overwrite=False,
reuse=False,
shrink_factor=4,
dtype=tf.float32,
model_fn=model_fn,
n_samples_per_ref=5,
lr_fn=lambda global_step: tf.train.exponential_decay(
hyper['initial_lr'], global_step, 100000, hyper['decay_factor']),
hyper=hyper,
)
sigopt_params = [
sigopt_double('initial_lr', 1e-5, 1e-3),
sigopt_double('decay_factor', 1e-3, 0.5),
sigopt_int('num_layers', 10, 20),
]
default_params = {
'initial_lr': 0.000965352400196344,
'decay_factor': 0.0017387361908150767,
'num_layers': 15,
}
def create_train_model(hyper, **kwargs):
model_setup = model_setup_params(hyper)
model_setup.update(kwargs)
return Model(**model_setup)
queue_cap=300,
overwrite=False,
reuse=False,
shrink_factor=4,
dtype=tf.float32,
model_fn=model_fn,
lr_fn=lambda global_step: tf.train.exponential_decay(
hyper['initial_lr'], global_step, 100000, hyper['decay_factor']),
hyper=hyper,
)
sigopt_params = [
sigopt_double('initial_lr', 1e-5, 1e-3),
sigopt_double('decay_factor', 1e-3, 0.5),
sigopt_int('num_layers', 10, 20),
sigopt_int('num_sub_layers', 1, 2),
]
default_params = {
'initial_lr': 0.000965352400196344,
'decay_factor': 0.0017387361908150767,
'num_layers': 20,
'num_sub_layers': 2
}
def create_train_model(hyper, **kwargs):
model_setup = model_setup_params(hyper)
model_setup.update(kwargs)
return Model(**model_setup)