def record_results(hyperparam_vect, i_hyper, g):
# print "Meta iter {0}. Recording results".format(i_hyper)
RS = RandomState((seed, i_hyper, "evaluation"))
new_seed = RS.int32()
def loss_fun(alphabets, report_train_loss):
return np.mean([
hyperloss(hyperparam_vect,
new_seed,
alphabets=alphabets,
verbose=False,
report_train_loss=report_train_loss)
for i in range(N_alphabets_eval)
])
cur_hyperparams = hyperparams_0.new_vect(hyperparam_vect.copy())
if i_hyper % N_hyper_thin == 0:
# Storing O(N_weights) is a bit expensive so we thin it out and store in low precision
for field in cur_hyperparams.names:
results[field].append(cur_hyperparams[field].astype(
np.float16))
results['train_loss'].append(
loss_fun(train_alphabets, report_train_loss=True))
results['valid_loss'].append(
loss_fun(train_alphabets, report_train_loss=False))
def record_results(hyperparam_vect, i_hyper, g):
# print "Meta iter {0}. Recording results".format(i_hyper)
RS = RandomState((seed, i_hyper, "evaluation"))
new_seed = RS.int32()
def loss_fun(alphabets, report_train_loss):
return np.mean(
[
hyperloss(
hyperparam_vect,
new_seed,
alphabets=alphabets,
verbose=False,
report_train_loss=report_train_loss,
)
for i in range(N_alphabets_eval)
]
)
cur_hyperparams = hyperparams_0.new_vect(hyperparam_vect.copy())
if i_hyper % N_hyper_thin == 0:
# Storing O(N_weights) is a bit expensive so we thin it out and store in low precision
for field in cur_hyperparams.names:
results[field].append(cur_hyperparams[field].astype(np.float16))
results["train_loss"].append(loss_fun(train_alphabets, report_train_loss=True))
results["valid_loss"].append(loss_fun(train_alphabets, report_train_loss=False))
def loss_fun(alphabets, report_train_loss):
RS = RandomState(
(seed, "evaluation")) # Same alphabet with i_hyper now
return np.mean([
hyperloss(hyperparam_vect,
RS.int32(),
alphabets=alphabets,
verbose=False,
report_train_loss=report_train_loss)
for i in range(N_alphabets_eval)
])
def loss_fun(alphabets, report_train_loss):
RS = RandomState((seed, "evaluation")) # Same alphabet with i_hyper now
return np.mean([hyperloss(hyperparam_vect, RS.int32(), alphabets=alphabets,
verbose=False, report_train_loss=report_train_loss)
for i in range(N_alphabets_eval)])
