def _get_model_desc(self):
model_desc = self.trainer.model_desc
if not model_desc or 'modules' not in model_desc:
if ModelConfig.model_desc_file is not None:
desc_file = ModelConfig.model_desc_file
desc_file = desc_file.replace("{local_base_path}",
self.trainer.local_base_path)
if ":" not in desc_file:
desc_file = os.path.abspath(desc_file)
if ":" in desc_file:
local_desc_file = FileOps.join_path(
self.trainer.local_output_path,
os.path.basename(desc_file))
FileOps.copy_file(desc_file, local_desc_file)
desc_file = local_desc_file
model_desc = Config(desc_file)
logger.info("net_desc:{}".format(model_desc))
elif ModelConfig.model_desc is not None:
model_desc = ModelConfig.model_desc
elif ModelConfig.models_folder is not None:
folder = ModelConfig.models_folder.replace(
"{local_base_path}", self.trainer.local_base_path)
pattern = FileOps.join_path(folder, "desc_*.json")
desc_file = glob.glob(pattern)[0]
model_desc = Config(desc_file)
else:
return None
return model_desc
def save_results(self):
"""Save the results of evolution contains the information of pupulation and elitism."""
_path = FileOps.join_path(self.local_output_path, General.step_name)
FileOps.make_dir(_path)
arch_file = FileOps.join_path(_path, 'arch.txt')
arch_child = FileOps.join_path(_path, 'arch_child.txt')
sel_arch_file = FileOps.join_path(_path, 'selected_arch.npy')
sel_arch = []
with open(arch_file, 'a') as fw_a, open(arch_child, 'a') as fw_ac:
writer_a = csv.writer(fw_a, lineterminator='\n')
writer_ac = csv.writer(fw_ac, lineterminator='\n')
writer_ac.writerow(
['Population Iteration: ' + str(self.evolution_count + 1)])
for c in range(self.individual_num):
writer_ac.writerow(
self._log_data(net_info_type='active_only',
pop=self.pop[c],
value=self.pop[c].fitness))
writer_a.writerow(
['Population Iteration: ' + str(self.evolution_count + 1)])
for c in range(self.elitism_num):
writer_a.writerow(
self._log_data(net_info_type='active_only',
pop=self.elitism[c],
value=self.elit_fitness[c]))
sel_arch.append(self.elitism[c].gene)
sel_arch = np.stack(sel_arch)
np.save(sel_arch_file, sel_arch)
if self.backup_base_path is not None:
FileOps.copy_folder(self.local_output_path, self.backup_base_path)
def _save_best_model(self):
"""Save best model."""
if zeus.is_torch_backend():
torch.save(self.trainer.model.state_dict(),
self.trainer.weights_file)
elif zeus.is_tf_backend():
worker_path = self.trainer.get_local_worker_path()
model_id = "model_{}".format(self.trainer.worker_id)
weights_folder = FileOps.join_path(worker_path, model_id)
FileOps.make_dir(weights_folder)
checkpoint_file = tf.train.latest_checkpoint(worker_path)
ckpt_globs = glob.glob("{}.*".format(checkpoint_file))
for _file in ckpt_globs:
dst_file = model_id + os.path.splitext(_file)[-1]
FileOps.copy_file(_file,
FileOps.join_path(weights_folder, dst_file))
FileOps.copy_file(FileOps.join_path(worker_path, 'checkpoint'),
weights_folder)
elif zeus.is_ms_backend():
worker_path = self.trainer.get_local_worker_path()
save_path = os.path.join(
worker_path, "model_{}.ckpt".format(self.trainer.worker_id))
for file in os.listdir(worker_path):
if file.startswith("CKP") and file.endswith(".ckpt"):
self.weights_file = FileOps.join_path(worker_path, file)
os.rename(self.weights_file, save_path)
def _save_checkpoint(self, epoch, best=False):
"""Save model weights.
:param epoch: current epoch
:type epoch: int
"""
save_dir = os.path.join(self.worker_path, str(epoch))
FileOps.make_dir(save_dir)
for name in self.model.model_names:
if isinstance(name, str):
save_filename = '%s_net_%s.pth' % (epoch, name)
save_path = FileOps.join_path(save_dir, save_filename)
net = getattr(self.model, 'net' + name)
best_file = FileOps.join_path(self.worker_path,
"model_{}.pth".format(name))
if self.cfg.cuda and torch.cuda.is_available():
# torch.save(net.module.cpu().state_dict(), save_path)
torch.save(net.module.state_dict(), save_path)
# net.cuda()
if best:
torch.save(net.module.state_dict(), best_file)
else:
torch.save(net.cpu().state_dict(), save_path)
if best:
torch.save(net.cpu().state_dict(), best_file)
def _init_next_rung(self):
"""Init next rung to search."""
next_rung_id = self.rung_id + 1
if next_rung_id >= self.total_rungs:
self.rung_id = self.rung_id + 1
return
for i in range(self.config_count):
self.all_config_dict[i][next_rung_id] = self.all_config_dict[i][self.rung_id]
current_score = []
for i in range(self.config_count):
current_score.append((i, self.best_score_dict[self.rung_id][i]))
current_score.sort(key=lambda current_score: current_score[1])
for i in range(4):
better_id = current_score[self.config_count - 1 - i][0]
worse_id = current_score[i][0]
better_worker_result_path = FileOps.join_path(self.local_base_path, 'cache', 'pba',
str(better_id), 'checkpoint')
FileOps.make_dir(better_worker_result_path)
worse_worker_result_path = FileOps.join_path(self.local_base_path, 'cache', 'pba',
str(worse_id), 'checkpoint')
FileOps.make_dir(worse_worker_result_path)
shutil.rmtree(worse_worker_result_path)
shutil.copytree(better_worker_result_path, worse_worker_result_path)
self.all_config_dict[worse_id] = self.all_config_dict[better_id]
policy_unchange = self.all_config_dict[worse_id][next_rung_id]
policy_changed = self.explore(policy_unchange)
self.all_config_dict[worse_id][next_rung_id] = policy_changed
for id in range(self.config_count):
self.best_score_dict[next_rung_id][id] = -1 * float('inf')
tmp_row_data = {'config_id': id,
'rung_id': next_rung_id,
'status': StatusType.WAITTING}
self._add_to_board(tmp_row_data)
self.rung_id = self.rung_id + 1
def _save_worker_record(cls, record):
step_name = record.get('step_name')
worker_id = record.get('worker_id')
_path = TaskOps().get_local_worker_path(step_name, worker_id)
for record_name in ["desc", "performance"]:
_file_name = None
_file = None
record_value = record.get(record_name)
if not record_value:
continue
_file = None
try:
# for cars/darts save multi-desc
if isinstance(record_value, list) and record_name == "desc":
for idx, value in enumerate(record_value):
_file_name = "desc_{}.json".format(idx)
_file = FileOps.join_path(_path, _file_name)
with open(_file, "w") as f:
json.dump(value, f)
else:
_file_name = None
if record_name == "desc":
_file_name = "desc_{}.json".format(worker_id)
if record_name == "performance":
_file_name = "performance_{}.json".format(worker_id)
_file = FileOps.join_path(_path, _file_name)
with open(_file, "w") as f:
json.dump(record_value, f)
except Exception as ex:
logging.error(
"Failed to save {}, file={}, desc={}, msg={}".format(
record_name, _file, record_value, str(ex)))
def load_model(self):
"""Load model."""
if not self.model_desc and not self.weights_file:
saved_folder = self.get_local_worker_path(self.step_name,
self.worker_id)
self.weights_file = FileOps.join_path(
saved_folder, 'model_{}.pth'.format(self.worker_id))
self.model_desc = FileOps.join_path(
saved_folder, 'desc_{}.json'.format(self.worker_id))
if 'modules' not in self.model_desc:
self.model_desc = ModelConfig.model_desc
self.model = ModelZoo.get_model(self.model_desc, self.weights_file)
def dump_model_visual_info(trainer, epoch, model, inputs):
"""Dump model to tensorboard event files.
:param trainer: trainer.
:type worker: object that the class was inherited from DistributedWorker.
:param model: model.
:type model: model.
:param inputs: input data.
:type inputs: data.
"""
(_, visual, interval, title, worker_id, output_path) = _get_trainer_info(trainer)
if visual is not True:
return
if epoch % interval != 0:
return
title = str(worker_id)
_path = FileOps.join_path(output_path, title)
FileOps.make_dir(_path)
try:
with SummaryWriter(_path) as writer:
writer.add_graph(model, (inputs,))
except Exception as e:
logging.error("Failed to dump model visual info, worker id: {}, epoch: {}, error: {}".format(
worker_id, epoch, str(e)
))
def load_model(self):
"""Load model."""
self.saved_folder = self.get_local_worker_path(self.step_name, self.worker_id)
if not self.model_desc:
self.model_desc = FileOps.join_path(self.saved_folder, 'desc_{}.json'.format(self.worker_id))
if not self.weights_file:
if zeus.is_torch_backend():
self.weights_file = FileOps.join_path(self.saved_folder, 'model_{}.pth'.format(self.worker_id))
elif zeus.is_ms_backend():
for file in os.listdir(self.saved_folder):
if file.startswith("CKP") and file.endswith(".ckpt"):
self.weights_file = FileOps.join_path(self.saved_folder, file)
if 'modules' not in self.model_desc:
self.model_desc = ModelConfig.model_desc
self.model = ModelZoo.get_model(self.model_desc, self.weights_file)
def search(self):
"""Search an id and hps from hpo."""
sample = self.hpo.propose()
if sample is None:
return None
re_hps = {}
sample = copy.deepcopy(sample)
sample_id = sample.get('config_id')
trans_para = sample.get('configs')
rung_id = sample.get('rung_id')
all_para = sample.get('all_configs')
re_hps['dataset.transforms'] = [{
'type': 'PBATransformer',
'para_array': trans_para,
'all_para': all_para,
'operation_names': self.operation_names
}]
checkpoint_path = FileOps.join_path(self.local_base_path, 'worker',
'cache', str(sample_id),
'checkpoint')
FileOps.make_dir(checkpoint_path)
if os.path.exists(checkpoint_path):
re_hps['trainer.checkpoint_path'] = checkpoint_path
if 'epoch' in sample:
re_hps['trainer.epochs'] = sample.get('epoch')
return dict(worker_id=sample_id, encoded_desc=re_hps, rung_id=rung_id)
def before_train(self, logs=None):
"""Call before_train of the managed callbacks."""
super().before_train(logs)
"""Be called before the training process."""
hpo_result = FileOps.load_pickle(
FileOps.join_path(self.trainer.local_output_path,
'best_config.pickle'))
logging.info("loading stage1_hpo_result \n{}".format(hpo_result))
feature_interaction_score = hpo_result['feature_interaction_score']
print('feature_interaction_score:', feature_interaction_score)
sorted_pairs = sorted(feature_interaction_score.items(),
key=lambda x: abs(x[1]),
reverse=True)
if ModelConfig.model_desc:
fis_ratio = ModelConfig.model_desc["custom"]["fis_ratio"]
else:
fis_ratio = 1.0
top_k = int(len(feature_interaction_score) * min(1.0, fis_ratio))
self.selected_pairs = list(map(lambda x: x[0], sorted_pairs[:top_k]))
# add selected_pairs
setattr(ModelConfig.model_desc['custom'], 'selected_pairs',
self.selected_pairs)
def _generate_init_model(self):
"""Generate init model by loading pretrained model.
:return: initial model after loading pretrained model
:rtype: torch.nn.Module
"""
model_init = self._new_model_init()
chn_mask = self._init_chn_node_mask()
if vega.is_torch_backend():
checkpoint = torch.load(self.config.init_model_file + '.pth')
model_init.load_state_dict(checkpoint)
model = PruneMobileNet(model_init).apply(chn_mask)
model.to(self.device)
elif vega.is_tf_backend():
model = model_init
with tf.compat.v1.Session(
config=self.trainer._init_session_config()) as sess:
saver = tf.compat.v1.train.import_meta_graph("{}.meta".format(
self.config.init_model_file))
saver.restore(sess, self.config.init_model_file)
all_weight = tf.compat.v1.get_collection(
tf.compat.v1.GraphKeys.VARIABLES)
all_weight = [
t for t in all_weight if not t.name.endswith('Momentum:0')
]
PruneMobileNet(all_weight).apply(chn_mask)
save_file = FileOps.join_path(
self.trainer.get_local_worker_path(), 'prune_model')
saver.save(sess, save_file)
elif vega.is_ms_backend():
parameter_dict = load_checkpoint(self.config.init_model_file)
load_param_into_net(model_init, parameter_dict)
model = PruneMobileNet(model_init).apply(chn_mask)
return model
def after_valid(self, logs=None):
"""Call after_valid of the managed callbacks."""
self.model = self.trainer.model
feature_interaction_score = self.model.get_feature_interaction_score()
print('get feature_interaction_score', feature_interaction_score)
feature_interaction = []
for feature in feature_interaction_score:
if abs(feature_interaction_score[feature]) > 0:
feature_interaction.append(feature)
print('get feature_interaction', feature_interaction)
curr_auc = float(self.trainer.valid_metrics.results['auc'])
if curr_auc > self.best_score:
best_config = {
'score': curr_auc,
'feature_interaction': feature_interaction
}
logging.info("BEST CONFIG IS\n{}".format(best_config))
pickle_result_file = FileOps.join_path(
self.trainer.local_output_path, 'best_config.pickle')
logging.info("Saved to {}".format(pickle_result_file))
FileOps.dump_pickle(best_config, pickle_result_file)
self.best_score = curr_auc
def _load_checkpoint(self):
"""Load checkpoint."""
if zeus.is_torch_backend():
checkpoint_file = FileOps.join_path(
self.trainer.get_local_worker_path(),
self.trainer.checkpoint_file_name)
if os.path.exists(checkpoint_file):
try:
logging.info("Load checkpoint file, file={}".format(
checkpoint_file))
checkpoint = torch.load(checkpoint_file)
self.trainer.model.load_state_dict(checkpoint["weight"])
self.trainer.optimizer.load_state_dict(
checkpoint["optimizer"])
self.trainer.lr_scheduler.load_state_dict(
checkpoint["lr_scheduler"])
if self.trainer._resume_training:
epoch = checkpoint["epoch"]
self.trainer._start_epoch = checkpoint["epoch"]
logging.info(
"Resume fully train, change start epoch to {}".
format(self.trainer._start_epoch))
except Exception as e:
logging.info("Load checkpoint failed {}".format(e))
else:
logging.info('Use default model')
def load_records_from_model_folder(cls, model_folder):
"""Transfer json_file to records."""
if not model_folder or not os.path.exists(model_folder):
logging.error(
"Failed to load records from model folder, folder={}".format(
model_folder))
return []
records = []
pattern = FileOps.join_path(model_folder, "desc_*.json")
files = glob.glob(pattern)
for _file in files:
try:
with open(_file) as f:
worker_id = _file.split(".")[-2].split("_")[-1]
weights_file = os.path.join(
os.path.dirname(_file),
"model_{}.pth".format(worker_id))
if os.path.exists(weights_file):
sample = dict(worker_id=worker_id,
desc=json.load(f),
weights_file=weights_file)
else:
sample = dict(worker_id=worker_id, desc=json.load(f))
record = ReportRecord().load_dict(sample)
records.append(record)
except Exception as ex:
logging.info(
'Can not read records from json because {}'.format(ex))
return records
def _backup(self):
"""Backup result worker folder."""
if self.need_backup is True and self.backup_base_path is not None:
backup_worker_path = FileOps.join_path(
self.backup_base_path, self.get_worker_subpath())
FileOps.copy_folder(
self.get_local_worker_path(self.step_name, self.worker_id), backup_worker_path)
def dump(self):
"""Dump report to file."""
try:
_file = FileOps.join_path(TaskOps().step_path, "reports.csv")
FileOps.make_base_dir(_file)
data = self.all_records
data_dict = {}
for step in data:
step_data = step.serialize().items()
for k, v in step_data:
if k in data_dict:
data_dict[k].append(v)
else:
data_dict[k] = [v]
data = pd.DataFrame(data_dict)
data.to_csv(_file, index=False)
_file = os.path.join(TaskOps().step_path, ".reports")
_dump_data = [
ReportServer._hist_records, ReportServer.__instances__
]
with open(_file, "wb") as f:
pickle.dump(_dump_data, f, protocol=pickle.HIGHEST_PROTOCOL)
self.backup_output_path()
except Exception:
logging.warning(traceback.format_exc())
def _get_current_step_records(self):
step_name = General.step_name
models_folder = PipeStepConfig.pipe_step.get("models_folder")
cur_index = PipelineConfig.steps.index(step_name)
if cur_index >= 1 or models_folder:
# records = Report().get_step_records(PipelineConfig.steps[cur_index - 1])
if not models_folder:
models_folder = FileOps.join_path(
TaskOps().local_output_path,
PipelineConfig.steps[cur_index - 1])
models_folder = models_folder.replace("{local_base_path}",
TaskOps().local_base_path)
records = Report().load_records_from_model_folder(models_folder)
else:
records = self._load_single_model_records()
final_records = []
for record in records:
if not record.weights_file:
logger.error("Model file is not existed, id={}".format(
record.worker_id))
else:
record.step_name = General.step_name
final_records.append(record)
logging.debug("Records: {}".format(final_records))
return final_records
def _copy_needed_file(self):
if self.config.pareto_front_file is None:
raise FileNotFoundError(
"Config item paretor_front_file not found in config file.")
init_pareto_front_file = self.config.pareto_front_file.replace(
"{local_base_path}", self.local_base_path)
self.pareto_front_file = FileOps.join_path(
self.local_output_path, self.step_name, "pareto_front.csv")
FileOps.make_base_dir(self.pareto_front_file)
FileOps.copy_file(init_pareto_front_file, self.pareto_front_file)
if self.config.random_file is None:
raise FileNotFoundError(
"Config item random_file not found in config file.")
init_random_file = self.config.random_file.replace(
"{local_base_path}", self.local_base_path)
self.random_file = FileOps.join_path(
self.local_output_path, self.step_name, "random.csv")
FileOps.copy_file(init_random_file, self.random_file)
def _save_best_model(self):
"""Save best model."""
if zeus.is_torch_backend():
torch.save(self.trainer.model.state_dict(),
self.trainer.weights_file)
elif zeus.is_tf_backend():
worker_path = self.trainer.get_local_worker_path()
model_id = "model_{}".format(self.trainer.worker_id)
weights_folder = FileOps.join_path(worker_path, model_id)
FileOps.make_dir(weights_folder)
checkpoint_file = tf.train.latest_checkpoint(worker_path)
ckpt_globs = glob.glob("{}.*".format(checkpoint_file))
for _file in ckpt_globs:
dst_file = model_id + os.path.splitext(_file)[-1]
FileOps.copy_file(_file,
FileOps.join_path(weights_folder, dst_file))
FileOps.copy_file(FileOps.join_path(worker_path, 'checkpoint'),
weights_folder)
def get_pareto_list_size(self):
"""Get the number of pareto list."""
pareto_list_size = 0
pareto_file_locate = FileOps.join_path(self.local_base_path, "result",
"pareto_front.csv")
if os.path.exists(pareto_file_locate):
pareto_front_df = pd.read_csv(pareto_file_locate)
pareto_list_size = pareto_front_df.size
return pareto_list_size
def __init__(self, **kwargs):
"""Construct the Imagenet class."""
Dataset.__init__(self, **kwargs)
self.args.data_path = FileOps.download_dataset(self.args.data_path)
split = 'train' if self.mode == 'train' else 'val'
local_data_path = FileOps.join_path(self.args.data_path, split)
ImageFolder.__init__(self,
root=local_data_path,
transform=Compose(self.transforms.__transform__))
def before_train(self, logs=None):
"""Be called before the whole train process."""
self.trainer.config.call_metrics_on_train = False
self.cfg = self.trainer.config
self.worker_id = self.trainer.worker_id
self.local_base_path = self.trainer.local_base_path
self.local_output_path = self.trainer.local_output_path
self.result_path = FileOps.join_path(self.trainer.local_base_path, "result")
FileOps.make_dir(self.result_path)
self.logger_patch()
def set_trainer(self, trainer):
"""Set trainer object for current callback."""
self.trainer = trainer
self.trainer._train_loop = self.train_process
self.cfg = self.trainer.config
self._worker_id = self.trainer._worker_id
self.worker_path = self.trainer.get_local_worker_path()
self.output_path = self.trainer.local_output_path
self.best_model_name = "model_best"
self.best_model_file = FileOps.join_path(
self.worker_path, "model_{}.pth".format(self.trainer.worker_id))
def from_json(cls, data, skip_check=True):
"""Restore config from a dictionary or a file."""
t_cls = super(ModelConfig, cls).from_json(data, skip_check)
if data.get("models_folder") and not data.get('model_desc'):
folder = data.models_folder.replace(
"{local_base_path}",
os.path.join(TaskConfig.local_base_path, TaskConfig.task_id))
pattern = FileOps.join_path(folder, "desc_*.json")
desc_file = glob.glob(pattern)[0]
t_cls.model_desc = Config(desc_file)
return t_cls
def logger_patch(self):
"""Patch the default logger."""
worker_path = self.trainer.get_local_worker_path()
worker_spec_log_file = FileOps.join_path(worker_path, 'current_worker.log')
logger = logging.getLogger(__name__)
for hdlr in logger.handlers:
logger.removeHandler(hdlr)
for hdlr in logging.root.handlers:
logging.root.removeHandler(hdlr)
logger.addHandler(logging.FileHandler(worker_spec_log_file))
logger.addHandler(logging.StreamHandler())
logger.setLevel(logging.INFO)
logging.root = logger
def after_train(self, logs=None):
"""Call after_train of the managed callbacks."""
curr_auc = float(self.trainer.valid_metrics.results['auc'])
self.sieve_board = self.sieve_board.append(
{
'selected_feature_pairs': self.selected_pairs,
'score': curr_auc
}, ignore_index=True)
result_file = FileOps.join_path(
self.trainer.local_output_path, '{}_result.csv'.format(self.trainer.__worker_id__))
self.sieve_board.to_csv(result_file, sep='\t')
def _get_search_space_list(self):
"""Get search space list from models folder."""
models_folder = PipeStepConfig.pipe_step.get("models_folder")
if not models_folder:
self.search_space_list = None
return
self.search_space_list = []
models_folder = models_folder.replace("{local_base_path}", TaskOps().local_base_path)
pattern = FileOps.join_path(models_folder, "*.json")
files = glob.glob(pattern)
for file in files:
with open(file) as f:
self.search_space_list.append(json.load(f))
def _output_records(self,
step_name,
records,
desc=True,
weights_file=False,
performance=False):
"""Dump records."""
columns = ["worker_id", "performance", "desc"]
outputs = []
for record in records:
record = record.serialize()
_record = {}
for key in columns:
_record[key] = record[key]
outputs.append(deepcopy(_record))
data = pd.DataFrame(outputs)
step_path = FileOps.join_path(TaskOps().local_output_path, step_name)
FileOps.make_dir(step_path)
_file = FileOps.join_path(step_path, "output.csv")
try:
data.to_csv(_file, index=False)
except Exception:
logging.error("Failed to save output file, file={}".format(_file))
for record in outputs:
worker_id = record["worker_id"]
worker_path = TaskOps().get_local_worker_path(step_name, worker_id)
outputs_globs = []
if desc:
outputs_globs += glob.glob(
FileOps.join_path(worker_path, "desc_*.json"))
if weights_file:
outputs_globs += glob.glob(
FileOps.join_path(worker_path, "model_*.pth"))
if performance:
outputs_globs += glob.glob(
FileOps.join_path(worker_path, "performance_*.json"))
for _file in outputs_globs:
FileOps.copy_file(_file, step_path)
def before_train(self, logs=None):
"""Call before_train of the managed callbacks."""
super().before_train(logs)
"""Be called before the training process."""
hpo_result = FileOps.load_pickle(FileOps.join_path(
self.trainer.local_output_path, 'best_config.pickle'))
logging.info("loading stage1_hpo_result \n{}".format(hpo_result))
self.selected_pairs = hpo_result['feature_interaction']
logging.info('feature_interaction:', self.selected_pairs)
# add selected_pairs
setattr(ModelConfig.model_desc['custom'], 'selected_pairs', self.selected_pairs)