def main():
if not torch.cuda.is_available():
logging.info('no gpu device available')
sys.exit(1)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
cudnn.enabled = True
cudnn.benchmark = False
cudnn.deterministic = True
args.steps = int(np.ceil(
45000 / args.child_batch_size)) * args.child_epochs
logging.info("args = %s", args)
if args.child_arch_pool is not None:
logging.info('Architecture pool is provided, loading')
with open(args.child_arch_pool) as f:
archs = f.read().splitlines()
archs = list(map(utils.build_dag, archs))
child_arch_pool = archs
elif os.path.exists(os.path.join(args.output_dir, 'arch_pool')):
logging.info('Architecture pool is founded, loading')
with open(os.path.join(args.output_dir, 'arch_pool')) as f:
archs = f.read().splitlines()
archs = list(map(utils.build_dag, archs))
child_arch_pool = archs
else:
child_arch_pool = None
child_eval_epochs = eval(args.child_eval_epochs)
build_fn = get_builder(args.dataset)
train_queue, valid_queue, model, train_criterion, eval_criterion, optimizer, scheduler = build_fn(
ratio=0.9, epoch=-1)
nao = NAO(
args.controller_encoder_layers,
args.controller_encoder_vocab_size,
args.controller_encoder_hidden_size,
args.controller_encoder_dropout,
args.controller_encoder_length,
args.controller_source_length,
args.controller_encoder_emb_size,
args.controller_mlp_layers,
args.controller_mlp_hidden_size,
args.controller_mlp_dropout,
args.controller_decoder_layers,
args.controller_decoder_vocab_size,
args.controller_decoder_hidden_size,
args.controller_decoder_dropout,
args.controller_decoder_length,
)
nao = nao.cuda()
logging.info("Encoder-Predictor-Decoder param size = %fMB",
utils.count_parameters_in_MB(nao))
# Train child model
if child_arch_pool is None:
logging.info(
'Architecture pool is not provided, randomly generating now')
child_arch_pool = utils.generate_arch(args.controller_seed_arch,
args.child_nodes,
5) # [[[conv],[reduc]]]
if args.child_sample_policy == 'params':
child_arch_pool_prob = []
for arch in child_arch_pool:
if args.dataset == 'cifar10':
tmp_model = NASNetworkCIFAR(
args, 10, args.child_layers, args.child_nodes,
args.child_channels, args.child_keep_prob,
args.child_drop_path_keep_prob, args.child_use_aux_head,
args.steps, arch)
elif args.dataset == 'cifar100':
tmp_model = NASNetworkCIFAR(
args, 100, args.child_layers, args.child_nodes,
args.child_channels, args.child_keep_prob,
args.child_drop_path_keep_prob, args.child_use_aux_head,
args.steps, arch)
else:
tmp_model = NASNetworkImageNet(
args, 1000, args.child_layers, args.child_nodes,
args.child_channels, args.child_keep_prob,
args.child_drop_path_keep_prob, args.child_use_aux_head,
args.steps, arch)
child_arch_pool_prob.append(
utils.count_parameters_in_MB(tmp_model))
del tmp_model
else:
child_arch_pool_prob = None
eval_points = utils.generate_eval_points(child_eval_epochs, 0,
args.child_epochs)
step = 0
for epoch in range(1, args.child_epochs + 1):
scheduler.step()
lr = scheduler.get_lr()[0]
logging.info('epoch %d lr %e', epoch, lr)
# sample an arch to train
train_acc, train_obj, step = child_train(train_queue, model, optimizer,
step, child_arch_pool,
child_arch_pool_prob,
train_criterion)
logging.info('train_acc %f', train_acc)
if epoch not in eval_points:
continue
# Evaluate seed archs
valid_accuracy_list = child_valid(valid_queue, model, child_arch_pool,
eval_criterion)
# Output archs and evaluated error rate
old_archs = child_arch_pool
old_archs_perf = valid_accuracy_list
old_archs_sorted_indices = np.argsort(old_archs_perf)[::-1]
old_archs = [old_archs[i] for i in old_archs_sorted_indices]
old_archs_perf = [old_archs_perf[i] for i in old_archs_sorted_indices]
with open(os.path.join(args.output_dir, 'arch_pool.{}'.format(epoch)),
'w') as fa:
with open(
os.path.join(args.output_dir,
'arch_pool.perf.{}'.format(epoch)),
'w') as fp:
with open(os.path.join(args.output_dir, 'arch_pool'),
'w') as fa_latest:
with open(os.path.join(args.output_dir, 'arch_pool.perf'),
'w') as fp_latest:
for arch, perf in zip(old_archs, old_archs_perf):
arch = ' '.join(map(str, arch[0] + arch[1]))
fa.write('{}\n'.format(arch))
fa_latest.write('{}\n'.format(arch))
fp.write('{}\n'.format(perf))
fp_latest.write('{}\n'.format(perf))
if epoch == args.child_epochs:
break
# Train Encoder-Predictor-Decoder
logging.info('Training Encoder-Predictor-Decoder')
encoder_input = list(
map(
lambda x: utils.parse_arch_to_seq(x[0], 2) + utils.
parse_arch_to_seq(x[1], 2), old_archs))
# [[conv, reduc]]
min_val = min(old_archs_perf)
max_val = max(old_archs_perf)
encoder_target = [(i - min_val) / (max_val - min_val)
for i in old_archs_perf]
if args.controller_expand is not None:
dataset = list(zip(encoder_input, encoder_target))
n = len(dataset)
ratio = 0.9
split = int(n * ratio)
np.random.shuffle(dataset)
encoder_input, encoder_target = list(zip(*dataset))
train_encoder_input = list(encoder_input[:split])
train_encoder_target = list(encoder_target[:split])
valid_encoder_input = list(encoder_input[split:])
valid_encoder_target = list(encoder_target[split:])
for _ in range(args.controller_expand - 1):
for src, tgt in zip(encoder_input[:split],
encoder_target[:split]):
a = np.random.randint(0, args.child_nodes)
b = np.random.randint(0, args.child_nodes)
src = src[:4 * a] + src[4 * a + 2:4 * a + 4] + \
src[4 * a:4 * a + 2] + src[4 * (a + 1):20 + 4 * b] + \
src[20 + 4 * b + 2:20 + 4 * b + 4] + src[20 + 4 * b:20 + 4 * b + 2] + \
src[20 + 4 * (b + 1):]
train_encoder_input.append(src)
train_encoder_target.append(tgt)
else:
train_encoder_input = encoder_input
train_encoder_target = encoder_target
valid_encoder_input = encoder_input
valid_encoder_target = encoder_target
logging.info('Train data: {}\tValid data: {}'.format(
len(train_encoder_input), len(valid_encoder_input)))
nao_train_dataset = utils.NAODataset(
train_encoder_input,
train_encoder_target,
True,
swap=True if args.controller_expand is None else False)
nao_valid_dataset = utils.NAODataset(valid_encoder_input,
valid_encoder_target, False)
nao_train_queue = torch.utils.data.DataLoader(
nao_train_dataset,
batch_size=args.controller_batch_size,
shuffle=True,
pin_memory=True)
nao_valid_queue = torch.utils.data.DataLoader(
nao_valid_dataset,
batch_size=args.controller_batch_size,
shuffle=False,
pin_memory=True)
nao_optimizer = torch.optim.Adam(nao.parameters(),
lr=args.controller_lr,
weight_decay=args.controller_l2_reg)
for nao_epoch in range(1, args.controller_epochs + 1):
nao_loss, nao_mse, nao_ce = nao_train(nao_train_queue, nao,
nao_optimizer)
logging.info("epoch %04d train loss %.6f mse %.6f ce %.6f",
nao_epoch, nao_loss, nao_mse, nao_ce)
if nao_epoch % 100 == 0:
pa, hs = nao_valid(nao_valid_queue, nao)
logging.info("Evaluation on valid data")
logging.info(
'epoch %04d pairwise accuracy %.6f hamming distance %.6f',
epoch, pa, hs)
# Generate new archs
new_archs = []
max_step_size = 50
predict_step_size = 0
top100_archs = list(
map(
lambda x: utils.parse_arch_to_seq(x[0], 2) + utils.
parse_arch_to_seq(x[1], 2), old_archs[:100]))
nao_infer_dataset = utils.NAODataset(top100_archs, None, False)
nao_infer_queue = torch.utils.data.DataLoader(
nao_infer_dataset,
batch_size=len(nao_infer_dataset),
shuffle=False,
pin_memory=True)
while len(new_archs) < args.controller_new_arch:
predict_step_size += 1
logging.info('Generate new architectures with step size %d',
predict_step_size)
new_arch = nao_infer(nao_infer_queue,
nao,
predict_step_size,
direction='+')
for arch in new_arch:
if arch not in encoder_input and arch not in new_archs:
new_archs.append(arch)
if len(new_archs) >= args.controller_new_arch:
break
logging.info('%d new archs generated now', len(new_archs))
if predict_step_size > max_step_size:
break
# [[conv, reduc]]
new_archs = list(
map(lambda x: utils.parse_seq_to_arch(x, 2),
new_archs)) # [[[conv],[reduc]]]
num_new_archs = len(new_archs)
logging.info("Generate %d new archs", num_new_archs)
# replace bottom archs
if args.controller_replace:
new_arch_pool = old_archs[:len(old_archs) - (num_new_archs + args.controller_random_arch)] + \
new_archs + utils.generate_arch(args.controller_random_arch, 5, 5)
# discard all archs except top k
elif args.controller_discard:
new_arch_pool = old_archs[:100] + new_archs + utils.generate_arch(
args.controller_random_arch, 5, 5)
# use all
else:
new_arch_pool = old_archs + new_archs + utils.generate_arch(
args.controller_random_arch, 5, 5)
logging.info("Totally %d architectures now to train",
len(new_arch_pool))
child_arch_pool = new_arch_pool
with open(os.path.join(args.output_dir, 'arch_pool'), 'w') as f:
for arch in new_arch_pool:
arch = ' '.join(map(str, arch[0] + arch[1]))
f.write('{}\n'.format(arch))
if args.child_sample_policy == 'params':
child_arch_pool_prob = []
for arch in child_arch_pool:
if args.dataset == 'cifar10':
tmp_model = NASNetworkCIFAR(
args, 10, args.child_layers, args.child_nodes,
args.child_channels, args.child_keep_prob,
args.child_drop_path_keep_prob,
args.child_use_aux_head, args.steps, arch)
elif args.dataset == 'cifar100':
tmp_model = NASNetworkCIFAR(
args, 100, args.child_layers, args.child_nodes,
args.child_channels, args.child_keep_prob,
args.child_drop_path_keep_prob,
args.child_use_aux_head, args.steps, arch)
else:
tmp_model = NASNetworkImageNet(
args, 1000, args.child_layers, args.child_nodes,
args.child_channels, args.child_keep_prob,
args.child_drop_path_keep_prob,
args.child_use_aux_head, args.steps, arch)
child_arch_pool_prob.append(
utils.count_parameters_in_MB(tmp_model))
del tmp_model
else:
child_arch_pool_prob = None
def main():
if not torch.cuda.is_available():
logging.info('No GPU found!')
sys.exit(1)
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
cudnn.enabled = True
cudnn.benchmark = False
cudnn.deterministic = True
logging.info("Args = %s", args)
nao = NAO(
args.encoder_layers,
args.encoder_vocab_size,
args.encoder_hidden_size,
args.encoder_dropout,
args.encoder_length,
args.source_length,
args.encoder_emb_size,
args.mlp_layers,
args.mlp_hidden_size,
args.mlp_dropout,
args.decoder_layers,
args.decoder_vocab_size,
args.decoder_hidden_size,
args.decoder_dropout,
args.decoder_length,
)
logging.info("param size = %fMB", utils.count_parameters_in_MB(nao))
nao = nao.cuda()
with open(
os.path.join(args.output_dir,
'arch_pool.{}'.format(args.iteration))) as f:
arch_pool = f.read().splitlines()
arch_pool = list(map(utils.build_dag, arch_pool))
with open(
os.path.join(args.output_dir,
'arch_pool.{}.perf'.format(args.iteration))) as f:
arch_pool_valid_acc = f.read().splitlines()
arch_pool_valid_acc = list(map(float, arch_pool_valid_acc))
logging.info('Training Encoder-Predictor-Decoder')
train_encoder_input = list(
map(
lambda x: utils.parse_arch_to_seq(x[0], 2) + utils.
parse_arch_to_seq(x[1], 2), arch_pool))
min_val = min(arch_pool_valid_acc)
max_val = max(arch_pool_valid_acc)
train_encoder_target = [(i - min_val) / (max_val - min_val)
for i in arch_pool_valid_acc]
if args.expand is not None:
buffer1, buffer2 = [], []
for _ in range(args.expand - 1):
for src, tgt in zip(train_encoder_input, train_encoder_target):
a = np.random.randint(0, 5)
b = np.random.randint(0, 5)
src = src[:4 * a] + src[4 * a + 2:4 * a + 4] + \
src[4 * a:4 * a + 2] + src[4 * (a + 1):20 + 4 * b] + \
src[20 + 4 * b + 2:20 + 4 * b + 4] + src[20 + 4 * b:20 + 4 * b + 2] + \
src[20 + 4 * (b + 1):]
buffer1.append(src)
buffer2.append(tgt)
train_encoder_input += buffer1
train_encoder_target += buffer2
nao_train_dataset = utils.NAODataset(train_encoder_input,
train_encoder_target,
True,
swap=True)
nao_valid_dataset = utils.NAODataset(train_encoder_input,
train_encoder_target, False)
nao_train_queue = torch.utils.data.DataLoader(nao_train_dataset,
batch_size=args.batch_size,
shuffle=True,
pin_memory=True)
nao_valid_queue = torch.utils.data.DataLoader(
nao_valid_dataset,
batch_size=len(nao_valid_dataset),
shuffle=False,
pin_memory=True)
nao_optimizer = torch.optim.Adam(nao.parameters(),
lr=args.lr,
weight_decay=args.l2_reg)
for nao_epoch in range(1, args.epochs + 1):
nao_loss, nao_mse, nao_ce = nao_train(nao_train_queue, nao,
nao_optimizer)
if nao_epoch % 10 == 0 or nao_epoch == 1:
logging.info("epoch %04d train loss %.6f mse %.6f ce %.6f",
nao_epoch, nao_loss, nao_mse, nao_ce)
if nao_epoch % 100 == 0 or nao_epoch == 1:
mse, pa, hs = nao_valid(nao_train_queue, nao)
logging.info("Evaluation on train data")
logging.info(
'epoch %04d mse %.6f pairwise accuracy %.6f hamming distance %.6f',
nao_epoch, mse, pa, hs)
mse, pa, hs = nao_valid(nao_valid_queue, nao)
logging.info("Evaluation on valid data")
logging.info(
'epoch %04d mse %.6f pairwise accuracy %.6f hamming distance %.6f',
nao_epoch, mse, pa, hs)
new_archs = []
predict_step_size = 0
top_archs = list(
map(
lambda x: utils.parse_arch_to_seq(x[0], 2) + utils.
parse_arch_to_seq(x[1], 2), arch_pool[:args.generate_topk]))
nao_infer_dataset = utils.NAODataset(top_archs, None, False)
nao_infer_queue = torch.utils.data.DataLoader(
nao_infer_dataset,
batch_size=len(nao_infer_dataset),
shuffle=False,
pin_memory=True)
while len(new_archs) < args.new_arch:
predict_step_size += 1
logging.info('Generate new architectures with step size %d',
predict_step_size)
new_arch = nao_infer(nao_infer_queue,
nao,
predict_step_size,
direction='+')
for arch in new_arch:
if arch not in train_encoder_input and arch not in new_archs:
new_archs.append(arch)
if len(new_archs) >= args.new_arch:
break
logging.info('%d new archs generated now', len(new_archs))
if predict_step_size > args.max_step_size:
break
logging.info("Generate %d new archs", len(new_archs))
new_arch_pool = list(
map(lambda x: utils.parse_seq_to_arch(x, 2), new_archs))
new_arch_pool = new_arch_pool + arch_pool[:args.remain_topk]
with open(
os.path.join(args.output_dir,
'new_arch_pool.{}'.format(args.iteration)), 'w') as f:
for arch in new_arch_pool:
arch = ' '.join(map(str, arch[0] + arch[1]))
f.write('{}\n'.format(arch))
logging.info('Finish training!')
def main():
arch_pool = utils.generate_arch(args.controller_seed_arch, 5, 5)
valid_arch_pool = utils.generate_arch(100, 5, 5)
train_encoder_input = list(
map(
lambda x: utils.parse_arch_to_seq(x[0], 2) + utils.
parse_arch_to_seq(x[1], 2), arch_pool))
valid_encoder_input = list(
map(
lambda x: utils.parse_arch_to_seq(x[0], 2) + utils.
parse_arch_to_seq(x[1], 2), valid_arch_pool))
train_encoder_target = [
np.random.random() for i in range(args.controller_seed_arch)
]
valid_encoder_target = [np.random.random() for i in range(100)]
nao = NAO(
args.controller_encoder_layers,
args.controller_encoder_vocab_size,
args.controller_encoder_hidden_size,
args.controller_encoder_dropout,
args.controller_encoder_length,
args.controller_source_length,
args.controller_encoder_emb_size,
args.controller_mlp_layers,
args.controller_mlp_hidden_size,
args.controller_mlp_dropout,
args.controller_decoder_layers,
args.controller_decoder_vocab_size,
args.controller_decoder_hidden_size,
args.controller_decoder_dropout,
args.controller_decoder_length,
)
logging.info("param size = %fMB", utils.count_parameters_in_MB(nao))
nao = nao.cuda()
nao_train_dataset = utils.NAODataset(train_encoder_input,
train_encoder_target,
True,
swap=True)
nao_valid_dataset = utils.NAODataset(valid_encoder_input,
valid_encoder_target, False)
nao_train_queue = torch.utils.data.DataLoader(
nao_train_dataset,
batch_size=args.controller_batch_size,
shuffle=True,
pin_memory=True)
nao_valid_queue = torch.utils.data.DataLoader(
nao_valid_dataset,
batch_size=len(nao_valid_dataset),
shuffle=False,
pin_memory=True)
nao_optimizer = torch.optim.Adam(nao.parameters(),
lr=args.controller_lr,
weight_decay=args.controller_l2_reg)
for nao_epoch in range(1, args.controller_epochs + 1):
nao_loss, nao_mse, nao_ce = nao_train(nao_train_queue, nao,
nao_optimizer)
if nao_epoch % 10 == 0:
logging.info("epoch %04d train loss %.6f mse %.6f ce %.6f",
nao_epoch, nao_loss, nao_mse, nao_ce)
if nao_epoch % 100 == 0:
pa, hs = nao_valid(nao_valid_queue, nao)
logging.info("Evaluation on training data")
logging.info(
'epoch %04d pairwise accuracy %.6f hamming distance %.6f',
nao_epoch, pa, hs)
new_archs = []
max_step_size = 100
predict_step_size = 0
top100_archs = list(
map(
lambda x: utils.parse_arch_to_seq(x[0], 2) + utils.
parse_arch_to_seq(x[1], 2), arch_pool[:100]))
nao_infer_dataset = utils.NAODataset(top100_archs, None, False)
nao_infer_queue = torch.utils.data.DataLoader(
nao_infer_dataset,
batch_size=len(nao_infer_dataset),
shuffle=False,
pin_memory=True)
while len(new_archs) < args.controller_new_arch:
predict_step_size += 1
logging.info('Generate new architectures with step size %d',
predict_step_size)
new_arch = nao_infer(nao_infer_queue, nao, predict_step_size)
for arch in new_arch:
if arch not in train_encoder_input and arch not in new_archs:
new_archs.append(arch)
if len(new_archs) >= args.controller_new_arch:
break
logging.info('%d new archs generated now', len(new_archs))
if predict_step_size > max_step_size:
break
# [[conv, reduc]]
new_archs = list(map(lambda x: utils.parse_seq_to_arch(x, 2),
new_archs)) # [[[conv],[reduc]]]
num_new_archs = len(new_archs)
logging.info("Generate %d new archs", num_new_archs)
new_arch_pool = arch_pool + new_archs + utils.generate_arch(
args.controller_random_arch, 5, 5)
logging.info("Totally %d archs now to train", len(new_arch_pool))
arch_pool = new_arch_pool
def main():
if not torch.cuda.is_available():
logging.info('No GPU found!')
sys.exit(1)
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
cudnn.enabled = True
cudnn.benchmark = True
logging.info("Args = %s", args)
args.source_length = args.encoder_length = args.decoder_length = (
args.nodes + 2) * (args.nodes - 1) // 2
nasbench = api.NASBench(os.path.join(args.data, 'nasbench_full.tfrecord'))
controller = NAO(
args.encoder_layers,
args.decoder_layers,
args.mlp_layers,
args.mlp_hidden_size,
args.hidden_size,
args.vocab_size,
args.dropout,
args.source_length,
args.encoder_length,
args.decoder_length,
)
logging.info("param size = %d", utils.count_parameters(controller))
controller = controller.cuda()
child_arch_pool, child_seq_pool, child_arch_pool_valid_acc = utils.generate_arch(
args.seed_arch, nasbench, need_perf=True)
arch_pool = []
seq_pool = []
arch_pool_valid_acc = []
for i in range(args.iteration + 1):
logging.info('Iteration {}'.format(i + 1))
if not child_arch_pool_valid_acc:
for arch in child_arch_pool:
data = nasbench.query(arch)
child_arch_pool_valid_acc.append(data['validation_accuracy'])
arch_pool += child_arch_pool
arch_pool_valid_acc += child_arch_pool_valid_acc
seq_pool += child_seq_pool
arch_pool_valid_acc_sorted_indices = np.argsort(
arch_pool_valid_acc)[::-1]
arch_pool = [arch_pool[i] for i in arch_pool_valid_acc_sorted_indices]
seq_pool = [seq_pool[i] for i in arch_pool_valid_acc_sorted_indices]
arch_pool_valid_acc = [
arch_pool_valid_acc[i] for i in arch_pool_valid_acc_sorted_indices
]
with open(os.path.join(args.output_dir, 'arch_pool.{}'.format(i)),
'w') as fa:
for arch, seq, valid_acc in zip(arch_pool, seq_pool,
arch_pool_valid_acc):
fa.write('{}\t{}\t{}\t{}\n'.format(arch.matrix, arch.ops, seq,
valid_acc))
for arch_index in range(10):
print('Top 10 architectures:')
print('Architecutre connection:{}'.format(
arch_pool[arch_index].matrix))
print('Architecture operations:{}'.format(
arch_pool[arch_index].ops))
print('Valid accuracy:{}'.format(arch_pool_valid_acc[arch_index]))
if i == args.iteration:
print('Final architectures:')
for arch_index in range(10):
print('Architecutre connection:{}'.format(
arch_pool[arch_index].matrix))
print('Architecture operations:{}'.format(
arch_pool[arch_index].ops))
print('Valid accuracy:{}'.format(
arch_pool_valid_acc[arch_index]))
fs, cs = nasbench.get_metrics_from_spec(arch_pool[arch_index])
test_acc = np.mean(
[cs[108][j]['final_test_accuracy'] for j in range(3)])
print('Mean test accuracy:{}'.format(test_acc))
break
train_encoder_input = seq_pool
min_val = min(arch_pool_valid_acc)
max_val = max(arch_pool_valid_acc)
train_encoder_target = [(i - min_val) / (max_val - min_val)
for i in arch_pool_valid_acc]
# Pre-train
logging.info('Pre-train EPD')
train_controller(controller, train_encoder_input, train_encoder_target,
args.pretrain_epochs)
logging.info('Finish pre-training EPD')
# Generate synthetic data
logging.info('Generate synthetic data for EPD')
synthetic_encoder_input, synthetic_encoder_target = generate_synthetic_controller_data(
nasbench, controller, train_encoder_input, args.random_arch)
if args.up_sample_ratio is None:
up_sample_ratio = np.ceil(args.random_arch /
len(train_encoder_input)).astype(np.int)
else:
up_sample_ratio = args.up_sample_ratio
all_encoder_input = train_encoder_input * up_sample_ratio + synthetic_encoder_input
all_encoder_target = train_encoder_target * up_sample_ratio + synthetic_encoder_target
# Train
logging.info('Train EPD')
train_controller(controller, all_encoder_input, all_encoder_target,
args.epochs)
logging.info('Finish training EPD')
new_archs = []
new_seqs = []
predict_step_size = 0
unique_input = train_encoder_input + synthetic_encoder_input
unique_target = train_encoder_target + synthetic_encoder_target
unique_indices = np.argsort(unique_target)[::-1]
unique_input = [unique_input[i] for i in unique_indices]
topk_archs = unique_input[:args.k]
controller_infer_dataset = utils.ControllerDataset(
topk_archs, None, False)
controller_infer_queue = torch.utils.data.DataLoader(
controller_infer_dataset,
batch_size=len(controller_infer_dataset),
shuffle=False,
pin_memory=True)
while len(new_archs) < args.new_arch:
predict_step_size += 1
logging.info('Generate new architectures with step size %d',
predict_step_size)
new_seq, new_perfs = controller_infer(controller_infer_queue,
controller,
predict_step_size,
direction='+')
for seq in new_seq:
matrix, ops = utils.convert_seq_to_arch(seq)
arch = api.ModelSpec(matrix=matrix, ops=ops)
if nasbench.is_valid(arch) and len(
arch.ops
) == 7 and seq not in train_encoder_input and seq not in new_seqs:
new_archs.append(arch)
new_seqs.append(seq)
if len(new_seqs) >= args.new_arch:
break
logging.info('%d new archs generated now', len(new_archs))
if predict_step_size > args.max_step_size:
break
child_arch_pool = new_archs
child_seq_pool = new_seqs
child_arch_pool_valid_acc = []
child_arch_pool_test_acc = []
logging.info("Generate %d new archs", len(child_arch_pool))
print(nasbench.get_budget_counters())
def main():
if not torch.cuda.is_available():
logging.info('no gpu device available')
sys.exit(1)
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
torch.cuda.manual_seed_all(args.seed)
cudnn.enabled = True
cudnn.benchmark = True
cudnn.deterministic = True
if args.dataset == 'cifar10':
args.num_class = 10
elif args.dataset == 'cifar100':
args.num_class = 100
else:
args.num_class = 10
if args.search_space == 'small':
OPERATIONS = OPERATIONS_search_small
elif args.search_space == 'middle':
OPERATIONS = OPERATIONS_search_middle
args.child_num_ops = len(OPERATIONS)
args.controller_encoder_vocab_size = 1 + (args.child_nodes + 2 -
1) + args.child_num_ops
args.controller_decoder_vocab_size = args.controller_encoder_vocab_size
args.steps = int(np.ceil(
45000 / args.child_batch_size)) * args.child_epochs
logging.info("args = %s", args)
if args.child_arch_pool is not None:
logging.info('Architecture pool is provided, loading')
with open(args.child_arch_pool) as f:
archs = f.read().splitlines()
archs = list(map(utils.build_dag, archs))
child_arch_pool = archs
elif os.path.exists(os.path.join(args.output_dir, 'arch_pool')):
logging.info('Architecture pool is founded, loading')
with open(os.path.join(args.output_dir, 'arch_pool')) as f:
archs = f.read().splitlines()
archs = list(map(utils.build_dag, archs))
child_arch_pool = archs
else:
child_arch_pool = None
build_fn = get_builder(args.dataset)
train_queue, valid_queue, model, train_criterion, eval_criterion, optimizer, scheduler = build_fn(
ratio=0.9, epoch=-1)
nao = NAO(
args.controller_encoder_layers,
args.controller_encoder_vocab_size,
args.controller_encoder_hidden_size,
args.controller_encoder_dropout,
args.controller_encoder_length,
args.controller_source_length,
args.controller_encoder_emb_size,
args.controller_mlp_layers,
args.controller_mlp_hidden_size,
args.controller_mlp_dropout,
args.controller_decoder_layers,
args.controller_decoder_vocab_size,
args.controller_decoder_hidden_size,
args.controller_decoder_dropout,
args.controller_decoder_length,
)
nao = nao.cuda()
logging.info("Encoder-Predictor-Decoder param size = %fMB",
utils.count_parameters_in_MB(nao))
if child_arch_pool is None:
logging.info(
'Architecture pool is not provided, randomly generating now')
child_arch_pool = utils.generate_arch(
args.controller_seed_arch, args.child_nodes,
args.child_num_ops) # [[[conv],[reduc]]]
arch_pool = []
arch_pool_valid_acc = []
for i in range(4):
logging.info('Iteration %d', i)
child_arch_pool_prob = []
for arch in child_arch_pool:
if args.dataset == 'cifar10':
tmp_model = NASNetworkCIFAR(
args, args.num_class, args.child_layers, args.child_nodes,
args.child_channels, args.child_keep_prob,
args.child_drop_path_keep_prob, args.child_use_aux_head,
args.steps, arch)
elif args.dataset == 'cifar100':
tmp_model = NASNetworkCIFAR(
args, args.num_class, args.child_layers, args.child_nodes,
args.child_channels, args.child_keep_prob,
args.child_drop_path_keep_prob, args.child_use_aux_head,
args.steps, arch)
else:
tmp_model = NASNetworkImageNet(
args, args.num_class, args.child_layers, args.child_nodes,
args.child_channels, args.child_keep_prob,
args.child_drop_path_keep_prob, args.child_use_aux_head,
args.steps, arch)
child_arch_pool_prob.append(
utils.count_parameters_in_MB(tmp_model))
del tmp_model
step = 0
scheduler = get_scheduler(optimizer, args.dataset)
for epoch in range(1, args.child_epochs + 1):
scheduler.step()
lr = scheduler.get_lr()[0]
logging.info('epoch %d lr %e', epoch, lr)
# sample an arch to train
train_acc, train_obj, step = child_train(train_queue, model,
optimizer, step,
child_arch_pool,
child_arch_pool_prob,
train_criterion)
logging.info('train_acc %f', train_acc)
logging.info("Evaluate seed archs")
arch_pool += child_arch_pool
arch_pool_valid_acc = child_valid(valid_queue, model, arch_pool,
eval_criterion)
arch_pool_valid_acc_sorted_indices = np.argsort(
arch_pool_valid_acc)[::-1]
arch_pool = [arch_pool[i] for i in arch_pool_valid_acc_sorted_indices]
arch_pool_valid_acc = [
arch_pool_valid_acc[i] for i in arch_pool_valid_acc_sorted_indices
]
with open(os.path.join(args.output_dir, 'arch_pool.{}'.format(i)),
'w') as fa:
with open(
os.path.join(args.output_dir,
'arch_pool.perf.{}'.format(i)), 'w') as fp:
for arch, perf in zip(arch_pool, arch_pool_valid_acc):
arch = ' '.join(map(str, arch[0] + arch[1]))
fa.write('{}\n'.format(arch))
fp.write('{}\n'.format(perf))
if i == 3:
break
# Train Encoder-Predictor-Decoder
logging.info('Train Encoder-Predictor-Decoder')
encoder_input = list(
map(
lambda x: utils.parse_arch_to_seq(x[0]) + utils.
parse_arch_to_seq(x[1]), arch_pool))
# [[conv, reduc]]
min_val = min(arch_pool_valid_acc)
max_val = max(arch_pool_valid_acc)
encoder_target = [(i - min_val) / (max_val - min_val)
for i in arch_pool_valid_acc]
if args.controller_expand:
dataset = list(zip(encoder_input, encoder_target))
n = len(dataset)
ratio = 0.9
split = int(n * ratio)
np.random.shuffle(dataset)
encoder_input, encoder_target = list(zip(*dataset))
train_encoder_input = list(encoder_input[:split])
train_encoder_target = list(encoder_target[:split])
valid_encoder_input = list(encoder_input[split:])
valid_encoder_target = list(encoder_target[split:])
for _ in range(args.controller_expand - 1):
for src, tgt in zip(encoder_input[:split],
encoder_target[:split]):
a = np.random.randint(0, args.child_nodes)
b = np.random.randint(0, args.child_nodes)
src = src[:4 * a] + src[4 * a + 2:4 * a + 4] + \
src[4 * a:4 * a + 2] + src[4 * (a + 1):20 + 4 * b] + \
src[20 + 4 * b + 2:20 + 4 * b + 4] + src[20 + 4 * b:20 + 4 * b + 2] + \
src[20 + 4 * (b + 1):]
train_encoder_input.append(src)
train_encoder_target.append(tgt)
else:
train_encoder_input = encoder_input
train_encoder_target = encoder_target
valid_encoder_input = encoder_input
valid_encoder_target = encoder_target
logging.info('Train data: {}\tValid data: {}'.format(
len(train_encoder_input), len(valid_encoder_input)))
nao_train_dataset = utils.NAODataset(
train_encoder_input,
train_encoder_target,
True,
swap=True if args.controller_expand is None else False)
nao_valid_dataset = utils.NAODataset(valid_encoder_input,
valid_encoder_target, False)
nao_train_queue = torch.utils.data.DataLoader(
nao_train_dataset,
batch_size=args.controller_batch_size,
shuffle=True,
pin_memory=True)
nao_valid_queue = torch.utils.data.DataLoader(
nao_valid_dataset,
batch_size=args.controller_batch_size,
shuffle=False,
pin_memory=True)
nao_optimizer = torch.optim.Adam(nao.parameters(),
lr=args.controller_lr,
weight_decay=args.controller_l2_reg)
for nao_epoch in range(1, args.controller_epochs + 1):
nao_loss, nao_mse, nao_ce = nao_train(nao_train_queue, nao,
nao_optimizer)
logging.info("epoch %04d train loss %.6f mse %.6f ce %.6f",
nao_epoch, nao_loss, nao_mse, nao_ce)
if nao_epoch % 100 == 0:
pa, hs = nao_valid(nao_valid_queue, nao)
logging.info("Evaluation on valid data")
logging.info(
'epoch %04d pairwise accuracy %.6f hamming distance %.6f',
nao_epoch, pa, hs)
# Generate new archs
new_archs = []
max_step_size = 50
predict_step_size = 0
top100_archs = list(
map(
lambda x: utils.parse_arch_to_seq(x[0]) + utils.
parse_arch_to_seq(x[1]), arch_pool[:100]))
nao_infer_dataset = utils.NAODataset(top100_archs, None, False)
nao_infer_queue = torch.utils.data.DataLoader(
nao_infer_dataset,
batch_size=len(nao_infer_dataset),
shuffle=False,
pin_memory=True)
while len(new_archs) < args.controller_new_arch:
predict_step_size += 1
logging.info('Generate new architectures with step size %d',
predict_step_size)
new_arch = nao_infer(nao_infer_queue,
nao,
predict_step_size,
direction='+')
for arch in new_arch:
if arch not in encoder_input and arch not in new_archs:
new_archs.append(arch)
if len(new_archs) >= args.controller_new_arch:
break
logging.info('%d new archs generated now', len(new_archs))
if predict_step_size > max_step_size:
break
child_arch_pool = list(
map(lambda x: utils.parse_seq_to_arch(x),
new_archs)) # [[[conv],[reduc]]]
logging.info("Generate %d new archs", len(child_arch_pool))
logging.info('Finish Searching')
logging.info('Reranking top 5 architectures')
# reranking top 5
top_archs = arch_pool[:5]
if args.dataset == 'cifar10':
top_archs_perf = train_and_evaluate_top_on_cifar10(
top_archs, train_queue, valid_queue)
elif args.dataset == 'cifar100':
top_archs_perf = train_and_evaluate_top_on_cifar100(
top_archs, train_queue, valid_queue)
else:
top_archs_perf = train_and_evaluate_top_on_imagenet(
top_archs, train_queue, valid_queue)
top_archs_sorted_indices = np.argsort(top_archs_perf)[::-1]
top_archs = [top_archs[i] for i in top_archs_sorted_indices]
top_archs_perf = [top_archs_perf[i] for i in top_archs_sorted_indices]
with open(os.path.join(args.output_dir, 'arch_pool.final'), 'w') as fa:
with open(os.path.join(args.output_dir, 'arch_pool.perf.final'),
'w') as fp:
for arch, perf in zip(top_archs, top_archs_perf):
arch = ' '.join(map(str, arch[0] + arch[1]))
fa.write('{}\n'.format(arch))
fp.write('{}\n'.format(perf))