def main(task_id, ensemble_dir, performance_range_threshold, ensemble_size,
max_keep_best, seed, only_portfolio_runs, call_from_cmd):
if max_keep_best > 1:
assert max_keep_best == int(max_keep_best)
max_keep_best = int(max_keep_best)
memory_limit = 4000
precision = 32
metric = make_scorer('balanced_accuracy_fast', BalancedAccuracy())
if not os.path.exists(ensemble_dir):
raise NotADirectoryError("%s does not exist")
if call_from_cmd:
assert str(task_id) in ensemble_dir
fl_name = "ensemble_results_%fthresh_%dsize_%fbest" % \
(performance_range_threshold, ensemble_size, max_keep_best)
if only_portfolio_runs:
fl_name += "_only_portfolio"
fl_name = os.path.join(ensemble_dir, fl_name)
if os.path.isfile(fl_name):
raise ValueError("Nothing left to do, %s already exists" % fl_name)
# figure out how many prediction files are in dir
if call_from_cmd:
pred_dir = os.path.join(ensemble_dir, "auto-sklearn-output",
".auto-sklearn", "predictions_ensemble")
n_models = glob.glob(pred_dir +
"/predictions_ensemble_%d_*.npy.gz" % seed)
else:
pred_dir = os.path.join(ensemble_dir, ".auto-sklearn",
"predictions_ensemble")
n_models = glob.glob(pred_dir +
"/predictions_ensemble_%d_*.npy" % seed)
n_models.sort(key=lambda x: int(float(x.split("_")[-2])))
print("\n".join(n_models))
print("Found %d ensemble predictions" % len(n_models))
if len(n_models) == 0:
raise ValueError("%s has no ensemble predictions" % pred_dir)
# Get start time of ensemble building: 1) load json 2) find key 3) get creation times
if call_from_cmd:
timestamps_fl = os.path.join(ensemble_dir, "auto-sklearn-output",
"timestamps.json")
else:
timestamps_fl = os.path.join(ensemble_dir, "timestamps.json")
with open(timestamps_fl, "r") as fh:
timestamps = json.load(fh)
model_timestamps = None
overall_start_time = None
for k in timestamps:
if "predictions_ensemble" in k:
model_timestamps = timestamps[k]
if "start_time_%d" % seed in timestamps[k]:
overall_start_time = timestamps[k]["start_time_%d" % seed]
timestamp_keys = list(model_timestamps.keys())
for timestamp_key in timestamp_keys:
if timestamp_key.endswith(
'lock') or 'predictions_ensemble' not in timestamp_key:
del model_timestamps[timestamp_key]
assert model_timestamps is not None and overall_start_time is not None
assert len(model_timestamps) == len(n_models), (len(model_timestamps),
len(n_models))
# Get overall timelimit
vanilla_results_fl = os.path.join(ensemble_dir, "result.json")
with open(vanilla_results_fl, "r") as fh:
vanilla_results = json.load(fh)
# If only portfolio configurations, read runhistory
if only_portfolio_runs:
if call_from_cmd:
runhistory_fl = os.path.join(ensemble_dir, "auto-sklearn-output",
"smac3-output", "run*",
"runhistory.json")
else:
runhistory_fl = os.path.join(ensemble_dir, "smac3-output", "run*",
"runhistory.json")
runhistory_fl = glob.glob(runhistory_fl)
assert len(runhistory_fl) == 1
with open(runhistory_fl[0], "r") as fh:
runhistory = json.load(fh)
init_design_num_runs = []
for i in runhistory["data"]:
if i[1][3]["configuration_origin"] == "Initial design":
if "error" in i[1][3]:
continue
init_design_num_runs.append(i[1][3]["num_run"])
print("Portfolio stopped after %s runs" % str(init_design_num_runs))
last_run = max(init_design_num_runs)
print("Cut down to only portfolio runs fom %d" % len(n_models))
for i, n in enumerate(n_models):
if int(float(n.split("_")[-2])) > last_run:
n_models = n_models[:i]
break
print("... to %d" % len(n_models))
# load data
X_train, y_train, X_test, y_test, cat = load_task(task_id)
if len(np.unique(y_test)) == 2:
task_type = BINARY_CLASSIFICATION
elif len(np.unique(y_test)) > 2:
task_type = MULTICLASS_CLASSIFICATION
else:
raise ValueError("Unknown task type for task %d" % task_id)
tmp_dir = tempfile.TemporaryDirectory()
loss_trajectory = []
# Construct ensemble builder
context = BackendContextMock(
temporary_directory=(ensemble_dir + "/auto-sklearn-output/"
if call_from_cmd else ensemble_dir),
output_directory=tmp_dir.name,
delete_tmp_folder_after_terminate=False,
delete_output_folder_after_terminate=False,
shared_mode=False)
backend = Backend(context)
ens_builder = EnsembleBuilder(
backend=backend,
dataset_name=str(task_id),
task_type=task_type,
metric=metric,
limit=np.inf,
ensemble_size=ensemble_size,
ensemble_nbest=max_keep_best,
performance_range_threshold=performance_range_threshold,
max_models_on_disc=None,
seed=seed,
shared_mode=False,
precision=precision,
max_iterations=1,
read_at_most=1,
memory_limit=memory_limit,
random_state=1,
sleep_duration=0)
try:
# iterate over all models, take construction time into account when creating new trajectory
current_ensemble_timestamp = 0
skipped = 1
for midx, model_path in enumerate(n_models):
tstamp = model_timestamps[model_path.split("/")[-1].replace(
'.gz', '')] - overall_start_time
if current_ensemble_timestamp > tstamp:
# while this model was built, the ensemble script was not yet done
skipped += 1
continue
# Do one ensemble building step
start = time.time()
ens_builder.random_state = check_random_state(1)
print("############## %d: Working on %s (skipped %d)" %
(midx + 1, model_path, skipped - 1))
logging.basicConfig(level=logging.DEBUG)
ens_builder.read_at_most = skipped
valid_pred, test_pred = ens_builder.main(return_pred=True)
last_dur = time.time() - start
current_ensemble_timestamp = tstamp + last_dur
if current_ensemble_timestamp >= vanilla_results["0"]["time_limit"]:
print("############## Went over time %f > %f; Stop here" %
(current_ensemble_timestamp,
vanilla_results["0"]["time_limit"]))
break
# Reset, since we have just read model files
skipped = 1
if test_pred is None:
# Adding this model did not change the ensemble, no new prediction
continue
if task_type == BINARY_CLASSIFICATION:
# Recreate nx2 array
test_pred = np.concatenate([
1 - test_pred.reshape([-1, 1]),
test_pred.reshape([-1, 1])
],
axis=1)
# Build trajectory entry
score = 1 - balanced_accuracy(y_true=y_test, y_pred=test_pred)
loss_trajectory.append((current_ensemble_timestamp, score))
print("############## Round %d took %g sec" %
(midx, time.time() - start))
except:
raise
finally:
tmp_dir.cleanup()
# Store results
result = dict()
result[ensemble_size] = {
'task_id': task_id,
'time_limit': vanilla_results["0"]["time_limit"],
'loss': loss_trajectory[-1][1],
'configuration': {
"n_models": n_models,
"performance_range_threshold": performance_range_threshold,
"ensemble_size": ensemble_size,
"max_keep_best": max_keep_best,
"seed": seed,
"memory_limit": memory_limit,
"precision": precision,
},
'n_models': len(n_models),
'trajectory': loss_trajectory,
}
with open(fl_name, 'wt') as fh:
json.dump(result, fh, indent=4)
print("Dumped to %s" % fl_name)
import random
import torch
import torch.nn as nn
from tqdm import tqdm
from utils import load_task, make_word_vector, to_var, load_glove_weights, frobenius
from utils import save_pickle, load_pickle
from models import SelfAttentiveNet
data = load_task('./dataset/review.json')
# data = load_pickle('data.pickle')
# save_pickle(data, 'data.pickle')
vocab = set()
for review, _ in data:
vocab |= set(review)
vocab = ['<PAD>'] + list(sorted(vocab))
w2i = dict((w, i) for i, w in enumerate(vocab, 0))
i2w = dict((i, w) for i, w in enumerate(vocab, 0))
print('vocab size', len(vocab))
n_dev = 2000
split_id = len(data) - n_dev
train_data = data[:split_id]
dev_data = data[split_id:]
n_epoch = 4
batch_size = 32
embd_size = 100
attn_hops = 60
taskFilename = sys.argv[1]
vectorsFilename = sys.argv[2]
pathToSVMFile = sys.argv[3]
clusterFile = sys.argv[4]
relFile = sys.argv[5]
pathToExpansionCache = sys.argv[6]
normalVectorsFile = sys.argv[7]
expansion = 5
window = 5
svmFileInfo = '_SVM_' + clusterFile.split('/')[-1] + "_expansionParam" + str(expansion) + "_window" + str(window)
expansionCacheInfo = "_expansionParam_" + str(expansion)
print "Loading rel, task, vector, words that have been disambiguated"
rel = shelve.open(relFile)
task, tralala = load_task(taskFilename)
vectors = load_vectors(vectorsFilename)
normalVectors = load_vectors(normalVectorsFile)
disambiguatedWords = [x.split("_")[0] for x in os.listdir(pathToSVMFile)]
print "Reading agglomerative cluster centers"
clusterCenters = [getAverageWordRep(x, vectors) for x in read_sets(clusterFile)]
print "Starting..."
# initiate empty ratings
methodsRating = []
humanRating = []
questions = task.values()
jointVocCache = dict()
partVoc = set(vectors.keys())
vectorsFilename = sys.argv[2]
pathToSVMFile = sys.argv[3]
clusterFile = sys.argv[4]
relFile = sys.argv[5]
pathToExpansionCache = sys.argv[6]
normalVectorsFile = sys.argv[7]
expansion = 5
window = 5
svmFileInfo = '_SVM_' + clusterFile.split(
'/')[-1] + "_expansionParam" + str(expansion) + "_window" + str(window)
expansionCacheInfo = "_expansionParam_" + str(expansion)
print "Loading rel, task, vector, words that have been disambiguated"
rel = shelve.open(relFile)
task, tralala = load_task(taskFilename)
vectors = load_vectors(vectorsFilename)
normalVectors = load_vectors(normalVectorsFile)
disambiguatedWords = [x.split("_")[0] for x in os.listdir(pathToSVMFile)]
print "Reading agglomerative cluster centers"
clusterCenters = [
getAverageWordRep(x, vectors) for x in read_sets(clusterFile)
]
print "Starting..."
# initiate empty ratings
methodsRating = []
humanRating = []
questions = task.values()
print "Loading stuf..." taskFilename = sys.argv[1] filename = sys.argv[2] # "../../../cluster_descriptors/enwiki8.clust-desc.shelve" vectorsFilename = sys.argv[3] vectors = load_vectors(vectorsFilename) d = shelve.open(filename) key_sets = [] newD = dict() vec_size = len(d.keys()) for i in xrange(vec_size): key_sets.append(set(d[str(i)].keys())) newD[i] = d[str(i)] task, _ = load_task(taskFilename) questions = task.values() methodsRating = [] humanRating = [] print "Answering", len(task), "questions..." for i in xrange(len(questions)): if i % 100 == 0 and not i == 0: print "\tIteration", i, ": ", spearman(methodsRating, humanRating) question = questions[i] word1 = Word(question['word1']).lemma() word2 = Word(question['word2']).lemma()
import sys from utils import load_task, Word from fast_utils import cosine_similarity, load_vectors, spearman if __name__ == "__main__": print "Baseline with wordvectors" if len(sys.argv) < 3: print "USAGE: python baselline_word2vec.py <PATH TO TASK> <PATH TO WORDVECTORS>" sys.exit() taskFilename = sys.argv[1] vectorsFilename = sys.argv[2] task, _ = load_task(taskFilename) vectors = load_vectors(vectorsFilename) methodsRating = [] humanRating = [] questions = task.values() coverage = 0 for i in xrange(len(questions)): question = questions[i] word1 = Word(question['word1']).lemma() word2 = Word(question['word2']).lemma() if word1 in vectors and word2 in vectors: vec1 = vectors[word1] vec2 = vectors[word2] methodsRating.append(cosine_similarity(vec1, vec2))
create_symlink = args.create_symlink
only_check_stats_file = args.only_check_stats_file
max_runtime_limit = args.max_runtime_limit
disable_fallback = args.disable_fallback
with open(selector_file, 'rb') as fh:
selector_dict = pickle.load(fh)
selector = selector_dict['selector']
methods_to_choose_from = selector_dict['methods_to_choose_from']
methods_information = selector_dict['methods_information']
if disable_fallback:
if hasattr(selector, 'default_strategy_idx'):
selector.default_strategy_idx = None
X_train, y_train, _, _, _ = load_task(task_id)
min_num_samples_per_class = np.min(np.unique(y_train, return_counts=True)[1])
meta_features = compute_meta_features(X_train, y_train)
del X_train
del y_train
meta_features = np.array([
meta_features['NumberOfClasses'], meta_features['NumberOfFeatures'],
meta_features['NumberOfInstances']
])
#meta_features = get_meta_features(
# task_id, '/home/eggenspk/PoSHAutosklearn/2020_IEEE_Autosklearn_experiments/experiment_scripts/60MIN/AutoAuto_build_more_metafeatures/metafeatures/')
#meta_features = pd.Series(meta_features)
#for name in ['NumberOfClasses',
# 'NumberOfFeatures',
EVALUATION_INTERVAL = 1
GLOVE_PATH = '/home/crli/crli/glove.840B.300d.txt'
DROPOUT = 0.3
LAYER_NUM = 3
FIXED_EMBEDDING_NUM = 1000
SPAN_LENGTH = 15
KERNEL_SIZES=[2,3]
TRAIN_PATH = './squad/train-v1.1.json'
DEV_PATH = './squad/dev-v1.1.json'
path_to_predictions= './predict/predict_result'
path_to_dev = DEV_PATH
if options.first_load_data:
train_exs,dev_exs = utils.load_task(TRAIN_PATH,DEV_PATH)
with open(options.train_data,'wb') as file:
pickle.dump(train_exs,file)
with open(options.dev_data,'wb') as file:
pickle.dump(dev_exs,file)
else:
with open(options.train_data,'rb') as file:
train_exs = pickle.load(file)
with open(options.dev_data,'rb') as file:
dev_exs = pickle.load(file)
print('The size of train_set:',len(train_exs))
print('The size of dev_set:',len(dev_exs))
TRAIN_SIZE = len(train_exs)
DEV_SIZE = len(dev_exs)
# open the vectors
print "Loading vectors"
vecs = load_vectors(vecFile)
# read clusters and get their cluster centers by taking the average...
print "Reading agglomerative cluster centers"
agglomerativeClusterCenters = [
getAverageWordRep(x, vecs) for x in read_sets(clusterFile)
]
# set some parameters
expansion = 5
window = 5
# get the words that occur in the task and need to be compared
_, wordsToSplit = load_task(pathToTask)
indexCache = dict()
wordsToSplit = filter(lambda x: x not in alreadyDisambiguatedWords,
wordsToSplit)
total = len(wordsToSplit)
for i, word in enumerate(wordsToSplit):
# progess
print "Working on word ", word, i, " / ", total
mySVM, availableSVM, expansionCache = getSVM(
word,
read_file(textfile),
rel,
vecs,
rel = shelve.open(relFile)
# open the vectors
print "Loading vectors"
vecs = load_vectors(vecFile)
# read clusters and get their cluster centers by taking the average...
print "Reading agglomerative cluster centers"
agglomerativeClusterCenters = [getAverageWordRep(x, vecs) for x in read_sets(clusterFile)]
# set some parameters
expansion = 5
window = 5
# get the words that occur in the task and need to be compared
_, wordsToSplit = load_task(pathToTask)
indexCache = dict()
wordsToSplit = filter(lambda x: x not in alreadyDisambiguatedWords, wordsToSplit)
total = len(wordsToSplit)
for i, word in enumerate(wordsToSplit):
# progess
print "Working on word ", word, i, " / ", total
mySVM, availableSVM, expansionCache = getSVM(word, read_file(textfile), rel, vecs, agglomerativeClusterCenters, indexCache, expansionParam=expansion, skipsize=window)
# if we found an svm
if availableSVM:
# dump the svm
pickle.dump(mySVM, open(pathToSVMFile + word + '_SVM_' + clusterFile.split('/')[-1] + "_expansionParam" + str(expansion) + "_window" + str(window), 'w'))
# open expansioncache shelve object
if os.path.exists(tmp_dir):
print('Output directory %s already exists - no need to run this again!' % tmp_dir)
exit(0)
else:
os.makedirs(tmp_dir)
software_stats_file = os.path.join(tmp_dir, 'software.txt')
with open(software_stats_file, 'wt') as fh:
fh.write(strio.read())
argparser_content = {}
for key, value in vars(args).items():
argparser_content[key] = value
argparser_content_file = os.path.join(tmp_dir, 'arguments.json')
with open(argparser_content_file, 'wt') as fh:
json.dump(argparser_content, fh, indent=4)
X_train, y_train, X_test, y_test, cat = load_task(task_id)
iterative_wo_early_stopping = ['extra_trees', 'PassiveAggressiveWOEarlyStopping', 'random_forest',
'SGDWOEarlyStopping', 'GradientBoostingClassifierWOEarlyStopping']
iterative_w_early_stopping = ['extra_trees', 'passive_aggressive', 'random_forest', 'sgd', 'gradient_boosting']
if not early_stopping:
add_classifier_wo_early_stopping()
if searchspace == "iterative":
include_estimator = iterative_w_early_stopping if early_stopping else iterative_wo_early_stopping
include_preprocessor = ["no_preprocessing", ]
elif searchspace == "iterative-preproc":
include_estimator = iterative_w_early_stopping if early_stopping else iterative_wo_early_stopping
include_preprocessor = None
elif searchspace == "full":