, logits = network(inputs)
val_top1, val_top5 = obtain_accuracy(logits.data, targets.data, topk=(1, 5))
val_top1 = val_top1.view(-1) / 100
reward = val_top1 + controller_entropy_weight * entropy
if prev_baseline is None:
baseline = val_top1
else:
baseline = prev_baseline - (1 - controller_bl_dec) * (prev_baseline - reward)
loss = -1 * log_prob * (reward - baseline)
# account
RewardMeter.update(reward.item())
BaselineMeter.update(baseline.item())
ValAccMeter.update(val_top1.item()*100)
LossMeter.update(loss.item())
EntropyMeter.update(entropy.item())
# Average gradient over controller_num_aggregate samples
loss = loss / controller_num_aggregate
loss.backward(retain_graph=True)
# measure elapsed time
batch_time.update(time.time() - xend)
xend = time.time()
if (step+1) % controller_num_aggregate == 0:
grad_norm = torch.nn.utils.clip_grad_norm_(network.controller.parameters(), 5.0)
GradnormMeter.update(grad_norm)
optimizer.step()
network.controller.zero_grad()
if step % print_freq == 0:
Sstr = '*Train-Controller* ' + time_string() + ' [{:}][{:03d}/{:03d}]'.format(epoch_str, step, controller_train_steps * controller_num_aggregate)
Tstr = 'Time {batch_time.val:.2f} ({batch_time.avg:.2f}) Data {data_time.val:.2f} ({data_time.avg:.2f})'.format(batch_time=batch_time, data_time=data_time)
Wstr = '[Loss {loss.val:.3f} ({loss.avg:.3f}) Prec@1 {top1.val:.2f} ({top1.avg:.2f}) Reward {reward.val:.2f} ({reward.avg:.2f})] Baseline {basel.val:.2f} ({basel.avg:.2f})'.format(loss=LossMeter, top1=ValAccMeter, reward=RewardMeter, basel=BaselineMeter)
Estr = 'Entropy={:.4f} ({:.4f})'.format(EntropyMeter.val, EntropyMeter.avg)
logger.log(Sstr + ' ' + Tstr + ' ' + Wstr + ' ' + Estr)
return LossMeter.avg, ValAccMeter.avg, BaselineMeter.avg, RewardMeter.avg
def valid_func(xloader, network, criterion, logger):
data_time, batch_time = AverageMeter(), AverageMeter()
arch_losses, arch_top1, arch_top5 = AverageMeter(), AverageMeter(), AverageMeter()
end = time.time()
with torch.no_grad():
network.eval()
for step, (arch_inputs, arch_targets) in enumerate(xloader):
arch_targets = arch_targets.cuda(non_blocking=True)
# measure data loading time
data_time.update(time.time() - end)
# prediction
_, logits = network(arch_inputs.cuda(non_blocking=True))
arch_loss = criterion(logits, arch_targets)
# record
arch_prec1, arch_prec5 = obtain_accuracy(logits.data, arch_targets.data, topk=(1, 5))
arch_losses.update(arch_loss.item(), arch_inputs.size(0))
arch_top1.update (arch_prec1.item(), arch_inputs.size(0))
arch_top5.update (arch_prec5.item(), arch_inputs.size(0))
# measure elapsed time
batch_time.update(time.time() - end)
end = time.time()
return arch_losses.avg, arch_top1.avg, arch_top5.avg
def main(xargs):
assert torch.cuda.is_available(), 'CUDA is not available.'
torch.backends.cudnn.enabled = True
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
torch.set_num_threads( xargs.workers )
prepare_seed(xargs.rand_seed)
logger = prepare_logger(args)
train_data, valid_data, xshape, class_num = get_datasets(xargs.dataset, xargs.data_path, -1)
if xargs.overwite_epochs is None:
extra_info = {'class_num': class_num, 'xshape': xshape}
else:
extra_info = {'class_num': class_num, 'xshape': xshape, 'epochs': xargs.overwite_epochs}
config = load_config(xargs.config_path, extra_info, logger)
search_loader, train_loader, valid_loader = get_nas_search_loaders(train_data, valid_data, xargs.dataset, 'configs/nas-benchmark/', (config.batch_size, config.test_batch_size), xargs.workers)
logger.log('||||||| {:10s} ||||||| Search-Loader-Num={:}, Valid-Loader-Num={:}, batch size={:}'.format(xargs.dataset, len(search_loader), len(valid_loader), config.batch_size))
logger.log('||||||| {:10s} ||||||| Config={:}'.format(xargs.dataset, config))
search_space = get_search_spaces(xargs.search_space, 'nas-bench-301')
model_config = dict2config(
dict(name='generic', super_type='search-shape', candidate_Cs=search_space['candidates'], max_num_Cs=search_space['numbers'], num_classes=class_num,
genotype=args.genotype, affine=bool(xargs.affine), track_running_stats=bool(xargs.track_running_stats)), None)
logger.log('search space : {:}'.format(search_space))
logger.log('model config : {:}'.format(model_config))
search_model = get_cell_based_tiny_net(model_config)
search_model.set_algo(xargs.algo)
logger.log('{:}'.format(search_model))
w_optimizer, w_scheduler, criterion = get_optim_scheduler(search_model.weights, config)
a_optimizer = torch.optim.Adam(search_model.alphas, lr=xargs.arch_learning_rate, betas=(0.5, 0.999), weight_decay=xargs.arch_weight_decay, eps=xargs.arch_eps)
logger.log('w-optimizer : {:}'.format(w_optimizer))
logger.log('a-optimizer : {:}'.format(a_optimizer))
logger.log('w-scheduler : {:}'.format(w_scheduler))
logger.log('criterion : {:}'.format(criterion))
params = count_parameters_in_MB(search_model)
logger.log('The parameters of the search model = {:.2f} MB'.format(params))
logger.log('search-space : {:}'.format(search_space))
try:
api = API(verbose=False)
except:
api = None
logger.log('{:} create API = {:} done'.format(time_string(), api))
last_info, model_base_path, model_best_path = logger.path('info'), logger.path('model'), logger.path('best')
network, criterion = search_model.cuda(), criterion.cuda() # use a single GPU
last_info, model_base_path, model_best_path = logger.path('info'), logger.path('model'), logger.path('best')
if last_info.exists(): # automatically resume from previous checkpoint
logger.log("=> loading checkpoint of the last-info '{:}' start".format(last_info))
last_info = torch.load(last_info)
start_epoch = last_info['epoch']
checkpoint = torch.load(last_info['last_checkpoint'])
genotypes = checkpoint['genotypes']
valid_accuracies = checkpoint['valid_accuracies']
search_model.load_state_dict( checkpoint['search_model'] )
w_scheduler.load_state_dict ( checkpoint['w_scheduler'] )
w_optimizer.load_state_dict ( checkpoint['w_optimizer'] )
a_optimizer.load_state_dict ( checkpoint['a_optimizer'] )
logger.log("=> loading checkpoint of the last-info '{:}' start with {:}-th epoch.".format(last_info, start_epoch))
else:
logger.log("=> do not find the last-info file : {:}".format(last_info))
start_epoch, valid_accuracies, genotypes = 0, {'best': -1}, {-1: network.random}
# start training
start_time, search_time, epoch_time, total_epoch = time.time(), AverageMeter(), AverageMeter(), config.epochs + config.warmup
for epoch in range(start_epoch, total_epoch):
w_scheduler.update(epoch, 0.0)
need_time = 'Time Left: {:}'.format(convert_secs2time(epoch_time.val * (total_epoch-epoch), True))
epoch_str = '{:03d}-{:03d}'.format(epoch, total_epoch)
logger.log('\n[Search the {:}-th epoch] {:}, LR={:}'.format(epoch_str, need_time, min(w_scheduler.get_lr())))
if xargs.algo == 'fbv2' or xargs.algo == 'tas':
network.set_tau( xargs.tau_max - (xargs.tau_max-xargs.tau_min) * epoch / (total_epoch-1) )
logger.log('[RESET tau as : {:}]'.format(network.tau))
search_w_loss, search_w_top1, search_w_top5, search_a_loss, search_a_top1, search_a_top5 \
= search_func(search_loader, network, criterion, w_scheduler, w_optimizer, a_optimizer, epoch_str, xargs.print_freq, logger)
search_time.update(time.time() - start_time)
logger.log('[{:}] search [base] : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%, time-cost={:.1f} s'.format(epoch_str, search_w_loss, search_w_top1, search_w_top5, search_time.sum))
logger.log('[{:}] search [arch] : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}%'.format(epoch_str, search_a_loss, search_a_top1, search_a_top5))
genotype = network.genotype
logger.log('[{:}] - [get_best_arch] : {:}'.format(epoch_str, genotype))
valid_a_loss , valid_a_top1 , valid_a_top5 = valid_func(valid_loader, network, criterion, logger)
logger.log('[{:}] evaluate : loss={:.2f}, accuracy@1={:.2f}%, accuracy@5={:.2f}% | {:}'.format(epoch_str, valid_a_loss, valid_a_top1, valid_a_top5, genotype))
valid_accuracies[epoch] = valid_a_top1
genotypes[epoch] = genotype
logger.log('<<<--->>> The {:}-th epoch : {:}'.format(epoch_str, genotypes[epoch]))
# save checkpoint
save_path = save_checkpoint({'epoch' : epoch + 1,
'args' : deepcopy(xargs),
'search_model': search_model.state_dict(),
'w_optimizer' : w_optimizer.state_dict(),
'a_optimizer' : a_optimizer.state_dict(),
'w_scheduler' : w_scheduler.state_dict(),
'genotypes' : genotypes,
'valid_accuracies' : valid_accuracies},
model_base_path, logger)
last_info = save_checkpoint({
'epoch': epoch + 1,
'args' : deepcopy(args),
'last_checkpoint': save_path,
}, logger.path('info'), logger)
with torch.no_grad():
logger.log('{:}'.format(search_model.show_alphas()))
if api is not None: logger.log('{:}'.format(api.query_by_arch(genotypes[epoch], '90')))
# measure elapsed time
epoch_time.update(time.time() - start_time)
start_time = time.time()
# the final post procedure : count the time
start_time = time.time()
genotype = network.genotype
search_time.update(time.time() - start_time)
valid_a_loss, valid_a_top1, valid_a_top5 = valid_func(valid_loader, network, criterion, logger)
logger.log('Last : the gentotype is : {:}, with the validation accuracy of {:.3f}%.'.format(genotype, valid_a_top1))
logger.log('\n' + '-'*100)
# check the performance from the architecture dataset
logger.log('[{:}] run {:} epochs, cost {:.1f} s, last-geno is {:}.'.format(xargs.algo, total_epoch, search_time.sum, genotype))
if api is not None: logger.log('{:}'.format(api.query_by_arch(genotype, '90') ))
logger.close()
if __name__ == '__main__':
parser = argparse.ArgumentParser("Weight sharing NAS methods to search for cells.")
parser.add_argument('--data_path' , type=str, help='Path to dataset')
parser.add_argument('--dataset' , type=str, choices=['cifar10', 'cifar100', 'ImageNet16-120'], help='Choose between Cifar10/100 and ImageNet-16.')
parser.add_argument('--search_space', type=str, default='sss', choices=['sss'], help='The search space name.')
parser.add_argument('--algo' , type=str, choices=['tas', 'fbv2', 'enas'], help='The search space name.')
parser.add_argument('--genotype' , type=str, default='|nor_conv_3x3~0|+|nor_conv_3x3~0|nor_conv_3x3~1|+|skip_connect~0|nor_conv_3x3~1|nor_conv_3x3~2|', help='The genotype.')
# FOR GDAS
parser.add_argument('--tau_min', type=float, default=0.1, help='The minimum tau for Gumbel Softmax.')
parser.add_argument('--tau_max', type=float, default=10, help='The maximum tau for Gumbel Softmax.')
#
parser.add_argument('--track_running_stats',type=int, default=0, choices=[0,1],help='Whether use track_running_stats or not in the BN layer.')
parser.add_argument('--affine' , type=int, default=0, choices=[0,1],help='Whether use affine=True or False in the BN layer.')
parser.add_argument('--config_path' , type=str, default='./configs/nas-benchmark/algos/weight-sharing.config', help='The path of configuration.')
parser.add_argument('--overwite_epochs', type=int, help='The number of epochs to overwrite that value in config files.')
# architecture leraning rate
parser.add_argument('--arch_learning_rate', type=float, default=3e-4, help='learning rate for arch encoding')
parser.add_argument('--arch_weight_decay' , type=float, default=1e-3, help='weight decay for arch encoding')
parser.add_argument('--arch_eps' , type=float, default=1e-8, help='weight decay for arch encoding')
# log
parser.add_argument('--workers', type=int, default=2, help='number of data loading workers (default: 2)')
parser.add_argument('--save_dir', type=str, default='./output/search', help='Folder to save checkpoints and log.')
parser.add_argument('--print_freq', type=int, default=200, help='print frequency (default: 200)')
parser.add_argument('--rand_seed', type=int, help='manual seed')
args = parser.parse_args()
if args.rand_seed is None or args.rand_seed < 0: args.rand_seed = random.randint(1, 100000)
dirname = '{:}-affine{:}_BN{:}'.format(args.algo, args.affine, args.track_running_stats)
if args.overwite_epochs is not None:
dirname = dirname + '-E{:}'.format(args.overwite_epochs)
args.save_dir = os.path.join('{:}-{:}'.format(args.save_dir, args.search_space), args.dataset, dirname)
main(args)

1
exps/experimental/vis-bench-ws.py
View File

@@ -33,6 +33,7 @@ def fetch_data(root_dir='./output/search', search_space='tss', dataset=None):
alg2name['GDAS'] = 'gdas-affine0_BN0-None'
alg2name['RSPS'] = 'random-affine0_BN0-None'
alg2name['DARTS (1st)'] = 'darts-v1-affine0_BN0-None'
alg2name['ENAS'] = 'enas-affine0_BN0-None'
"""
alg2name['DARTS (2nd)'] = 'darts-v2-affine1_BN0-None'
alg2name['SETN'] = 'setn-affine1_BN0-None'