def parse_yaml(yaml_path, model_id):
from tensorflow.contrib.training import HParams
from ruamel.yaml import YAML
hparams = HParams()
hparams.add_hparam('model_id', model_id)
with open(yaml_path) as fp:
customized = YAML().load(fp)
for k, v in customized.items():
if k in hparams:
hparams.set_hparam(k, v)
else:
hparams.add_hparam(k, v)
return hparams
def _load_hparams(path):
with open(os.path.join(path, 'hparams.json'), 'rb') as json_file:
hparams_dict = {
k.encode('utf-8'): v.encode('utf-8') if type(v) == unicode else v
for k, v in json.load(json_file).iteritems()
}
hparams = HParams(**hparams_dict)
hparams.set_hparam('data_dir', path)
trainer_lib.add_problem_hparams(hparams, 'translate_mmt')
# Removing dropout from HParams even on TRAIN mode
for key in hparams.values():
if key.endswith("dropout"):
setattr(hparams, key, 0.0)
return hparams
def build_network(self, build_encoder=True):
if build_encoder:
self.encoder = Encoder(self._hp)
self.decoder = DecoderModule(self._hp, # infer actions in decoder if not using SH-Pred model
regress_actions=self._hp.regress_actions and self._hp.one_step_planner is not 'sh_pred')
self.build_inference()
if self._hp.regress_length:
self.length_pred = LengthPredictorModule(self._hp)
self.build_inference_encoder()
if self._hp.attach_inv_mdl:
self.inv_mdl = InverseModel(self._hp.inv_mdl_params, self._logger)
if self._hp.attach_cost_mdl:
self.cost_mdl = CostModel(self._hp.cost_mdl_params, self._logger)
if self._hp.attach_state_regressor:
self.state_regressor = BaseProcessingNet(self._hp.nz_enc, self._hp.nz_mid, self._hp.state_dim,
self._hp.n_processing_layers, self._hp.fc_builder)
if self._hp.separate_cnn_start_goal_encoder:
from blox.torch.layers import LayerBuilderParams
from tensorflow.contrib.training import HParams
with_conv_hp = HParams()
for k in self._hp.values().keys():
with_conv_hp.add_hparam(k, self._hp.values()[k])
#with_conv_hp = self._hp
with_conv_hp.set_hparam('use_convs', True)
with_conv_hp.set_hparam('input_nc', 3)
with_conv_hp.set_hparam('builder', LayerBuilderParams(
True, self._hp.use_batchnorm, self._hp.normalization, self._hp.predictor_normalization))
self.start_goal_enc = Encoder(with_conv_hp)
# update tensorflow's hparams object using values from
# python's ArgumentParser.
parser = argparse.ArgumentParser()
params, pargs = update_params_from_parser(params, parser)
logger.info('Load echonest dataset')
dataset = fetch_echonest(data_home='data/echonest',
min_playcount=params.min_playcount,
min_interactions=params.min_interactions)
n_users = dataset['interactions'].shape[0]
n_items = dataset['interactions'].shape[1]
# update n_users, n_items parameters
if hasattr(params, 'n_users'):
params.set_hparam('n_users', n_users)
else:
params.add_hparam('n_users', n_users)
if hasattr(params, 'n_items'):
params.set_hparam('n_items', n_items)
else:
params.add_hparam('n_items', n_items)
model_path = os.path.join(
'echonest_pltcnt{}_minint{}'.format(params.min_playcount,
params.min_interactions),
'batch_{}'.format(params.batch_size),
'lr{}_dim{}_nepochs{}_sampler-{}_batchsize{}_negs{}'.format(
params.learning_rate, params.embedding_dim, params.n_epochs,
params.sampler, params.batch_size, params.n_negatives))
def parse_args(yaml_path, model_id, default_set, followup=None):
logger = logging.getLogger(APP_NAME)
hparams = HParams()
hparams.add_hparam('model_id', model_id)
with open('default.yaml') as fp:
configs = YAML().load(fp)
default_cfg = configs[default_set]
add_param_recur(hparams, default_cfg)
if yaml_path:
logger.info('loading parameters...')
with open(yaml_path) as fp:
customized = YAML().load(fp)
for k, v in customized.items():
if k in hparams and hparams.get(k) != v:
logger.info('%20s: %20s -> %20s' %
(k, hparams.get(k), v))
hparams.set_hparam(k, v)
elif k not in hparams: # add new parameter
hparams.add_hparam(k, v)
logger.info(
'%30s %20s: %20s' %
("[add from %s]" % yaml_path, k, hparams.get(k)))
if followup:
# useful when changing args for prediction
logger.info('override args with follow-up args...')
for k, v in followup.items():
if k in hparams and hparams.get(k) != v:
logger.info('%20s: %20s -> %20s' % (k, hparams.get(k), v))
hparams.set_hparam(k, v)
elif k not in hparams:
logger.warning('%s is not a valid attribute! ignore!' % k)
if 'save_dir' not in hparams:
hparams.add_hparam(
'save_dir',
os.path.join(hparams.get('model_dir'), hparams.get('model_id')))
if 'code_dir' not in hparams:
hparams.add_hparam('code_dir',
os.path.join(hparams.get('save_dir'), 'code'))
hparams.set_hparam('summary_dir',
os.path.join(hparams.get('save_dir'), 'summary'))
# reset logger model id
logger = set_logger(model_id='%s:%s' %
(DEVICE_ID, hparams.get('model_id')))
try:
shutil.copytree('./',
hparams.get('code_dir'),
ignore=shutil.ignore_patterns(*IGNORE_PATTERNS))
logger.info('current code base is copied to %s' %
hparams.get('save_dir'))
except FileExistsError:
logger.info('code base exist, no need to copy!')
# if hparams.get('model_id') != model_id:
# logger.warning('model id is changed %s -> %s! '
# 'This happens when you train a pretrained model' % (
# hparams.get('model_id'), model_id))
# hparams.set_hparam('model_id', model_id)
if 'loss_csv_file' not in hparams:
hparams.add_hparam('loss_csv_file',
os.path.join(hparams.get('save_dir'), 'loss.csv'))
if 'is_serving' not in hparams:
hparams.add_hparam('is_serving', False)
logger.info('current parameters')
for k, v in sorted(vars(hparams).items()):
if not k.startswith('_'):
logger.info('%20s = %-20s' % (k, v))
return hparams
src_max_len=200,
tgt_max_len=200,
infer_src_max_len=200,
infer_tgt_max_len=200,
optimizer='adam',
learning_rate=0.0001,
final_learning_rate=0.00002,
warmup_steps=4000,
bucket=0,
use_bpe=False,
attention_mechanism='multi_head',
beam_width=1,
encoder_type='bi',
unit_type='gru',
dropout=0.2,
num_residual_layers=0,
sos='<s>',
eos='</s>',
pad='<pad>',
unk='<unk>',
)
if params_scale == 'medium':
hparams.set_hparam('batch_size', 32)
hparams.set_hparam('bpe_num_symbols', 24000)
hparams.set_hparam('num_units', 256)
elif params_scale == 'large':
hparams.set_hparam('batch_size', 64)
hparams.set_hparam('bpe_num_symbols', 32000)
hparams.set_hparam('num_units', 512)