def get_nasbench201_api(dataset=None):
"""
Load the original nasbench201 dataset (which does not include full LC info)
TODO: this is a subset of the full LC datasets, so it is possible to get rid of this dataset.
"""
with open(os.path.join(get_project_root(), 'data', 'nb201_all.pickle'),
'rb') as f:
nb201_data = pickle.load(f)
"""
Now load the full LC info. These files are large, so we only load one for the specific dataset.
"""
if dataset == 'cifar10':
with open(
os.path.join(get_project_root(), 'data',
'nb201_cifar10_full_training.pickle'), 'rb') as f:
full_lc_data = pickle.load(f)
elif dataset == 'cifar100':
with open(
os.path.join(get_project_root(), 'data',
'nb201_cifar100_full_training.pickle'),
'rb') as f:
full_lc_data = pickle.load(f)
elif dataset == 'ImageNet16-120':
with open(
os.path.join(get_project_root(), 'data',
'nb201_ImageNet16_full_training.pickle'),
'rb') as f:
full_lc_data = pickle.load(f)
return {'raw_data': nb201_data, 'full_lc_data': full_lc_data}
def get_nlp_api(dataset=None):
"""
Load the NAS-Bench-NLP data
"""
with open(os.path.join(get_project_root(), 'data', 'nb_nlp.pickle'), 'rb') as f:
nlp_data = pickle.load(f)
nlp_arches = list(nlp_data.keys())
return {'nlp_data':nlp_data, 'nlp_arches':nlp_arches}
def get_nasbench201_api(dataset=None):
"""
Load the NAS-Bench-201 data
"""
if dataset == 'cifar10':
with open(os.path.join(get_project_root(), 'data', 'nb201_cifar10_full_training.pickle'), 'rb') as f:
data = pickle.load(f)
elif dataset == 'cifar100':
with open(os.path.join(get_project_root(), 'data', 'nb201_cifar100_full_training.pickle'), 'rb') as f:
data = pickle.load(f)
elif dataset == 'ImageNet16-120':
with open(os.path.join(get_project_root(), 'data', 'nb201_ImageNet16_full_training.pickle'), 'rb') as f:
data = pickle.load(f)
return {'nb201_data':data}
def get_darts_api(dataset=None):
"""
Load the nb301 training data (which contains full learning curves) and the nb301 models
"""
import nasbench301
data_folder = os.path.join(get_project_root(), 'data/')
with open(os.path.join(data_folder, 'nb301_full_training.pickle'), 'rb') as f:
nb301_data = pickle.load(f)
nb301_arches = list(nb301_data.keys())
performance_model = nasbench301.load_ensemble(os.path.join(data_folder + 'nb301_models/xgb_v1.0'))
runtime_model = nasbench301.load_ensemble(os.path.join(data_folder + 'nb301_models/lgb_runtime_v1.0'))
nb301_model = [performance_model, runtime_model]
return {'nb301_data': nb301_data, 'nb301_arches':nb301_arches, 'nb301_model':nb301_model}
def get_configspace(path_to_configspace_obj=os.path.join(
get_project_root(), "search_spaces/darts/configspace.json")):
"""
Returns the ConfigSpace object for the search space
Args:
path_to_configspace_obj: path to ConfigSpace json encoding
Returns:
ConfigSpace.ConfigutationSpace: a ConfigSpace object
"""
with open(path_to_configspace_obj, 'r') as fh:
json_string = fh.read()
config_space = config_space_json_r_w.read(json_string)
return config_space
def __init__(self, config, batch_size=64, method_type='jacov'):
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
self.batch_size = batch_size
self.method_type = method_type
self.device = torch.device(
"cuda:0" if torch.cuda.is_available() else "cpu")
config.data = "{}/data".format(get_project_root())
self.config = config
if method_type == 'jacov':
self.num_classes = 1
else:
num_classes_dic = {
'cifar10': 10,
'cifar100': 100,
'ImageNet16-120': 120
}
self.num_classes = num_classes_dic[self.config.dataset]
def __init__(self, config, batch_size=64, method_type='jacov'):
# available zero-cost method types: 'jacov', 'snip', 'synflow', 'grad_norm', 'fisher', 'grasp'
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
self.batch_size = batch_size
self.dataload = 'random'
self.num_imgs_or_batches = 1
self.method_type = method_type
self.device = torch.device(
"cuda:0" if torch.cuda.is_available() else "cpu")
config.data = "{}/data".format(get_project_root())
self.config = config
num_classes_dic = {
'cifar10': 10,
'cifar100': 100,
'ImageNet16-120': 120
}
self.num_classes = None
if self.config.dataset in num_classes_dic:
self.num_classes = num_classes_dic[self.config.dataset]
import unittest
import logging
import torch
import os
from naslib.search_spaces import HierarchicalSearchSpace
from naslib.optimizers import DARTSOptimizer, GDASOptimizer
from naslib.utils import utils, setup_logger
logger = setup_logger(
os.path.join(utils.get_project_root().parent, "tmp", "tests.log"))
logger.handlers[0].setLevel(logging.FATAL)
config = utils.AttrDict()
config.dataset = 'cifar10'
config.search = utils.AttrDict()
config.search.grad_clip = None
config.search.learning_rate = 0.01
config.search.momentum = 0.1
config.search.weight_decay = 0.1
config.search.arch_learning_rate = 0.01
config.search.arch_weight_decay = 0.1
config.search.tau_max = 10
config.search.tau_min = 1
config.search.epochs = 2
data_train = (torch.ones([2, 3, 32, 32]), torch.ones([2]).long())
data_val = (torch.ones([2, 3, 32, 32]), torch.ones([2]).long())
if torch.cuda.is_available():
data_train = tuple(x.cuda() for x in data_train)
def get_nasbench101_api(dataset=None):
# load nasbench101
from nasbench import api
nb101_data = api.NASBench(
os.path.join(get_project_root(), 'data', 'nasbench_only108.tfrecord'))
return {'api': api, 'nb101_data': nb101_data}
import os
import pickle
import torch.nn as nn
from naslib.search_spaces.core import primitives as ops
from naslib.search_spaces.core.graph import Graph, EdgeData
from naslib.search_spaces.core.primitives import AbstractPrimitive
from naslib.search_spaces.core.query_metrics import Metric
from naslib.utils.utils import get_project_root
from .primitives import ResNetBasicblock
# load the nasbench201 data
with open(os.path.join(get_project_root(), 'data', 'nb201_all.pickle'), 'rb') as f:
nb201_data = pickle.load(f)
class NasBench201SearchSpace(Graph):
"""
Implementation of the nasbench 201 search space.
It also has an interface to the tabular benchmark of nasbench 201.
"""
OPTIMIZER_SCOPE = [
"stage_1",
"stage_2",
"stage_3",
]
QUERYABLE = True
import numpy as np
from naslib.search_spaces.core import primitives as ops
from naslib.search_spaces.core.graph import Graph, EdgeData
from naslib.search_spaces.core.primitives import AbstractPrimitive
from naslib.search_spaces.core.query_metrics import Metric
from naslib.utils.utils import get_project_root
from .primitives import ReLUConvBN
# load the nasbench101 data -- requires TF 1.x
from nasbench import api
nb101_datadir = os.path.join(get_project_root(), 'data', 'nasbench_only108.tfrecord')
nasbench = api.NASBench(nb101_datadir)
# data = nasbench.query(cell)
class NasBench101SearchSpace(Graph):
"""
Contains the interface to the tabular benchmark of nasbench 101.
"""
OPTIMIZER_SCOPE = [
"stack_1",
"stack_2",
"stack_3",
]
