def __init__(self, num_classes, wd, name='AuxiliaryHeadCIFAR', **kwargs):
super(AuxiliaryHeadCIFAR, self).__init__(name=name, **kwargs)
self.features = Sequential([
ReLU(),
AveragePooling2D(5, strides=3, padding='valid'),
Conv2D(filters=128,
kernel_size=1,
strides=1,
padding='valid',
kernel_initializer=kernel_init(),
kernel_regularizer=regularizer(wd),
use_bias=False),
BatchNormalization(affine=True),
ReLU(),
Conv2D(filters=768,
kernel_size=2,
strides=1,
padding='valid',
kernel_initializer=kernel_init(),
kernel_regularizer=regularizer(wd),
use_bias=False),
BatchNormalization(affine=True),
ReLU()
])
self.classifier = Dense(num_classes,
kernel_initializer=kernel_init(),
kernel_regularizer=regularizer(wd))
def CifarModel(cfg, training=True, stem_multiplier=3, name='CifarModel'):
"""Cifar Model"""
logging.info(f"buliding {name}...")
input_size = cfg['input_size']
ch_init = cfg['init_channels']
layers = cfg['layers']
num_cls = cfg['num_classes']
wd = cfg['weights_decay']
genotype = eval("genotypes.%s" % cfg['arch'])
# define model
inputs = Input([input_size, input_size, 3], name='input_image')
if training:
drop_path_prob = Input([], name='drop_prob')
else:
drop_path_prob = None
ch_curr = stem_multiplier * ch_init
s0 = s1 = Sequential([
Conv2D(filters=ch_curr, kernel_size=3, strides=1, padding='same',
kernel_initializer=kernel_init(),
kernel_regularizer=regularizer(wd), use_bias=False),
BatchNormalization(affine=True)], name='stem')(inputs)
ch_curr = ch_init
reduction_prev = False
logits_aux = None
for layer_index in range(layers):
if layer_index in [layers // 3, 2 * layers // 3]:
ch_curr *= 2
reduction = True
else:
reduction = False
cell = Cell(genotype, ch_curr, reduction, reduction_prev, wd,
name=f'Cell_{layer_index}')
s0, s1 = s1, cell(s0, s1, drop_path_prob)
reduction_prev = reduction
if layer_index == 2 * layers // 3 and training:
logits_aux = AuxiliaryHeadCIFAR(num_cls, wd=wd)(s1)
fea = GlobalAveragePooling2D()(s1)
logits = Dense(num_cls, kernel_initializer=kernel_init(),
kernel_regularizer=regularizer(wd))(Flatten()(fea))
if training:
return Model((inputs, drop_path_prob), (logits, logits_aux), name=name)
else:
return Model(inputs, logits, name=name)
def _build_model(self):
"""Model"""
logging.info(f"buliding {self.name}...")
input_size = self.cfg['input_size']
ch_init = self.cfg['init_channels']
layers = self.cfg['layers']
num_cls = self.cfg['num_classes']
wd = self.cfg['weights_decay']
# define model
inputs = Input([input_size, input_size, 3], name='input_image')
alphas_normal = Input([None], name='alphas_normal')
alphas_reduce = Input([None], name='alphas_reduce')
betas_normal = Input([], name='betas_normal')
betas_reduce = Input([], name='betas_reduce')
alphas_reduce_weights = Softmax(
name='AlphasReduceSoftmax')(alphas_reduce)
alphas_normal_weights = Softmax(
name='AlphasNormalSoftmax')(alphas_normal)
betas_reduce_weights = SplitSoftmax(
range(2, 2 + self.steps), name='BetasReduceSoftmax')(betas_reduce)
betas_normal_weights = SplitSoftmax(
range(2, 2 + self.steps), name='BetasNormalSoftmax')(betas_normal)
ch_curr = self.stem_multiplier * ch_init
s0 = s1 = Sequential([
Conv2D(filters=ch_curr,
kernel_size=3,
strides=1,
padding='same',
kernel_initializer=kernel_init(),
kernel_regularizer=regularizer(wd),
use_bias=False),
BatchNormalization(affine=True)
],
name='stem')(inputs)
ch_curr = ch_init
reduction_prev = False
for layer_index in range(layers):
if layer_index in [layers // 3, 2 * layers // 3]:
ch_curr *= 2
reduction = True
weights = alphas_reduce_weights
edge_weights = betas_reduce_weights
else:
reduction = False
weights = alphas_normal_weights
edge_weights = betas_normal_weights
cell = Cell(self.steps,
self.multiplier,
ch_curr,
reduction,
reduction_prev,
wd,
name=f'Cell_{layer_index}')
s0, s1 = s1, cell(s0, s1, weights, edge_weights)
reduction_prev = reduction
fea = GlobalAveragePooling2D()(s1)
logits = Dense(num_cls,
kernel_initializer=kernel_init(),
kernel_regularizer=regularizer(wd))(Flatten()(fea))
return Model(
(inputs, alphas_normal, alphas_reduce, betas_normal, betas_reduce),
logits,
name=self.name)