mirror of
https://github.com/iperov/DeepFaceLab.git
synced 2025-07-06 13:02:15 -07:00
revert back Adam
This commit is contained in:
iperov
parent
e4637336ef
commit
ee8dbcbc35
3 changed files with 56 additions and 66 deletions
|
|
@ -121,7 +121,8 @@ class ModelBase(object):
|
|||
nnlib.import_all ( nnlib.DeviceConfig(allow_growth=False, **self.device_args) )
|
||||
self.device_config = nnlib.active_DeviceConfig
|
||||
self.keras = nnlib.keras
|
||||
|
||||
self.K = nnlib.keras.backend
|
||||
|
||||
self.onInitialize()
|
||||
|
||||
self.options['batch_size'] = self.batch_size
|
||||
|
|
@ -282,17 +283,20 @@ class ModelBase(object):
|
|||
if len(optimizer_filename_list) != 0:
|
||||
opt_filename = self.get_strpath_storage_for_file('opt.h5')
|
||||
if Path(opt_filename).exists():
|
||||
h5dict = self.keras.utils.io_utils.H5Dict(opt_filename, mode='r')
|
||||
try:
|
||||
with open(opt_filename, "rb") as f:
|
||||
d = pickle.loads(f.read())
|
||||
|
||||
for x in optimizer_filename_list:
|
||||
opt, filename = x
|
||||
if filename in h5dict:
|
||||
opt = opt.__class__.from_config( json.loads(h5dict[filename]) )
|
||||
x[0] = opt
|
||||
finally:
|
||||
h5dict.close()
|
||||
|
||||
return [x[0] for x in optimizer_filename_list]
|
||||
if filename in d:
|
||||
weights = d[filename].get('weights', None)
|
||||
if weights:
|
||||
opt.set_weights(weights)
|
||||
print("set ok")
|
||||
except Exception as e:
|
||||
print ("Unable to load ", opt_filename)
|
||||
|
||||
|
||||
def save_weights_safe(self, model_filename_list, optimizer_filename_list=[]):
|
||||
for model, filename in model_filename_list:
|
||||
|
|
@ -302,13 +306,23 @@ class ModelBase(object):
|
|||
rename_list = model_filename_list
|
||||
if len(optimizer_filename_list) != 0:
|
||||
opt_filename = self.get_strpath_storage_for_file('opt.h5')
|
||||
h5dict = self.keras.utils.io_utils.H5Dict(opt_filename + '.tmp', mode='w')
|
||||
|
||||
try:
|
||||
d = {}
|
||||
for opt, filename in optimizer_filename_list:
|
||||
h5dict[filename] = json.dumps(opt.get_config())
|
||||
finally:
|
||||
h5dict.close()
|
||||
fd = {}
|
||||
symbolic_weights = getattr(opt, 'weights')
|
||||
if symbolic_weights:
|
||||
fd['weights'] = self.K.batch_get_value(symbolic_weights)
|
||||
|
||||
d[filename] = fd
|
||||
|
||||
with open(opt_filename+'.tmp', 'wb') as f:
|
||||
f.write( pickle.dumps(d) )
|
||||
|
||||
rename_list += [('', 'opt.h5')]
|
||||
except Exception as e:
|
||||
print ("Unable to save ", opt_filename)
|
||||
|
||||
for _, filename in rename_list:
|
||||
filename = self.get_strpath_storage_for_file(filename)
|
||||
|
|
|
|||
|
|
@ -223,13 +223,6 @@ class SAEModel(ModelBase):
|
|||
pred_dst_dstm = self.decoder_dstm(warped_dst_code)
|
||||
pred_src_dstm = self.decoder_srcm(warped_dst_code)
|
||||
|
||||
self.src_dst_opt, \
|
||||
self.src_dst_mask_opt = self.load_weights_safe(
|
||||
weights_to_load,
|
||||
[ [Adam(lr=5e-5, beta_1=0.5, beta_2=0.999), 'src_dst_opt'],
|
||||
[Adam(lr=5e-5, beta_1=0.5, beta_2=0.999), 'src_dst_mask_opt']
|
||||
])
|
||||
|
||||
pred_src_src, pred_dst_dst, pred_src_dst, = [ [x] if type(x) != list else x for x in [pred_src_src, pred_dst_dst, pred_src_dst, ] ]
|
||||
|
||||
if self.options['learn_mask']:
|
||||
|
|
@ -267,6 +260,9 @@ class SAEModel(ModelBase):
|
|||
psd_target_dst_anti_masked_ar = [ pred_src_dst_sigm_ar[i]*target_dstm_anti_sigm_ar[i] for i in range(len(pred_src_dst_sigm_ar))]
|
||||
|
||||
if self.is_training_mode:
|
||||
self.src_dst_opt = Adam(lr=5e-5, beta_1=0.5, beta_2=0.999)
|
||||
self.src_dst_mask_opt = Adam(lr=5e-5, beta_1=0.5, beta_2=0.999)
|
||||
|
||||
if self.options['archi'] == 'liae':
|
||||
src_dst_loss_train_weights = self.encoder.trainable_weights + self.inter_B.trainable_weights + self.inter_AB.trainable_weights + self.decoder.trainable_weights
|
||||
if self.options['learn_mask']:
|
||||
|
|
@ -325,14 +321,17 @@ class SAEModel(ModelBase):
|
|||
self.AE_view = K.function ([warped_src, warped_dst], [pred_src_src[-1], pred_dst_dst[-1], pred_src_dst[-1], pred_src_dstm[-1]])
|
||||
else:
|
||||
self.AE_view = K.function ([warped_src, warped_dst], [pred_src_src[-1], pred_dst_dst[-1], pred_src_dst[-1] ] )
|
||||
|
||||
|
||||
self.load_weights_safe(weights_to_load)#, [ [self.src_dst_opt, 'src_dst_opt'], [self.src_dst_mask_opt, 'src_dst_mask_opt']])
|
||||
else:
|
||||
self.load_weights_safe(weights_to_load)
|
||||
if self.options['learn_mask']:
|
||||
self.AE_convert = K.function ([warped_dst],[ pred_src_dst[-1], pred_src_dstm[-1] ])
|
||||
else:
|
||||
self.AE_convert = K.function ([warped_dst],[ pred_src_dst[-1] ])
|
||||
|
||||
if self.is_training_mode:
|
||||
|
||||
|
||||
if self.is_training_mode:
|
||||
self.src_sample_losses = []
|
||||
self.dst_sample_losses = []
|
||||
|
||||
|
|
@ -353,6 +352,7 @@ class SAEModel(ModelBase):
|
|||
sample_process_options=SampleProcessor.Options(random_flip=self.random_flip, normalize_tanh = True),
|
||||
output_sample_types=output_sample_types )
|
||||
])
|
||||
|
||||
#override
|
||||
def onSave(self):
|
||||
opt_ar = [ [self.src_dst_opt, 'src_dst_opt'],
|
||||
|
|
|
|||
|
|
@ -71,8 +71,8 @@ ZeroPadding2D = keras.layers.ZeroPadding2D
|
|||
RandomNormal = keras.initializers.RandomNormal
|
||||
Model = keras.models.Model
|
||||
|
||||
#Adam = keras.optimizers.Adam
|
||||
Adam = nnlib.Adam
|
||||
Adam = keras.optimizers.Adam
|
||||
FastAdam = nnlib.FastAdam
|
||||
|
||||
modelify = nnlib.modelify
|
||||
gaussian_blur = nnlib.gaussian_blur
|
||||
|
|
@ -434,21 +434,16 @@ NLayerDiscriminator = nnlib.NLayerDiscriminator
|
|||
return dict(list(base_config.items()) + list(config.items()))
|
||||
nnlib.Scale = Scale
|
||||
|
||||
class Adam(keras.optimizers.Optimizer):
|
||||
def __init__(self, lr=0.001, beta_1=0.9, beta_2=0.999,
|
||||
epsilon=None, decay=0., amsgrad=False, iterations=0, **kwargs):
|
||||
super(Adam, self).__init__(**kwargs)
|
||||
class FastAdam(keras.optimizers.Optimizer):
|
||||
def __init__(self, lr=0.001, beta_1=0.9, beta_2=0.999, iterations=0, **kwargs):
|
||||
super(FastAdam, self).__init__(**kwargs)
|
||||
with K.name_scope(self.__class__.__name__):
|
||||
self.iterations = K.variable(iterations, dtype='int64', name='iterations')
|
||||
self.lr = K.variable(lr, name='lr')
|
||||
self.beta_1 = K.variable(beta_1, name='beta_1')
|
||||
self.beta_2 = K.variable(beta_2, name='beta_2')
|
||||
self.decay = K.variable(decay, name='decay')
|
||||
if epsilon is None:
|
||||
epsilon = K.epsilon()
|
||||
self.epsilon = epsilon
|
||||
self.initial_decay = decay
|
||||
self.amsgrad = amsgrad
|
||||
|
||||
self.epsilon = K.epsilon()
|
||||
|
||||
@keras.legacy.interfaces.legacy_get_updates_support
|
||||
def get_updates(self, loss, params):
|
||||
|
|
@ -456,34 +451,16 @@ NLayerDiscriminator = nnlib.NLayerDiscriminator
|
|||
self.updates = [K.update_add(self.iterations, 1)]
|
||||
|
||||
lr = self.lr
|
||||
if self.initial_decay > 0:
|
||||
lr = lr * (1. / (1. + self.decay * K.cast(self.iterations,
|
||||
K.dtype(self.decay))))
|
||||
|
||||
t = K.cast(self.iterations, K.floatx()) + 1
|
||||
lr_t = lr * (K.sqrt(1. - K.pow(self.beta_2, t)) /
|
||||
(1. - K.pow(self.beta_1, t)))
|
||||
(1. - K.pow(self.beta_1, t)))
|
||||
self.weights = [self.iterations]
|
||||
|
||||
ms = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
|
||||
vs = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
|
||||
if self.amsgrad:
|
||||
vhats = [K.zeros(K.int_shape(p), dtype=K.dtype(p)) for p in params]
|
||||
else:
|
||||
vhats = [K.zeros(1) for _ in params]
|
||||
self.weights = [self.iterations] + ms + vs + vhats
|
||||
for p, g in zip(params, grads):
|
||||
|
||||
for p, g, m, v, vhat in zip(params, grads, ms, vs, vhats):
|
||||
m_t = (self.beta_1 * m) + (1. - self.beta_1) * g
|
||||
v_t = (self.beta_2 * v) + (1. - self.beta_2) * K.square(g)
|
||||
if self.amsgrad:
|
||||
vhat_t = K.maximum(vhat, v_t)
|
||||
p_t = p - lr_t * m_t / (K.sqrt(vhat_t) + self.epsilon)
|
||||
self.updates.append(K.update(vhat, vhat_t))
|
||||
else:
|
||||
p_t = p - lr_t * m_t / (K.sqrt(v_t) + self.epsilon)
|
||||
|
||||
self.updates.append(K.update(m, m_t))
|
||||
self.updates.append(K.update(v, v_t))
|
||||
m_t = (1. - self.beta_1) * g
|
||||
v_t = (1. - self.beta_2) * K.square(g)
|
||||
p_t = p - lr_t * m_t / (K.sqrt(v_t) + self.epsilon)
|
||||
new_p = p_t
|
||||
|
||||
# Apply constraints.
|
||||
|
|
@ -497,15 +474,14 @@ NLayerDiscriminator = nnlib.NLayerDiscriminator
|
|||
config = {'iterations': int(K.get_value(self.iterations)),
|
||||
'lr': float(K.get_value(self.lr)),
|
||||
'beta_1': float(K.get_value(self.beta_1)),
|
||||
'beta_2': float(K.get_value(self.beta_2)),
|
||||
'decay': float(K.get_value(self.decay)),
|
||||
'epsilon': self.epsilon,
|
||||
'amsgrad': self.amsgrad}
|
||||
base_config = super(Adam, self).get_config()
|
||||
'beta_2': float(K.get_value(self.beta_2))
|
||||
}
|
||||
base_config = super(FastAdam, self).get_config()
|
||||
return dict(list(base_config.items()) + list(config.items()))
|
||||
nnlib.Adam = Adam
|
||||
|
||||
'''
|
||||
nnlib.FastAdam = FastAdam
|
||||
|
||||
'''
|
||||
not implemented in plaidML
|
||||
class ReflectionPadding2D(keras.layers.Layer):
|
||||
def __init__(self, padding=(1, 1), **kwargs):
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue