def on_epoch_end(self, epoch, logs=None):
#get current validation mae
if cf.dataset == 'PPG_Dalia':
current = logs.get("val_mean_absolute_error")
elif cf.dataset == 'copy_memory':
current = logs.get("val_loss")
elif cf.dataset == 'poly_music':
current = logs.get("val_loss")
#compare with previous best one
if np.less(current, self.best):
self.best = current
# Record the best model if current results is better (less).
names = [
weight.name for layer in model.layers
for weight in layer.weights
]
weights = model.get_weights()
for name, weight in zip(names, weights):
if re.search('learned_conv2d.+_?[0-9]/gamma', name):
self.gamma[name] = weight
self.gamma[name] = np.array(
self.gamma[name] > cf.threshold, dtype=bool)
self.gamma[name] = utils.dil_fact(self.gamma[name],
op='mul')
print("New best MAE, update file. \n")
print(self.gamma)
utils.save_dil_fact(cf.saving_path, self.gamma)
def train_gammas_Char_PTB(model, epochs, n_words, X_train, X_valid, X_test):
train_losses = list()
valid_losses = list()
test_losses = list()
best_v_l = np.Inf
patience = 30
wait = 0
gamma = dict()
for ep in range(epochs):
tr_l, v_l, t_l = train_Char_PTB(model, ep, n_words, X_train, X_valid,
X_test, train_losses, valid_losses,
test_losses)
names = [
weight.name for layer in model.layers for weight in layer.weights
]
weights = model.get_weights()
for name, weight in zip(names, weights):
if re.search('learned_conv2d.+_?[0-9]/gamma', name):
gamma[name] = weight
print(gamma)
# early_stop + export_structure
if np.less(v_l, best_v_l):
if v_l >= 2:
best_v_l = v_l
wait = 0
# Record the best model if current results is better (less).
names = [
weight.name for layer in model.layers
for weight in layer.weights
]
weights = model.get_weights()
for name, weight in zip(names, weights):
if re.search('learned_conv2d.+_?[0-9]/gamma', name):
#gamma[name] = weight
gamma[name] = np.array(gamma[name] > cf.threshold,
dtype=bool)
gamma[name] = utils.dil_fact(gamma[name], op='mul')
print("New best MAE, update file. \n")
print(gamma)
utils.save_dil_fact(cf.saving_path, gamma)
else:
wait += 1
print("Val loss did not improve from {}".format(best_v_l))
print("Iter for at least {} epochs".format(patience - wait))
if wait >= patience:
print("Early Stop")
break
ep += 1
def on_epoch_end(self, epoch, logs=None):
#get current validation mae
if cf.dataset == 'PPG_Dalia':
current = logs.get(val_mae)
l = 1
h = 0
wait = 0
elif cf.dataset == 'Nottingham' or cf.dataset == 'JSB_Chorales':
current = logs.get("val_loss")
l = 1
h = 0
elif cf.dataset == 'SeqMNIST' or cf.dataset == 'PerMNIST':
current = logs.get("val_accuracy")
l = 0
h = 1
wait = 20
else:
print("{} is not supported".format(cf.dataset))
sys.exit()
if self.i > wait:
# compare with previous best one
if bool(np.less(current, self.best) * l) ^ \
bool((current > self.best) * h):
self.best = current
# Record the best model if current results is better.
names = [
weight.name for layer in self.model.layers
for weight in layer.weights
]
weights = self.model.get_weights()
for name, weight in zip(names, weights):
if re.search('learned_conv2d.+_?[0-9]/gamma', name):
self.gamma[name] = weight
self.gamma[name] = np.array(
self.gamma[name] > cf.threshold, dtype=bool)
self.gamma[name] = utils.dil_fact(self.gamma[name],
op='mul')
elif re.search('weight_norm.+_?[0-9]/gamma', name):
self.gamma[name] = weight
self.gamma[name] = np.array(
self.gamma[name] > cf.threshold, dtype=bool)
self.gamma[name] = utils.dil_fact(self.gamma[name],
op='mul')
print("New best model, update file. \n")
print(self.gamma)
utils.save_dil_fact(cf.saving_path, self.gamma)
else:
self.i += 1
# Record the best model if current results is better (less).
names = [
weight.name for layer in model.layers
for weight in layer.weights
]
weights = model.get_weights()
for name, weight in zip(names, weights):
if re.search('learned_conv2d.+_?[0-9]/gamma', name):
gamma[name] = weight
gamma[name] = np.array(gamma[name] > cf.threshold,
dtype=bool)
gamma[name] = utils.dil_fact(gamma[name], op='mul')
print("New best MAE, update file. \n")
print(gamma)
utils.save_dil_fact(cf.saving_path, gamma)
else:
wait += 1
print("Val loss did not improve from {}".format(best_v_l))
print("Iter for at least {} epochs".format(patience - wait))
if wait >= patience:
print("Early Stop")
break
ep += 1
model.save_weights(cf.saving_path + 'autodil/test2_trained_weights_warmup' +
str(cf.warmup) + '.h5')