def generate_meta_info(save_dir, max_node, divide=40):
aa_nas_bench_ss = get_search_spaces("cell", "nas-bench-201")
archs = CellStructure.gen_all(aa_nas_bench_ss, max_node, False)
print("There are {:} archs vs {:}.".format(
len(archs),
len(aa_nas_bench_ss)**((max_node - 1) * max_node / 2)))
random.seed(88) # please do not change this line for reproducibility
random.shuffle(archs)
# to test fixed-random shuffle
# print ('arch [0] : {:}\n---->>>> {:}'.format( archs[0], archs[0].tostr() ))
# print ('arch [9] : {:}\n---->>>> {:}'.format( archs[9], archs[9].tostr() ))
assert (
archs[0].tostr() ==
"|avg_pool_3x3~0|+|nor_conv_1x1~0|skip_connect~1|+|nor_conv_1x1~0|skip_connect~1|skip_connect~2|"
), "please check the 0-th architecture : {:}".format(archs[0])
assert (
archs[9].tostr() ==
"|avg_pool_3x3~0|+|none~0|none~1|+|skip_connect~0|none~1|nor_conv_3x3~2|"
), "please check the 9-th architecture : {:}".format(archs[9])
assert (
archs[123].tostr() ==
"|avg_pool_3x3~0|+|avg_pool_3x3~0|nor_conv_1x1~1|+|none~0|avg_pool_3x3~1|nor_conv_3x3~2|"
), "please check the 123-th architecture : {:}".format(archs[123])
total_arch = len(archs)
num = 50000
indexes_5W = list(range(num))
random.seed(1021)
random.shuffle(indexes_5W)
train_split = sorted(list(set(indexes_5W[:num // 2])))
valid_split = sorted(list(set(indexes_5W[num // 2:])))
assert len(train_split) + len(valid_split) == num
assert (train_split[0] == 0 and train_split[10] == 26
and train_split[111] == 203 and valid_split[0] == 1
and valid_split[10] == 18
and valid_split[111] == 242), "{:} {:} {:} - {:} {:} {:}".format(
train_split[0],
train_split[10],
train_split[111],
valid_split[0],
valid_split[10],
valid_split[111],
)
splits = {num: {"train": train_split, "valid": valid_split}}
info = {
"archs": [x.tostr() for x in archs],
"total": total_arch,
"max_node": max_node,
"splits": splits,
}
save_dir = Path(save_dir)
save_dir.mkdir(parents=True, exist_ok=True)
save_name = save_dir / "meta-node-{:}.pth".format(max_node)
assert not save_name.exists(), "{:} already exist".format(save_name)
torch.save(info, save_name)
print("save the meta file into {:}".format(save_name))
def config2structure(config):
genotypes = []
for i in range(1, max_nodes):
xlist = []
for j in range(i):
node_str = "{:}<-{:}".format(i, j)
op_name = config[node_str]
xlist.append((op_name, j))
genotypes.append(tuple(xlist))
return CellStructure(genotypes)
def random_architecture():
genotypes = []
for i in range(1, max_nodes):
xlist = []
for j in range(i):
node_str = "{:}<-{:}".format(i, j)
op_name = random.choice(op_names)
xlist.append((op_name, j))
genotypes.append(tuple(xlist))
return CellStructure(genotypes)
def generate_arch(self, actions):
genotypes = []
for i in range(1, self.max_nodes):
xlist = []
for j in range(i):
node_str = "{:}<-{:}".format(i, j)
op_name = self.search_space[actions[self.edge2index[node_str]]]
xlist.append((op_name, j))
genotypes.append(tuple(xlist))
return CellStructure(genotypes)
def genotype(self):
genotypes = []
for i in range(1, self.max_nodes):
xlist = []
for j in range(i):
node_str = "{:}<-{:}".format(i, j)
with torch.no_grad():
weights = self.arch_parameters[self.edge2index[node_str]]
op_name = self.search_space[weights.argmax().item()]
xlist.append((op_name, j))
genotypes.append(tuple(xlist))
return CellStructure(genotypes)
def check_unique_arch(meta_file):
api = API(str(meta_file))
arch_strs = deepcopy(api.meta_archs)
xarchs = [CellStructure.str2structure(x) for x in arch_strs]
def get_unique_matrix(archs, consider_zero):
UniquStrs = [arch.to_unique_str(consider_zero) for arch in archs]
print("{:} create unique-string ({:}/{:}) done".format(
time_string(), len(set(UniquStrs)), len(UniquStrs)))
Unique2Index = dict()
for index, xstr in enumerate(UniquStrs):
if xstr not in Unique2Index:
Unique2Index[xstr] = list()
Unique2Index[xstr].append(index)
sm_matrix = torch.eye(len(archs)).bool()
for _, xlist in Unique2Index.items():
for i in xlist:
for j in xlist:
sm_matrix[i, j] = True
unique_ids, unique_num = [-1 for _ in archs], 0
for i in range(len(unique_ids)):
if unique_ids[i] > -1:
continue
neighbours = sm_matrix[i].nonzero().view(-1).tolist()
for nghb in neighbours:
assert unique_ids[nghb] == -1, "impossible"
unique_ids[nghb] = unique_num
unique_num += 1
return sm_matrix, unique_ids, unique_num
print("There are {:} valid-archs".format(
sum(arch.check_valid() for arch in xarchs)))
sm_matrix, uniqueIDs, unique_num = get_unique_matrix(xarchs, None)
print(
"{:} There are {:} unique architectures (considering nothing).".format(
time_string(), unique_num))
sm_matrix, uniqueIDs, unique_num = get_unique_matrix(xarchs, False)
print("{:} There are {:} unique architectures (not considering zero).".
format(time_string(), unique_num))
sm_matrix, uniqueIDs, unique_num = get_unique_matrix(xarchs, True)
print("{:} There are {:} unique architectures (considering zero).".format(
time_string(), unique_num))
def traverse_net(max_node):
aa_nas_bench_ss = get_search_spaces("cell", "nats-bench")
archs = CellStructure.gen_all(aa_nas_bench_ss, max_node, False)
print("There are {:} archs vs {:}.".format(
len(archs),
len(aa_nas_bench_ss)**((max_node - 1) * max_node / 2)))
random.seed(88) # please do not change this line for reproducibility
random.shuffle(archs)
assert (
archs[0].tostr() ==
"|avg_pool_3x3~0|+|nor_conv_1x1~0|skip_connect~1|+|nor_conv_1x1~0|skip_connect~1|skip_connect~2|"
), "please check the 0-th architecture : {:}".format(archs[0])
assert (
archs[9].tostr() ==
"|avg_pool_3x3~0|+|none~0|none~1|+|skip_connect~0|none~1|nor_conv_3x3~2|"
), "please check the 9-th architecture : {:}".format(archs[9])
assert (
archs[123].tostr() ==
"|avg_pool_3x3~0|+|avg_pool_3x3~0|nor_conv_1x1~1|+|none~0|avg_pool_3x3~1|nor_conv_3x3~2|"
), "please check the 123-th architecture : {:}".format(archs[123])
return [x.tostr() for x in archs]
def generate_meta_info(save_dir, max_node, divide=40):
aa_nas_bench_ss = get_search_spaces("cell", "nas-bench-201")
archs = CellStructure.gen_all(aa_nas_bench_ss, max_node, False)
print("There are {:} archs vs {:}.".format(
len(archs),
len(aa_nas_bench_ss)**((max_node - 1) * max_node / 2)))
random.seed(88) # please do not change this line for reproducibility
random.shuffle(archs)
# to test fixed-random shuffle
# print ('arch [0] : {:}\n---->>>> {:}'.format( archs[0], archs[0].tostr() ))
# print ('arch [9] : {:}\n---->>>> {:}'.format( archs[9], archs[9].tostr() ))
assert (
archs[0].tostr() ==
"|avg_pool_3x3~0|+|nor_conv_1x1~0|skip_connect~1|+|nor_conv_1x1~0|skip_connect~1|skip_connect~2|"
), "please check the 0-th architecture : {:}".format(archs[0])
assert (
archs[9].tostr() ==
"|avg_pool_3x3~0|+|none~0|none~1|+|skip_connect~0|none~1|nor_conv_3x3~2|"
), "please check the 9-th architecture : {:}".format(archs[9])
assert (
archs[123].tostr() ==
"|avg_pool_3x3~0|+|avg_pool_3x3~0|nor_conv_1x1~1|+|none~0|avg_pool_3x3~1|nor_conv_3x3~2|"
), "please check the 123-th architecture : {:}".format(archs[123])
total_arch = len(archs)
num = 50000
indexes_5W = list(range(num))
random.seed(1021)
random.shuffle(indexes_5W)
train_split = sorted(list(set(indexes_5W[:num // 2])))
valid_split = sorted(list(set(indexes_5W[num // 2:])))
assert len(train_split) + len(valid_split) == num
assert (train_split[0] == 0 and train_split[10] == 26
and train_split[111] == 203 and valid_split[0] == 1
and valid_split[10] == 18
and valid_split[111] == 242), "{:} {:} {:} - {:} {:} {:}".format(
train_split[0],
train_split[10],
train_split[111],
valid_split[0],
valid_split[10],
valid_split[111],
)
splits = {num: {"train": train_split, "valid": valid_split}}
info = {
"archs": [x.tostr() for x in archs],
"total": total_arch,
"max_node": max_node,
"splits": splits,
}
save_dir = Path(save_dir)
save_dir.mkdir(parents=True, exist_ok=True)
save_name = save_dir / "meta-node-{:}.pth".format(max_node)
assert not save_name.exists(), "{:} already exist".format(save_name)
torch.save(info, save_name)
print("save the meta file into {:}".format(save_name))
script_name_full = save_dir / "BENCH-201-N{:}.opt-full.script".format(
max_node)
script_name_less = save_dir / "BENCH-201-N{:}.opt-less.script".format(
max_node)
full_file = open(str(script_name_full), "w")
less_file = open(str(script_name_less), "w")
gaps = total_arch // divide
for start in range(0, total_arch, gaps):
xend = min(start + gaps, total_arch)
full_file.write(
"bash ./scripts-search/NAS-Bench-201/train-models.sh 0 {:5d} {:5d} -1 '777 888 999'\n"
.format(start, xend - 1))
less_file.write(
"bash ./scripts-search/NAS-Bench-201/train-models.sh 1 {:5d} {:5d} -1 '777 888 999'\n"
.format(start, xend - 1))
print("save the training script into {:} and {:}".format(
script_name_full, script_name_less))
full_file.close()
less_file.close()
script_name = save_dir / "meta-node-{:}.cal-script.txt".format(max_node)
macro = "OMP_NUM_THREADS=6 CUDA_VISIBLE_DEVICES=0"
with open(str(script_name), "w") as cfile:
for start in range(0, total_arch, gaps):
xend = min(start + gaps, total_arch)
cfile.write(
"{:} python exps/NAS-Bench-201/statistics.py --mode cal --target_dir {:06d}-{:06d}-C16-N5\n"
.format(macro, start, xend - 1))
print("save the post-processing script into {:}".format(script_name))
def train_single_model(save_dir, workers, datasets, xpaths, splits, use_less,
seeds, model_str, arch_config):
assert torch.cuda.is_available(), "CUDA is not available."
torch.backends.cudnn.enabled = True
torch.backends.cudnn.deterministic = True
# torch.backends.cudnn.benchmark = True
torch.set_num_threads(workers)
save_dir = (Path(save_dir) / "specifics" / "{:}-{:}-{:}-{:}".format(
"LESS" if use_less else "FULL",
model_str,
arch_config["channel"],
arch_config["num_cells"],
))
logger = Logger(str(save_dir), 0, False)
if model_str in CellArchitectures:
arch = CellArchitectures[model_str]
logger.log(
"The model string is found in pre-defined architecture dict : {:}".
format(model_str))
else:
try:
arch = CellStructure.str2structure(model_str)
except:
raise ValueError(
"Invalid model string : {:}. It can not be found or parsed.".
format(model_str))
assert arch.check_valid_op(get_search_spaces(
"cell", "full")), "{:} has the invalid op.".format(arch)
logger.log("Start train-evaluate {:}".format(arch.tostr()))
logger.log("arch_config : {:}".format(arch_config))
start_time, seed_time = time.time(), AverageMeter()
for _is, seed in enumerate(seeds):
logger.log(
"\nThe {:02d}/{:02d}-th seed is {:} ----------------------<.>----------------------"
.format(_is, len(seeds), seed))
to_save_name = save_dir / "seed-{:04d}.pth".format(seed)
if to_save_name.exists():
logger.log("Find the existing file {:}, directly load!".format(
to_save_name))
checkpoint = torch.load(to_save_name)
else:
logger.log(
"Does not find the existing file {:}, train and evaluate!".
format(to_save_name))
checkpoint = evaluate_all_datasets(
arch,
datasets,
xpaths,
splits,
use_less,
seed,
arch_config,
workers,
logger,
)
torch.save(checkpoint, to_save_name)
# log information
logger.log("{:}".format(checkpoint["info"]))
all_dataset_keys = checkpoint["all_dataset_keys"]
for dataset_key in all_dataset_keys:
logger.log("\n{:} dataset : {:} {:}".format(
"-" * 15, dataset_key, "-" * 15))
dataset_info = checkpoint[dataset_key]
# logger.log('Network ==>\n{:}'.format( dataset_info['net_string'] ))
logger.log("Flops = {:} MB, Params = {:} MB".format(
dataset_info["flop"], dataset_info["param"]))
logger.log("config : {:}".format(dataset_info["config"]))
logger.log("Training State (finish) = {:}".format(
dataset_info["finish-train"]))
last_epoch = dataset_info["total_epoch"] - 1
train_acc1es, train_acc5es = (
dataset_info["train_acc1es"],
dataset_info["train_acc5es"],
)
valid_acc1es, valid_acc5es = (
dataset_info["valid_acc1es"],
dataset_info["valid_acc5es"],
)
logger.log(
"Last Info : Train = Acc@1 {:.2f}% Acc@5 {:.2f}% Error@1 {:.2f}%, Test = Acc@1 {:.2f}% Acc@5 {:.2f}% Error@1 {:.2f}%"
.format(
train_acc1es[last_epoch],
train_acc5es[last_epoch],
100 - train_acc1es[last_epoch],
valid_acc1es[last_epoch],
valid_acc5es[last_epoch],
100 - valid_acc1es[last_epoch],
))
# measure elapsed time
seed_time.update(time.time() - start_time)
start_time = time.time()
need_time = "Time Left: {:}".format(
convert_secs2time(seed_time.avg * (len(seeds) - _is - 1), True))
logger.log(
"\n<<<***>>> The {:02d}/{:02d}-th seed is {:} <finish> other procedures need {:}"
.format(_is, len(seeds), seed, need_time))
logger.close()
def main(
save_dir,
workers,
datasets,
xpaths,
splits,
use_less,
srange,
arch_index,
seeds,
cover_mode,
meta_info,
arch_config,
):
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(workers)
assert (len(srange) == 2 and
0 <= srange[0] <= srange[1]), "invalid srange : {:}".format(srange)
if use_less:
sub_dir = Path(save_dir) / "{:06d}-{:06d}-C{:}-N{:}-LESS".format(
srange[0], srange[1], arch_config["channel"],
arch_config["num_cells"])
else:
sub_dir = Path(save_dir) / "{:06d}-{:06d}-C{:}-N{:}".format(
srange[0], srange[1], arch_config["channel"],
arch_config["num_cells"])
logger = Logger(str(sub_dir), 0, False)
all_archs = meta_info["archs"]
assert srange[1] < meta_info[
"total"], "invalid range : {:}-{:} vs. {:}".format(
srange[0], srange[1], meta_info["total"])
assert (arch_index == -1 or srange[0] <= arch_index <= srange[1]
), "invalid range : {:} vs. {:} vs. {:}".format(
srange[0], arch_index, srange[1])
if arch_index == -1:
to_evaluate_indexes = list(range(srange[0], srange[1] + 1))
else:
to_evaluate_indexes = [arch_index]
logger.log("xargs : seeds = {:}".format(seeds))
logger.log("xargs : arch_index = {:}".format(arch_index))
logger.log("xargs : cover_mode = {:}".format(cover_mode))
logger.log("-" * 100)
logger.log(
"Start evaluating range =: {:06d} vs. {:06d} vs. {:06d} / {:06d} with cover-mode={:}"
.format(srange[0], arch_index, srange[1], meta_info["total"],
cover_mode))
for i, (dataset, xpath, split) in enumerate(zip(datasets, xpaths, splits)):
logger.log(
"--->>> Evaluate {:}/{:} : dataset={:9s}, path={:}, split={:}".
format(i, len(datasets), dataset, xpath, split))
logger.log("--->>> architecture config : {:}".format(arch_config))
start_time, epoch_time = time.time(), AverageMeter()
for i, index in enumerate(to_evaluate_indexes):
arch = all_archs[index]
logger.log(
"\n{:} evaluate {:06d}/{:06d} ({:06d}/{:06d})-th architecture [seeds={:}] {:}"
.format(
"-" * 15,
i,
len(to_evaluate_indexes),
index,
meta_info["total"],
seeds,
"-" * 15,
))
# logger.log('{:} {:} {:}'.format('-'*15, arch.tostr(), '-'*15))
logger.log("{:} {:} {:}".format("-" * 15, arch, "-" * 15))
# test this arch on different datasets with different seeds
has_continue = False
for seed in seeds:
to_save_name = sub_dir / "arch-{:06d}-seed-{:04d}.pth".format(
index, seed)
if to_save_name.exists():
if cover_mode:
logger.log(
"Find existing file : {:}, remove it before evaluation"
.format(to_save_name))
os.remove(str(to_save_name))
else:
logger.log(
"Find existing file : {:}, skip this evaluation".
format(to_save_name))
has_continue = True
continue
results = evaluate_all_datasets(
CellStructure.str2structure(arch),
datasets,
xpaths,
splits,
use_less,
seed,
arch_config,
workers,
logger,
)
torch.save(results, to_save_name)
logger.log(
"{:} --evaluate-- {:06d}/{:06d} ({:06d}/{:06d})-th seed={:} done, save into {:}"
.format(
"-" * 15,
i,
len(to_evaluate_indexes),
index,
meta_info["total"],
seed,
to_save_name,
))
# measure elapsed time
if not has_continue:
epoch_time.update(time.time() - start_time)
start_time = time.time()
need_time = "Time Left: {:}".format(
convert_secs2time(
epoch_time.avg * (len(to_evaluate_indexes) - i - 1), True))
logger.log("This arch costs : {:}".format(
convert_secs2time(epoch_time.val, True)))
logger.log("{:}".format("*" * 100))
logger.log("{:} {:74s} {:}".format(
"*" * 10,
"{:06d}/{:06d} ({:06d}/{:06d})-th done, left {:}".format(
i, len(to_evaluate_indexes), index, meta_info["total"],
need_time),
"*" * 10,
))
logger.log("{:}".format("*" * 100))
logger.close()
def main(
save_dir: Path,
workers: int,
datasets: List[Text],
xpaths: List[Text],
splits: List[int],
seeds: List[int],
nets: List[str],
opt_config: Dict[Text, Any],
to_evaluate_indexes: tuple,
cover_mode: bool,
arch_config: Dict[Text, Any],
):
log_dir = save_dir / "logs"
log_dir.mkdir(parents=True, exist_ok=True)
logger = Logger(str(log_dir), os.getpid(), False)
logger.log("xargs : seeds = {:}".format(seeds))
logger.log("xargs : cover_mode = {:}".format(cover_mode))
logger.log("-" * 100)
logger.log("Start evaluating range =: {:06d} - {:06d}".format(
min(to_evaluate_indexes), max(to_evaluate_indexes)) +
"({:} in total) / {:06d} with cover-mode={:}".format(
len(to_evaluate_indexes), len(nets), cover_mode))
for i, (dataset, xpath, split) in enumerate(zip(datasets, xpaths, splits)):
logger.log(
"--->>> Evaluate {:}/{:} : dataset={:9s}, path={:}, split={:}".
format(i, len(datasets), dataset, xpath, split))
logger.log("--->>> optimization config : {:}".format(opt_config))
start_time, epoch_time = time.time(), AverageMeter()
for i, index in enumerate(to_evaluate_indexes):
arch = nets[index]
logger.log(
"\n{:} evaluate {:06d}/{:06d} ({:06d}/{:06d})-th arch [seeds={:}] {:}"
.format(
time_string(),
i,
len(to_evaluate_indexes),
index,
len(nets),
seeds,
"-" * 15,
))
logger.log("{:} {:} {:}".format("-" * 15, arch, "-" * 15))
# test this arch on different datasets with different seeds
has_continue = False
for seed in seeds:
to_save_name = save_dir / "arch-{:06d}-seed-{:04d}.pth".format(
index, seed)
if to_save_name.exists():
if cover_mode:
logger.log(
"Find existing file : {:}, remove it before evaluation"
.format(to_save_name))
os.remove(str(to_save_name))
else:
logger.log(
"Find existing file : {:}, skip this evaluation".
format(to_save_name))
has_continue = True
continue
results = evaluate_all_datasets(
CellStructure.str2structure(arch),
datasets,
xpaths,
splits,
opt_config,
seed,
arch_config,
workers,
logger,
)
torch.save(results, to_save_name)
logger.log(
"\n{:} evaluate {:06d}/{:06d} ({:06d}/{:06d})-th arch [seeds={:}] ===>>> {:}"
.format(
time_string(),
i,
len(to_evaluate_indexes),
index,
len(nets),
seeds,
to_save_name,
))
# measure elapsed time
if not has_continue:
epoch_time.update(time.time() - start_time)
start_time = time.time()
need_time = "Time Left: {:}".format(
convert_secs2time(
epoch_time.avg * (len(to_evaluate_indexes) - i - 1), True))
logger.log("This arch costs : {:}".format(
convert_secs2time(epoch_time.val, True)))
logger.log("{:}".format("*" * 100))
logger.log("{:} {:74s} {:}".format(
"*" * 10,
"{:06d}/{:06d} ({:06d}/{:06d})-th done, left {:}".format(
i, len(to_evaluate_indexes), index, len(nets), need_time),
"*" * 10,
))
logger.log("{:}".format("*" * 100))
logger.close()
def create_result_count(used_seed, dataset, arch_config, results,
dataloader_dict):
xresult = ResultsCount(
dataset,
results["net_state_dict"],
results["train_acc1es"],
results["train_losses"],
results["param"],
results["flop"],
arch_config,
used_seed,
results["total_epoch"],
None,
)
net_config = dict2config(
{
"name": "infer.tiny",
"C": arch_config["channel"],
"N": arch_config["num_cells"],
"genotype": CellStructure.str2structure(arch_config["arch_str"]),
"num_classes": arch_config["class_num"],
},
None,
)
network = get_cell_based_tiny_net(net_config)
network.load_state_dict(xresult.get_net_param())
if "train_times" in results: # new version
xresult.update_train_info(
results["train_acc1es"],
results["train_acc5es"],
results["train_losses"],
results["train_times"],
)
xresult.update_eval(results["valid_acc1es"], results["valid_losses"],
results["valid_times"])
else:
if dataset == "cifar10-valid":
xresult.update_OLD_eval("x-valid", results["valid_acc1es"],
results["valid_losses"])
loss, top1, top5, latencies = pure_evaluate(
dataloader_dict["{:}@{:}".format("cifar10", "test")],
network.cuda())
xresult.update_OLD_eval(
"ori-test",
{results["total_epoch"] - 1: top1},
{results["total_epoch"] - 1: loss},
)
xresult.update_latency(latencies)
elif dataset == "cifar10":
xresult.update_OLD_eval("ori-test", results["valid_acc1es"],
results["valid_losses"])
loss, top1, top5, latencies = pure_evaluate(
dataloader_dict["{:}@{:}".format(dataset, "test")],
network.cuda())
xresult.update_latency(latencies)
elif dataset == "cifar100" or dataset == "ImageNet16-120":
xresult.update_OLD_eval("ori-test", results["valid_acc1es"],
results["valid_losses"])
loss, top1, top5, latencies = pure_evaluate(
dataloader_dict["{:}@{:}".format(dataset, "valid")],
network.cuda())
xresult.update_OLD_eval(
"x-valid",
{results["total_epoch"] - 1: top1},
{results["total_epoch"] - 1: loss},
)
loss, top1, top5, latencies = pure_evaluate(
dataloader_dict["{:}@{:}".format(dataset, "test")],
network.cuda())
xresult.update_OLD_eval(
"x-test",
{results["total_epoch"] - 1: top1},
{results["total_epoch"] - 1: loss},
)
xresult.update_latency(latencies)
else:
raise ValueError("invalid dataset name : {:}".format(dataset))
return xresult
