def validate(model, args, *, arch_loader=None):
assert arch_loader is not None
objs = AvgrageMeter()
top1 = AvgrageMeter()
top5 = AvgrageMeter()
val_dataloader = args.val_dataloader
model.eval()
t1 = time.time()
result_dict = {}
# base_model = mutableResNet20().cuda()
with torch.no_grad():
for key, arch in tqdm(arch_loader):
# print(key, arch)
# max_val_iters += 1
# print('\r ', key, ' iter:', max_val_iters, end='')
retrain_bn(model,
max_iters=5,
dataprovider=DataIterator(val_dataloader),
device=0,
cand=arch[0])
for data, target in val_dataloader: # 过一遍数据集
target = target.type(torch.LongTensor)
data, target = data.cuda(args.gpu), target.cuda(args.gpu)
output = model(data, arch[0])
prec1, prec5 = accuracy(output, target, topk=(1, 5))
n = data.size(0)
top1.update(prec1.item(), n)
top5.update(prec5.item(), n)
print("\t acc1: ", top1.avg)
tmp_dict = {}
tmp_dict['arch'] = arch[0]
tmp_dict['acc'] = top1.avg
result_dict[key[0]] = tmp_dict
with open("acc_result_rank_%d.json" % args.local_rank, "w") as f:
json.dump(result_dict, f)
def infer(train_loader, val_loader, model, criterion, archloader, args):
objs_, top1_, top5_ = AvgrageMeter(), AvgrageMeter(), AvgrageMeter()
model.eval()
now = time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time()))
# [16, 16, 16, 16, 16, 16, 16, 32, 32, 32, 32, 32, 32, 64, 64, 64, 64, 64, 64, 64]
fair_arc_list = archloader.generate_niu_fair_batch(random.randint(
0, 100))[-1].tolist()
# archloader.generate_spos_like_batch().tolist()
print('{} |=> Test rng = {}'.format(now, fair_arc_list)) # 只测试最后一个模型
# BN calibration
retrain_bn(model, 15, train_loader, fair_arc_list, device=0)
with torch.no_grad():
for step, (image, target) in enumerate(val_loader):
t0 = time.time()
datatime = time.time() - t0
image = Variable(image,
requires_grad=False).cuda(args.local_rank,
non_blocking=True)
target = Variable(target,
requires_grad=False).cuda(args.local_rank,
non_blocking=True)
logits = model(image, fair_arc_list)
loss = criterion(logits, target)
top1, top5 = accuracy(logits, target, topk=(1, 5))
if torch.cuda.device_count() > 1:
torch.distributed.barrier()
loss = reduce_mean(loss, args.nprocs)
top1 = reduce_mean(top1, image.size(0))
top5 = reduce_mean(top5, image.size(0))
n = image.size(0)
objs_.update(loss.data.item(), n)
top1_.update(top1.data.item(), n)
top5_.update(top5.data.item(), n)
now = time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time()))
print(
'{} |=> valid: step={}, loss={:.2f}, val_acc1={:.2f}, val_acc5={:2f}, datatime={:.2f}'
.format(now, step, objs_.avg, top1_.avg, top5_.avg, datatime))
return top1_.avg, top5_.avg, objs_.avg
def infer(train_dataprovider, val_dataprovider, model, criterion, fair_arc_list, val_iters, archloader):
objs = AvgrageMeter()
top1 = AvgrageMeter()
model.eval()
now = time.strftime('%Y-%m-%d %H:%M:%S', time.localtime(time.time()))
print('{} |=> Test rng = {}'.format(now, fair_arc_list[-1])) # 只测试最后一个模型
# BN calibration
retrain_bn(model, 5, train_dataprovider,
archloader.convert_list_arc_str(fair_arc_list[-1]), device=0)
with torch.no_grad():
for step in range(val_iters):
t0 = time.time()
image, target = val_dataprovider.next()
datatime = time.time() - t0
image = Variable(image, requires_grad=False).cuda()
target = Variable(target, requires_grad=False).cuda()
logits = model(
image, archloader.convert_list_arc_str(fair_arc_list[-1]))
loss = criterion(logits, target)
prec1, _ = accuracy(logits, target, topk=(1, 5))
n = image.size(0)
objs.update(loss.data.item(), n)
top1.update(prec1.data.item(), n)
# for arc in fair_arc_list:
# logits = model(image, archloader.convert_list_arc_str(arc))
# loss = criterion(logits, target)
# prec1, _ = accuracy(logits, target, topk=(1, 5))
# n = image.size(0)
# objs.update(loss.data.item(), n)
# top1.update(prec1.data.item(), n)
now = time.strftime('%Y-%m-%d %H:%M:%S',
time.localtime(time.time()))
print('{} |=> valid: step={}, loss={:.2f}, acc={:.2f}, datatime={:.2f}'.format(
now, step, objs.avg, top1.avg, datatime))
return top1.avg, objs.avg