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

Replace nats_bench by soft link

Browse Source
This commit is contained in:
D-X-Y
2021-03-17 10:06:29 +00:00
parent 9046a4e87c
commit 95e304495f
13 changed files with 48 additions and 2047 deletions
Download Patch File Download Diff File Expand all files Collapse all files

3
.gitmodules vendored
View File

@@ -1,3 +1,6 @@
[submodule ".latent-data/qlib"]
path = .latent-data/qlib
url = git@github.com:D-X-Y/qlib.git
[submodule ".latent-data/NATS-Bench"]
path = .latent-data/NATS-Bench
url = git@github.com:D-X-Y/NATS-Bench.git

1
.latent-data/NATS-Bench Submodule

Submodule .latent-data/NATS-Bench added at 51187c1e91

6
README.md
View File

@@ -94,6 +94,12 @@ Some visualization codes may require `opencv`.
CIFAR and ImageNet should be downloaded and extracted into `$TORCH_HOME`.
Some methods use knowledge distillation (KD), which require pre-trained models. Please download these models from [Google Drive](https://drive.google.com/open?id=1ANmiYEGX-IQZTfH8w0aSpj-Wypg-0DR-) (or train by yourself) and save into `.latent-data`.
Please use
```
git clone --recurse-submodules git@github.com:D-X-Y/AutoDL-Projects.git
```
to download this repo with submodules.
## Citation
If you find that this project helps your research, please consider citing the related paper:

1
docs/NATS-Bench.md
View File

@@ -8,6 +8,7 @@ We also show the versatility of NATS-Bench by benchmarking 13 recent state-of-th
This facilitates a much larger community of researchers to focus on developing better NAS algorithms in a more comparable and computationally effective environment.
**You can use `pip install nats_bench` to install the library of NATS-Bench.**
or install from the [source codes](https://github.com/D-X-Y/NATS-Bench) via `python setup.py install`.
The structure of this Markdown file:
- [How to use NATS-Bench?](#How-to-Use-NATS-Bench)

5
exps/NATS-algos/README.md
View File

@@ -9,6 +9,11 @@ The Python files in this folder are used to re-produce the results in ``NATS-Ben
- [`regularized_ea.py`](https://github.com/D-X-Y/AutoDL-Projects/blob/main/exps/NATS-algos/regularized_ea.py) contains the REA algorithm for both search spaces.
- [`reinforce.py`](https://github.com/D-X-Y/AutoDL-Projects/blob/main/exps/NATS-algos/reinforce.py) contains the REINFORCE algorithm for both search spaces.
## Requirements
- `nats_bench`>=v1.1 : you can use `pip install nats_bench` to install or from [sources](https://github.com/D-X-Y/NATS-Bench)
- `hpbandster` : if you want to run BOHB
## Citation
If you find that this project helps your research, please consider citing the related paper:

24
lib/layers/super_mlp.py Normal file
View File

@@ -0,0 +1,24 @@
import torch.nn as nn
from typing import Optional
class MLP(nn.Module):
# MLP: FC -> Activation -> Drop -> FC -> Drop
def __init__(self, in_features, hidden_features: Optional[int] = None,
out_features: Optional[int] = None,
act_layer=nn.GELU,
drop: Optional[float] = None):
super(MLP, self).__init__()
out_features = out_features or in_features
hidden_features = hidden_features or in_features
self.fc1 = nn.Linear(in_features, hidden_features)
self.act = act_layer()
self.fc2 = nn.Linear(hidden_features, out_features)
self.drop = nn.Dropout(drop or 0)
def forward(self, x):
x = self.fc1(x)
x = self.act(x)
x = self.drop(x)
x = self.fc2(x)
x = self.drop(x)
return x

7
lib/layers/super_module.py Normal file
View File

@@ -0,0 +1,7 @@
import torch.nn as nn
class SuperModule(nn.Module):
def __init__(self):
super(SuperModule, self).__init__()

70
lib/nats_bench/__init__.py
View File

@@ -1,70 +0,0 @@
##############################################################################
# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2020.08 ##########################
##############################################################################
# NATS-Bench: Benchmarking NAS algorithms for Architecture Topology and Size #
##############################################################################
"""The official Application Programming Interface (API) for NATS-Bench."""
from nats_bench.api_size import NATSsize
from nats_bench.api_topology import NATStopology
from nats_bench.api_utils import ArchResults
from nats_bench.api_utils import pickle_load
from nats_bench.api_utils import pickle_save
from nats_bench.api_utils import ResultsCount
NATS_BENCH_API_VERSIONs = ['v1.0', # [2020.08.31]
'v1.1'] # [2020.12.20] adding unit tests
NATS_BENCH_SSS_NAMEs = ('sss', 'size')
NATS_BENCH_TSS_NAMEs = ('tss', 'topology')
def version():
return NATS_BENCH_API_VERSIONs[-1]
def create(file_path_or_dict, search_space, fast_mode=False, verbose=True):
"""Create the instead for NATS API.
Args:
file_path_or_dict: None or a file path or a directory path.
search_space: This is a string indicates the search space in NATS-Bench.
fast_mode: If True, we will not load all the data at initialization,
instead, the data for each candidate architecture will be loaded when
quering it; If False, we will load all the data during initialization.
verbose: This is a flag to indicate whether log additional information.
Raises:
ValueError: If not find the matched serach space description.
Returns:
The created NATS-Bench API.
"""
if search_space in NATS_BENCH_TSS_NAMEs:
return NATStopology(file_path_or_dict, fast_mode, verbose)
elif search_space in NATS_BENCH_SSS_NAMEs:
return NATSsize(file_path_or_dict, fast_mode, verbose)
else:
raise ValueError('invalid search space : {:}'.format(search_space))
def search_space_info(main_tag, aux_tag):
"""Obtain the search space information."""
nats_sss = dict(candidates=[8, 16, 24, 32, 40, 48, 56, 64],
num_layers=5)
nats_tss = dict(op_names=['none', 'skip_connect',
'nor_conv_1x1', 'nor_conv_3x3',
'avg_pool_3x3'],
num_nodes=4)
if main_tag == 'nats-bench':
if aux_tag in NATS_BENCH_SSS_NAMEs:
return nats_sss
elif aux_tag in NATS_BENCH_TSS_NAMEs:
return nats_tss
else:
raise ValueError('Unknown auxiliary tag: {:}'.format(aux_tag))
elif main_tag == 'nas-bench-201':
if aux_tag is not None:
raise ValueError('For NAS-Bench-201, the auxiliary tag should be None.')
return nats_tss
else:
raise ValueError('Unknown main tag: {:}'.format(main_tag))

291
lib/nats_bench/api_size.py
View File

@@ -1,291 +0,0 @@
#####################################################
# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2020.08 #
##############################################################################
# NATS-Bench: Benchmarking NAS algorithms for Architecture Topology and Size #
##############################################################################
# The history of benchmark files are as follows, #
# where the format is (the name is NATS-sss-[version]-[md5].pickle.pbz2) #
# [2020.08.31] NATS-sss-v1_0-50262.pickle.pbz2 #
##############################################################################
# pylint: disable=line-too-long
"""The API for size search space in NATS-Bench."""
import collections
import copy
import os
import random
from typing import Dict, Optional, Text, Union, Any
from nats_bench.api_utils import ArchResults
from nats_bench.api_utils import NASBenchMetaAPI
from nats_bench.api_utils import get_torch_home
from nats_bench.api_utils import nats_is_dir
from nats_bench.api_utils import nats_is_file
from nats_bench.api_utils import PICKLE_EXT
from nats_bench.api_utils import pickle_load
from nats_bench.api_utils import time_string
ALL_BASE_NAMES = ['NATS-sss-v1_0-50262']
def print_information(information, extra_info=None, show=False):
"""print out the information of a given ArchResults."""
dataset_names = information.get_dataset_names()
strings = [
information.arch_str,
'datasets : {:}, extra-info : {:}'.format(dataset_names, extra_info)
]
def metric2str(loss, acc):
return 'loss = {:.3f} & top1 = {:.2f}%'.format(loss, acc)
for dataset in dataset_names:
metric = information.get_compute_costs(dataset)
flop, param, latency = metric['flops'], metric['params'], metric['latency']
str1 = '{:14s} FLOP={:6.2f} M, Params={:.3f} MB, latency={:} ms.'.format(
dataset, flop, param,
'{:.2f}'.format(latency *
1000) if latency is not None and latency > 0 else None)
train_info = information.get_metrics(dataset, 'train')
if dataset == 'cifar10-valid':
valid_info = information.get_metrics(dataset, 'x-valid')
test__info = information.get_metrics(dataset, 'ori-test')
str2 = '{:14s} train : [{:}], valid : [{:}], test : [{:}]'.format(
dataset, metric2str(train_info['loss'], train_info['accuracy']),
metric2str(valid_info['loss'], valid_info['accuracy']),
metric2str(test__info['loss'], test__info['accuracy']))
elif dataset == 'cifar10':
test__info = information.get_metrics(dataset, 'ori-test')
str2 = '{:14s} train : [{:}], test : [{:}]'.format(
dataset, metric2str(train_info['loss'], train_info['accuracy']),
metric2str(test__info['loss'], test__info['accuracy']))
else:
valid_info = information.get_metrics(dataset, 'x-valid')
test__info = information.get_metrics(dataset, 'x-test')
str2 = '{:14s} train : [{:}], valid : [{:}], test : [{:}]'.format(
dataset, metric2str(train_info['loss'], train_info['accuracy']),
metric2str(valid_info['loss'], valid_info['accuracy']),
metric2str(test__info['loss'], test__info['accuracy']))
strings += [str1, str2]
if show: print('\n'.join(strings))
return strings
class NATSsize(NASBenchMetaAPI):
"""This is the class for the API of size search space in NATS-Bench."""
def __init__(self,
file_path_or_dict: Optional[Union[Text, Dict[Text, Any]]] = None,
fast_mode: bool = False,
verbose: bool = True):
"""The initialization function that takes the dataset file path (or a dict loaded from that path) as input."""
self._all_base_names = ALL_BASE_NAMES
self.filename = None
self._search_space_name = 'size'
self._fast_mode = fast_mode
self._archive_dir = None
self._full_train_epochs = 90
self.reset_time()
if file_path_or_dict is None:
if self._fast_mode:
self._archive_dir = os.path.join(
get_torch_home(), '{:}-simple'.format(ALL_BASE_NAMES[-1]))
else:
file_path_or_dict = os.path.join(
get_torch_home(), '{:}.{:}'.format(
ALL_BASE_NAMES[-1], PICKLE_EXT))
print('{:} Try to use the default NATS-Bench (size) path from '
'fast_mode={:} and path={:}.'.format(time_string(), self._fast_mode,
file_path_or_dict))
if isinstance(file_path_or_dict, str):
file_path_or_dict = str(file_path_or_dict)
if verbose:
print('{:} Try to create the NATS-Bench (size) api '
'from {:} with fast_mode={:}'.format(
time_string(), file_path_or_dict, fast_mode))
if not nats_is_file(file_path_or_dict) and not nats_is_dir(
file_path_or_dict):
raise ValueError('{:} is neither a file or a dir.'.format(
file_path_or_dict))
self.filename = os.path.basename(file_path_or_dict)
if fast_mode:
if nats_is_file(file_path_or_dict):
raise ValueError('fast_mode={:} must feed the path for directory '
': {:}'.format(fast_mode, file_path_or_dict))
else:
self._archive_dir = file_path_or_dict
else:
if nats_is_dir(file_path_or_dict):
raise ValueError('fast_mode={:} must feed the path for file '
': {:}'.format(fast_mode, file_path_or_dict))
else:
file_path_or_dict = pickle_load(file_path_or_dict)
elif isinstance(file_path_or_dict, dict):
file_path_or_dict = copy.deepcopy(file_path_or_dict)
self.verbose = verbose
if isinstance(file_path_or_dict, dict):
keys = ('meta_archs', 'arch2infos', 'evaluated_indexes')
for key in keys:
if key not in file_path_or_dict:
raise ValueError('Can not find key[{:}] in the dict'.format(key))
self.meta_archs = copy.deepcopy(file_path_or_dict['meta_archs'])
# NOTE(xuanyidong): This is a dict mapping each architecture to a dict,
# where the key is #epochs and the value is ArchResults
self.arch2infos_dict = collections.OrderedDict()
self._avaliable_hps = set()
for xkey in sorted(list(file_path_or_dict['arch2infos'].keys())):
all_infos = file_path_or_dict['arch2infos'][xkey]
hp2archres = collections.OrderedDict()
for hp_key, results in all_infos.items():
hp2archres[hp_key] = ArchResults.create_from_state_dict(results)
self._avaliable_hps.add(hp_key) # save the avaliable hyper-parameter
self.arch2infos_dict[xkey] = hp2archres
self.evaluated_indexes = set(file_path_or_dict['evaluated_indexes'])
elif self.archive_dir is not None:
benchmark_meta = pickle_load('{:}/meta.{:}'.format(
self.archive_dir, PICKLE_EXT))
self.meta_archs = copy.deepcopy(benchmark_meta['meta_archs'])
self.arch2infos_dict = collections.OrderedDict()
self._avaliable_hps = set()
self.evaluated_indexes = set()
else:
raise ValueError('file_path_or_dict [{:}] must be a dict or archive_dir '
'must be set'.format(type(file_path_or_dict)))
self.archstr2index = {}
for idx, arch in enumerate(self.meta_archs):
if arch in self.archstr2index:
raise ValueError('This [{:}]-th arch {:} already in the '
'dict ({:}).'.format(
idx, arch, self.archstr2index[arch]))
self.archstr2index[arch] = idx
if self.verbose:
print('{:} Create NATS-Bench (size) done with {:}/{:} architectures '
'avaliable.'.format(time_string(),
len(self.evaluated_indexes),
len(self.meta_archs)))
def query_info_str_by_arch(self, arch, hp: Text = '12'):
"""Query the information of a specific architecture.
Args:
arch: it can be an architecture index or an architecture string.
hp: the hyperparamete indicator, could be 01, 12, or 90. The difference
between these three configurations are the number of training epochs.
Returns:
ArchResults instance
"""
if self.verbose:
print('{:} Call query_info_str_by_arch with arch={:}'
'and hp={:}'.format(time_string(), arch, hp))
return self._query_info_str_by_arch(arch, hp, print_information)
def get_more_info(self,
index,
dataset,
iepoch=None,
hp: Text = '12',
is_random: bool = True):
"""Return the metric for the `index`-th architecture.
Args:
index: the architecture index.
dataset:
'cifar10-valid' : using the proposed train set of CIFAR-10 as the training set
'cifar10' : using the proposed train+valid set of CIFAR-10 as the training set
'cifar100' : using the proposed train set of CIFAR-100 as the training set
'ImageNet16-120' : using the proposed train set of ImageNet-16-120 as the training set
iepoch: the index of training epochs from 0 to 11/199.
When iepoch=None, it will return the metric for the last training epoch
When iepoch=11, it will return the metric for the 11-th training epoch (starting from 0)
hp: indicates different hyper-parameters for training
When hp=01, it trains the network with 01 epochs and the LR decayed from 0.1 to 0 within 01 epochs
When hp=12, it trains the network with 01 epochs and the LR decayed from 0.1 to 0 within 12 epochs
When hp=90, it trains the network with 01 epochs and the LR decayed from 0.1 to 0 within 90 epochs
is_random:
When is_random=True, the performance of a random architecture will be returned
When is_random=False, the performanceo of all trials will be averaged.
Returns:
a dict, where key is the metric name and value is its value.
"""
if self.verbose:
print('{:} Call the get_more_info function with index={:}, dataset={:}, '
'iepoch={:}, hp={:}, and is_random={:}.'.format(
time_string(), index, dataset, iepoch, hp, is_random))
index = self.query_index_by_arch(index) # To avoid the input is a string or an instance of a arch object
self._prepare_info(index)
if index not in self.arch2infos_dict:
raise ValueError('Did not find {:} from arch2infos_dict.'.format(index))
archresult = self.arch2infos_dict[index][str(hp)]
# if randomly select one trial, select the seed at first
if isinstance(is_random, bool) and is_random:
seeds = archresult.get_dataset_seeds(dataset)
is_random = random.choice(seeds)
# collect the training information
train_info = archresult.get_metrics(
dataset, 'train', iepoch=iepoch, is_random=is_random)
total = train_info['iepoch'] + 1
xinfo = {
'train-loss': train_info['loss'],
'train-accuracy': train_info['accuracy'],
'train-per-time': train_info['all_time'] / total,
'train-all-time': train_info['all_time']
}
# collect the evaluation information
if dataset == 'cifar10-valid':
valid_info = archresult.get_metrics(
dataset, 'x-valid', iepoch=iepoch, is_random=is_random)
try:
test_info = archresult.get_metrics(
dataset, 'ori-test', iepoch=iepoch, is_random=is_random)
except Exception as unused_e: # pylint: disable=broad-except
test_info = None
valtest_info = None
xinfo['comment'] = 'In this dict, train-loss/accuracy/time is the metric on the train set of CIFAR-10. The test-loss/accuracy/time is the performance of the CIFAR-10 test set after training on the train set by {:} epochs. The per-time and total-time indicate the per epoch and total time costs, respectively.'.format(hp)
else:
if dataset == 'cifar10':
xinfo['comment'] = 'In this dict, train-loss/accuracy/time is the metric on the train+valid sets of CIFAR-10. The test-loss/accuracy/time is the performance of the CIFAR-10 test set after training on the train+valid sets by {:} epochs. The per-time and total-time indicate the per epoch and total time costs, respectively.'.format(hp)
try: # collect results on the proposed test set
if dataset == 'cifar10':
test_info = archresult.get_metrics(
dataset, 'ori-test', iepoch=iepoch, is_random=is_random)
else:
test_info = archresult.get_metrics(
dataset, 'x-test', iepoch=iepoch, is_random=is_random)
except Exception as unused_e: # pylint: disable=broad-except
test_info = None
try: # collect results on the proposed validation set
valid_info = archresult.get_metrics(
dataset, 'x-valid', iepoch=iepoch, is_random=is_random)
except Exception as unused_e: # pylint: disable=broad-except
valid_info = None
try:
if dataset != 'cifar10':
valtest_info = archresult.get_metrics(
dataset, 'ori-test', iepoch=iepoch, is_random=is_random)
else:
valtest_info = None
except Exception as unused_e: # pylint: disable=broad-except
valtest_info = None
if valid_info is not None:
xinfo['valid-loss'] = valid_info['loss']
xinfo['valid-accuracy'] = valid_info['accuracy']
xinfo['valid-per-time'] = valid_info['all_time'] / total
xinfo['valid-all-time'] = valid_info['all_time']
if test_info is not None:
xinfo['test-loss'] = test_info['loss']
xinfo['test-accuracy'] = test_info['accuracy']
xinfo['test-per-time'] = test_info['all_time'] / total
xinfo['test-all-time'] = test_info['all_time']
if valtest_info is not None:
xinfo['valtest-loss'] = valtest_info['loss']
xinfo['valtest-accuracy'] = valtest_info['accuracy']
xinfo['valtest-per-time'] = valtest_info['all_time'] / total
xinfo['valtest-all-time'] = valtest_info['all_time']
return xinfo
def show(self, index: int = -1) -> None:
"""Print the information of a specific (or all) architecture(s)."""
self._show(index, print_information)

131
lib/nats_bench/api_test.py
View File

@@ -1,131 +0,0 @@
##############################################################################
# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2020.08 ##########################
##############################################################################
# NATS-Bench: Benchmarking NAS algorithms for Architecture Topology and Size #
##############################################################################
# pytest --capture=tee-sys #
##############################################################################
"""This file is used to quickly test the API."""
import os
import pytest
import random
from nats_bench.api_size import NATSsize
from nats_bench.api_size import ALL_BASE_NAMES as sss_base_names
from nats_bench.api_topology import NATStopology
from nats_bench.api_topology import ALL_BASE_NAMES as tss_base_names
def get_fake_torch_home_dir():
print('This file is {:}'.format(os.path.abspath(__file__)))
print('The current directory is {:}'.format(os.path.abspath(os.getcwd())))
xname = 'FAKE_TORCH_HOME'
if xname in os.environ:
return os.environ['FAKE_TORCH_HOME']
else:
return os.path.join(os.path.dirname(os.path.abspath(__file__)), '..', 'fake_torch_dir')
class TestNATSBench(object):
def test_nats_bench_tss(self, benchmark_dir=None, fake_random=True):
if benchmark_dir is None:
benchmark_dir = os.path.join(get_fake_torch_home_dir(), sss_base_names[-1] + '-simple')
return _test_nats_bench(benchmark_dir, True, fake_random)
def test_nats_bench_sss(self, benchmark_dir=None, fake_random=True):
if benchmark_dir is None:
benchmark_dir = os.path.join(get_fake_torch_home_dir(), tss_base_names[-1] + '-simple')
return _test_nats_bench(benchmark_dir, False, fake_random)
def prepare_fake_tss(self):
print('')
tss_benchmark_dir = os.path.join(get_fake_torch_home_dir(), tss_base_names[-1] + '-simple')
api = NATStopology(tss_benchmark_dir, True, False)
return api
def test_01_th_issue(self):
# Link: https://github.com/D-X-Y/NATS-Bench/issues/1
api = self.prepare_fake_tss()
# The performance of 0-th architecture on CIFAR-10 (trained by 12 epochs)
info = api.get_more_info(0, 'cifar10', hp=12)
# First of all, the data split in NATS-Bench is different from that in the official CIFAR paper.
# In NATS-Bench, we split the original CIFAR-10 training set into two parts, i.e., a training set and a validation set.
# In the following, we will use the splits of NATS-Bench to explain.
print(info['comment'])
print('The loss on the training + validation sets of CIFAR-10: {:}'.format(info['train-loss']))
print('The total training time for 12 epochs on the training + validation sets of CIFAR-10: {:}'.format(info['train-all-time']))
print('The per-epoch training time on CIFAR-10: {:}'.format(info['train-per-time']))
print('The total evaluation time on the test set of CIFAR-10 for 12 times: {:}'.format(info['test-all-time']))
print('The evaluation time on the test set of CIFAR-10: {:}'.format(info['test-per-time']))
cost_info = api.get_cost_info(0, 'cifar10')
xkeys = ['T-train@epoch', # The per epoch training time on the training + validation sets of CIFAR-10.
'T-train@total',
'T-ori-test@epoch', # The time cost for the evaluation on CIFAR-10 test set.
'T-ori-test@total'] # T-ori-test@epoch * 12 times.
for xkey in xkeys:
print('The cost info [{:}] for 0-th architecture on CIFAR-10 is {:}'.format(xkey, cost_info[xkey]))
def test_02_th_issue(self):
# https://github.com/D-X-Y/NATS-Bench/issues/2
api = self.prepare_fake_tss()
data = api.query_by_index(284, dataname='cifar10', hp=200)
for xkey, xvalue in data.items():
print('{:} : {:}'.format(xkey, xvalue))
xinfo = data[777].get_train()
print(xinfo)
print(data[777].train_acc1es)
info_012_epochs = api.get_more_info(284, 'cifar10', hp= 12)
print('Train accuracy for 12 epochs is {:}'.format(info_012_epochs['train-accuracy']))
info_200_epochs = api.get_more_info(284, 'cifar10', hp=200)
print('Train accuracy for 200 epochs is {:}'.format(info_200_epochs['train-accuracy']))
def _test_nats_bench(benchmark_dir, is_tss, fake_random, verbose=False):
"""The main test entry for NATS-Bench."""
if is_tss:
api = NATStopology(benchmark_dir, True, verbose)
else:
api = NATSsize(benchmark_dir, True, verbose)
if fake_random:
test_indexes = [0, 11, 284]
else:
test_indexes = [random.randint(0, len(api) - 1) for _ in range(10)]
key2dataset = {'cifar10': 'CIFAR-10',
'cifar100': 'CIFAR-100',
'ImageNet16-120': 'ImageNet16-120'}
for index in test_indexes:
print('\n\nEvaluate the {:5d}-th architecture.'.format(index))
for key, dataset in key2dataset.items():
# Query the loss / accuracy / time for the `index`-th candidate
# architecture on CIFAR-10
# info is a dict, where you can easily figure out the meaning by key
info = api.get_more_info(index, key)
print(' -->> The performance on {:}: {:}'.format(dataset, info))
# Query the flops, params, latency. info is a dict.
info = api.get_cost_info(index, key)
print(' -->> The cost info on {:}: {:}'.format(dataset, info))
# Simulate the training of the `index`-th candidate:
validation_accuracy, latency, time_cost, current_total_time_cost = api.simulate_train_eval(
index, dataset=key, hp='12')
print(' -->> The validation accuracy={:}, latency={:}, '
'the current time cost={:} s, accumulated time cost={:} s'
.format(validation_accuracy, latency, time_cost,
current_total_time_cost))
# Print the configuration of the `index`-th architecture on CIFAR-10
config = api.get_net_config(index, key)
print(' -->> The configuration on {:} is {:}'.format(dataset, config))
# Show the information of the `index`-th architecture
api.show(index)
with pytest.raises(ValueError):
api.get_more_info(100000, 'cifar10')

338
lib/nats_bench/api_topology.py
View File

@@ -1,338 +0,0 @@
#####################################################
# Copyright (c) Xuanyi Dong [GitHub D-X-Y], 2020.08 #
##############################################################################
# NATS-Bench: Benchmarking NAS algorithms for Architecture Topology and Size #
##############################################################################
# The history of benchmark files are as follows, #
# where the format is (the name is NATS-tss-[version]-[md5].pickle.pbz2) #
# [2020.08.31] NATS-tss-v1_0-3ffb9.pickle.pbz2 #
##############################################################################
# pylint: disable=line-too-long
"""The API for topology search space in NATS-Bench."""
import collections
import copy
import os
import random
from typing import Any, Dict, List, Optional, Text, Union
from nats_bench.api_utils import ArchResults
from nats_bench.api_utils import NASBenchMetaAPI
from nats_bench.api_utils import get_torch_home
from nats_bench.api_utils import nats_is_dir
from nats_bench.api_utils import nats_is_file
from nats_bench.api_utils import PICKLE_EXT
from nats_bench.api_utils import pickle_load
from nats_bench.api_utils import time_string
import numpy as np
ALL_BASE_NAMES = ['NATS-tss-v1_0-3ffb9']
def print_information(information, extra_info=None, show=False):
"""print out the information of a given ArchResults."""
dataset_names = information.get_dataset_names()
strings = [
information.arch_str,
'datasets : {:}, extra-info : {:}'.format(dataset_names, extra_info)
]
def metric2str(loss, acc):
return 'loss = {:.3f} & top1 = {:.2f}%'.format(loss, acc)
for dataset in dataset_names:
metric = information.get_compute_costs(dataset)
flop, param, latency = metric['flops'], metric['params'], metric['latency']
str1 = '{:14s} FLOP={:6.2f} M, Params={:.3f} MB, latency={:} ms.'.format(
dataset, flop, param,
'{:.2f}'.format(latency *
1000) if latency is not None and latency > 0 else None)
train_info = information.get_metrics(dataset, 'train')
if dataset == 'cifar10-valid':
valid_info = information.get_metrics(dataset, 'x-valid')
str2 = '{:14s} train : [{:}], valid : [{:}]'.format(
dataset, metric2str(train_info['loss'], train_info['accuracy']),
metric2str(valid_info['loss'], valid_info['accuracy']))
elif dataset == 'cifar10':
test__info = information.get_metrics(dataset, 'ori-test')
str2 = '{:14s} train : [{:}], test : [{:}]'.format(
dataset, metric2str(train_info['loss'], train_info['accuracy']),
metric2str(test__info['loss'], test__info['accuracy']))
else:
valid_info = information.get_metrics(dataset, 'x-valid')
test__info = information.get_metrics(dataset, 'x-test')
str2 = '{:14s} train : [{:}], valid : [{:}], test : [{:}]'.format(
dataset, metric2str(train_info['loss'], train_info['accuracy']),
metric2str(valid_info['loss'], valid_info['accuracy']),
metric2str(test__info['loss'], test__info['accuracy']))
strings += [str1, str2]
if show: print('\n'.join(strings))
return strings
class NATStopology(NASBenchMetaAPI):
"""This is the class for the API of topology search space in NATS-Bench."""
def __init__(self,
file_path_or_dict: Optional[Union[Text, Dict[Text, Any]]] = None,
fast_mode: bool = False,
verbose: bool = True):
"""The initialization function that takes the dataset file path (or a dict loaded from that path) as input."""
self._all_base_names = ALL_BASE_NAMES
self.filename = None
self._search_space_name = 'topology'
self._fast_mode = fast_mode
self._archive_dir = None
self._full_train_epochs = 200
self.reset_time()
if file_path_or_dict is None:
if self._fast_mode:
self._archive_dir = os.path.join(
get_torch_home(), '{:}-simple'.format(ALL_BASE_NAMES[-1]))
else:
file_path_or_dict = os.path.join(
get_torch_home(), '{:}.{:}'.format(
ALL_BASE_NAMES[-1], PICKLE_EXT))
print('{:} Try to use the default NATS-Bench (topology) path from '
'fast_mode={:} and path={:}.'.format(time_string(), self._fast_mode, file_path_or_dict))
if isinstance(file_path_or_dict, str):
file_path_or_dict = str(file_path_or_dict)
if verbose:
print('{:} Try to create the NATS-Bench (topology) api '
'from {:} with fast_mode={:}'.format(
time_string(), file_path_or_dict, fast_mode))
if not nats_is_file(file_path_or_dict) and not nats_is_dir(
file_path_or_dict):
raise ValueError('{:} is neither a file or a dir.'.format(
file_path_or_dict))
self.filename = os.path.basename(file_path_or_dict)
if fast_mode:
if nats_is_file(file_path_or_dict):
raise ValueError('fast_mode={:} must feed the path for directory '
': {:}'.format(fast_mode, file_path_or_dict))
else:
self._archive_dir = file_path_or_dict
else:
if nats_is_dir(file_path_or_dict):
raise ValueError('fast_mode={:} must feed the path for file '
': {:}'.format(fast_mode, file_path_or_dict))
else:
file_path_or_dict = pickle_load(file_path_or_dict)
elif isinstance(file_path_or_dict, dict):
file_path_or_dict = copy.deepcopy(file_path_or_dict)
self.verbose = verbose
if isinstance(file_path_or_dict, dict):
keys = ('meta_archs', 'arch2infos', 'evaluated_indexes')
for key in keys:
if key not in file_path_or_dict:
raise ValueError('Can not find key[{:}] in the dict'.format(key))
self.meta_archs = copy.deepcopy(file_path_or_dict['meta_archs'])
# NOTE(xuanyidong): This is a dict mapping each architecture to a dict,
# where the key is #epochs and the value is ArchResults
self.arch2infos_dict = collections.OrderedDict()
self._avaliable_hps = set()
for xkey in sorted(list(file_path_or_dict['arch2infos'].keys())):
all_infos = file_path_or_dict['arch2infos'][xkey]
hp2archres = collections.OrderedDict()
for hp_key, results in all_infos.items():
hp2archres[hp_key] = ArchResults.create_from_state_dict(results)
self._avaliable_hps.add(hp_key) # save the avaliable hyper-parameter
self.arch2infos_dict[xkey] = hp2archres
self.evaluated_indexes = set(file_path_or_dict['evaluated_indexes'])
elif self.archive_dir is not None:
benchmark_meta = pickle_load('{:}/meta.{:}'.format(
self.archive_dir, PICKLE_EXT))
self.meta_archs = copy.deepcopy(benchmark_meta['meta_archs'])
self.arch2infos_dict = collections.OrderedDict()
self._avaliable_hps = set()
self.evaluated_indexes = set()
else:
raise ValueError('file_path_or_dict [{:}] must be a dict or archive_dir '
'must be set'.format(type(file_path_or_dict)))
self.archstr2index = {}
for idx, arch in enumerate(self.meta_archs):
if arch in self.archstr2index:
raise ValueError('This [{:}]-th arch {:} already in the '
'dict ({:}).'.format(
idx, arch, self.archstr2index[arch]))
self.archstr2index[arch] = idx
if self.verbose:
print('{:} Create NATS-Bench (topology) done with {:}/{:} architectures '
'avaliable.'.format(time_string(),
len(self.evaluated_indexes),
len(self.meta_archs)))
def query_info_str_by_arch(self, arch, hp: Text = '12'):
"""Query the information of a specific architecture.
Args:
arch: it can be an architecture index or an architecture string.
hp: the hyperparamete indicator, could be 12 or 200. The difference
between these three configurations are the number of training epochs.
Returns:
ArchResults instance
"""
if self.verbose:
print('{:} Call query_info_str_by_arch with arch={:}'
'and hp={:}'.format(time_string(), arch, hp))
return self._query_info_str_by_arch(arch, hp, print_information)
def get_more_info(self,
index,
dataset,
iepoch=None,
hp: Text = '12',
is_random: bool = True):
"""Return the metric for the `index`-th architecture."""
if self.verbose:
print('{:} Call the get_more_info function with index={:}, dataset={:}, '
'iepoch={:}, hp={:}, and is_random={:}.'.format(
time_string(), index, dataset, iepoch, hp, is_random))
index = self.query_index_by_arch(index) # To avoid the input is a string or an instance of a arch object
self._prepare_info(index)
if index not in self.arch2infos_dict:
raise ValueError('Did not find {:} from arch2infos_dict.'.format(index))
archresult = self.arch2infos_dict[index][str(hp)]
# if randomly select one trial, select the seed at first
if isinstance(is_random, bool) and is_random:
seeds = archresult.get_dataset_seeds(dataset)
is_random = random.choice(seeds)
# collect the training information
train_info = archresult.get_metrics(dataset, 'train', iepoch=iepoch, is_random=is_random)
total = train_info['iepoch'] + 1
xinfo = {
'train-loss':
train_info['loss'],
'train-accuracy':
train_info['accuracy'],
'train-per-time':
train_info['all_time'] /
total if train_info['all_time'] is not None else None,
'train-all-time':
train_info['all_time']
}
# collect the evaluation information
if dataset == 'cifar10-valid':
valid_info = archresult.get_metrics(dataset, 'x-valid', iepoch=iepoch, is_random=is_random)
try:
test_info = archresult.get_metrics(dataset, 'ori-test', iepoch=iepoch, is_random=is_random)
except Exception as unused_e: # pylint: disable=broad-except
test_info = None
valtest_info = None
xinfo['comment'] = 'In this dict, train-loss/accuracy/time is the metric on the train set of CIFAR-10. The test-loss/accuracy/time is the performance of the CIFAR-10 test set after training on the train set by {:} epochs. The per-time and total-time indicate the per epoch and total time costs, respectively.'.format(hp)
else:
if dataset == 'cifar10':
xinfo['comment'] = 'In this dict, train-loss/accuracy/time is the metric on the train+valid sets of CIFAR-10. The test-loss/accuracy/time is the performance of the CIFAR-10 test set after training on the train+valid sets by {:} epochs. The per-time and total-time indicate the per epoch and total time costs, respectively.'.format(hp)
try: # collect results on the proposed test set
if dataset == 'cifar10':
test_info = archresult.get_metrics(dataset, 'ori-test', iepoch=iepoch, is_random=is_random)
else:
test_info = archresult.get_metrics(dataset, 'x-test', iepoch=iepoch, is_random=is_random)
except Exception as unused_e: # pylint: disable=broad-except
test_info = None
try: # collect results on the proposed validation set
valid_info = archresult.get_metrics(dataset, 'x-valid', iepoch=iepoch, is_random=is_random)
except Exception as unused_e: # pylint: disable=broad-except
valid_info = None
try:
if dataset != 'cifar10':
valtest_info = archresult.get_metrics(dataset, 'ori-test', iepoch=iepoch, is_random=is_random)
else:
valtest_info = None
except Exception as unused_e: # pylint: disable=broad-except
valtest_info = None
if valid_info is not None:
xinfo['valid-loss'] = valid_info['loss']
xinfo['valid-accuracy'] = valid_info['accuracy']
xinfo['valid-per-time'] = valid_info['all_time'] / total if valid_info['all_time'] is not None else None
xinfo['valid-all-time'] = valid_info['all_time']
if test_info is not None:
xinfo['test-loss'] = test_info['loss']
xinfo['test-accuracy'] = test_info['accuracy']
xinfo['test-per-time'] = test_info['all_time'] / total if test_info['all_time'] is not None else None
xinfo['test-all-time'] = test_info['all_time']
if valtest_info is not None:
xinfo['valtest-loss'] = valtest_info['loss']
xinfo['valtest-accuracy'] = valtest_info['accuracy']
xinfo['valtest-per-time'] = valtest_info['all_time'] / total if valtest_info['all_time'] is not None else None
xinfo['valtest-all-time'] = valtest_info['all_time']
return xinfo
def show(self, index: int = -1) -> None:
"""This function will print the information of a specific (or all) architecture(s)."""
self._show(index, print_information)
@staticmethod
def str2lists(arch_str: Text) -> List[Any]:
"""Shows how to read the string-based architecture encoding.
Args:
arch_str: the input is a string indicates the architecture topology, such as
|nor_conv_1x1~0|+|none~0|none~1|+|none~0|none~1|skip_connect~2|
Returns:
a list of tuple, contains multiple (op, input_node_index) pairs.
[USAGE]
It is the same as the `str2structure` func in AutoDL-Projects:
`github.com/D-X-Y/AutoDL-Projects/lib/models/cell_searchs/genotypes.py`
```
arch = api.str2lists( '|nor_conv_1x1~0|+|none~0|none~1|+|none~0|none~1|skip_connect~2|' )
print ('there are {:} nodes in this arch'.format(len(arch)+1)) # arch is a list
for i, node in enumerate(arch):
print('the {:}-th node is the sum of these {:} nodes with op: {:}'.format(i+1, len(node), node))
```
"""
node_strs = arch_str.split('+')
genotypes = []
for unused_i, node_str in enumerate(node_strs):
inputs = list(filter(lambda x: x != '', node_str.split('|'))) # pylint: disable=g-explicit-bool-comparison
for xinput in inputs:
assert len(
xinput.split('~')) == 2, 'invalid input length : {:}'.format(xinput)
inputs = (xi.split('~') for xi in inputs)
input_infos = tuple((op, int(idx)) for (op, idx) in inputs)
genotypes.append(input_infos)
return genotypes
@staticmethod
def str2matrix(arch_str: Text,
search_space: List[Text] = ('none', 'skip_connect', 'nor_conv_1x1', 'nor_conv_3x3', 'avg_pool_3x3')) -> np.ndarray:
"""Convert the string-based architecture encoding to the encoding strategy in NAS-Bench-101.
Args:
arch_str: the input is a string indicates the architecture topology, such as
|nor_conv_1x1~0|+|none~0|none~1|+|none~0|none~1|skip_connect~2|
search_space: a list of operation string, the default list is the topology search space for NATS-BENCH.
the default value should be be consistent with this line https://github.com/D-X-Y/AutoDL-Projects/blob/main/lib/models/cell_operations.py#L24
Returns:
the numpy matrix (2-D np.ndarray) representing the DAG of this architecture topology
[USAGE]
matrix = api.str2matrix( '|nor_conv_1x1~0|+|none~0|none~1|+|none~0|none~1|skip_connect~2|' )
This matrix is 4-by-4 matrix representing a cell with 4 nodes (only the lower left triangle is useful).
[ [0, 0, 0, 0], # the first line represents the input (0-th) node
[2, 0, 0, 0], # the second line represents the 1-st node, is calculated by 2-th-op( 0-th-node )
[0, 0, 0, 0], # the third line represents the 2-nd node, is calculated by 0-th-op( 0-th-node ) + 0-th-op( 1-th-node )
[0, 0, 1, 0] ] # the fourth line represents the 3-rd node, is calculated by 0-th-op( 0-th-node ) + 0-th-op( 1-th-node ) + 1-th-op( 2-th-node )
In the topology search space in NATS-BENCH, 0-th-op is 'none', 1-th-op is 'skip_connect',
2-th-op is 'nor_conv_1x1', 3-th-op is 'nor_conv_3x3', 4-th-op is 'avg_pool_3x3'.
[NOTE]
If a node has two input-edges from the same node, this function does not work. One edge will be overlapped.
"""
node_strs = arch_str.split('+')
num_nodes = len(node_strs) + 1
matrix = np.zeros((num_nodes, num_nodes))
for i, node_str in enumerate(node_strs):
inputs = list(filter(lambda x: x != '', node_str.split('|'))) # pylint: disable=g-explicit-bool-comparison
for xinput in inputs:
assert len(xinput.split('~')) == 2, 'invalid input length : {:}'.format(xinput)
for xi in inputs:
op, idx = xi.split('~')
if op not in search_space: raise ValueError('this op ({:}) is not in {:}'.format(op, search_space))
op_idx, node_idx = search_space.index(op), int(idx)
matrix[i+1, node_idx] = op_idx
return matrix

1217
lib/nats_bench/api_utils.py
View File

File diff suppressed because it is too large Load Diff

1
lib/spaces/__init__.py Normal file
View File

@@ -0,0 +1 @@
#