def _input(self, inputs):
with tf.variable_scope('input'):
u = UpSampling2D('upsampling_u',
inputs=inputs[self.KEYS.TENSOR.INPUT],
size=self.config(
self.KEYS.CONFIG.UPSAMPLE_RATIO))()
if self.KEYS.TENSOR.LABEL in inputs:
l = inputs[self.KEYS.TENSOR.LABEL]
else:
l = None
if self.config(self.KEYS.CONFIG.INTERP):
return u, None, l
if SRKeys.REPRESENTS in inputs:
r = UpSampling2D('upsampling_r',
inputs=inputs[SRKeys.REPRESENTS],
size=self.config(
self.KEYS.CONFIG.UPSAMPLE_RATIO))()
r = align_crop(r, u)
else:
r = tf.layers.conv2d(inputs=u,
filters=self.config(
self.KEYS.CONFIG.FILTERS),
kernel_size=5,
padding='same',
name='stem',
reuse=tf.AUTO_REUSE)
return u, r, l
def kernel(self, inputs):
with tf.variable_scope('input'):
u = UpSampling2D(inputs=inputs[self.KEYS.TENSOR.INPUT],
size=(2, 2))()
if SRKeys.REPRESENTS in inputs:
r = UpSampling2D(inputs=inputs[SRKeys.REPRESENTS],
size=(2, 2))()
r = align_crop(r, u)
r = tf.concat([r, u], axis=3)
else:
r = tf.layers.conv2d(inputs=u,
filters=self.config(
self.KEYS.CONFIG.FILTERS),
kernel_size=5,
name='stem0')
key = self.KEYS.GRAPHS.SHORT_CUT
x = self.get_or_create_graph(key, self._short_cut(key))(r)
with tf.variable_scope('inference'):
res = tf.layers.conv2d(
inputs=x,
filters=1,
kernel_size=3,
padding='same',
name='stem1',
)
res = boundary_crop(res,
self.config(self.KEYS.CONFIG.BOUNDARY_CROP))
u_c = align_crop(u, res)
y = res + u_c
result = {
self.KEYS.TENSOR.INFERENCE: y,
SRKeys.REPRESENTS: x,
SRKeys.RESIDUAL: res,
SRKeys.INTERP: u_c
}
if self.KEYS.TENSOR.LABEL in inputs:
with tf.name_scope('loss'):
aligned_label = align_crop(inputs[self.KEYS.TENSOR.LABEL], y)
l = mean_square_error(aligned_label, y)
result.update({
self.KEYS.TENSOR.LOSS: l,
SRKeys.ALIGNED_LABEL: aligned_label
})
return result
def _loss(self, label, infer):
if label is not None:
with tf.name_scope('loss'):
align_label = align_crop(label, infer)
if self.config(self.KEYS.CONFIG.USE_COMBINED_LOSS):
with tf.name_scope('use_combine_loss'):
stdv = tf.constant(
self.config(self.KEYS.CONFIG.DENORM_STD),
tf.float32)
meanv = tf.constant(
self.config(self.KEYS.CONFIG.DENORM_MEAN,
tf.float32))
labeld = align_label * stdv + meanv
inferd = infer * stdv + meanv
result = CombinedSupervisedLoss(self.name / 'loss',
inputs={
self.KEYS.TENSOR.INPUT:
inferd,
self.KEYS.TENSOR.LABEL:
labeld
})()
result.update({SRKeys.ALIGNED_LABEL: align_label})
result.update({
self.KEYS.TENSOR.LOSS:
result[self.KEYS.TENSOR.OUTPUT]
})
result.pop(self.KEYS.TENSOR.OUTPUT)
return result
else:
loss_mse = mean_square_error(
align_label, infer) * self.config(
self.KEYS.CONFIG.MES_LOSS_WEIGHT)
loss = loss_mse
if self.config(self.KEYS.CONFIG.WITH_POI_LOSS):
loss_poi = poission_loss(
align_label, infer) * self.config(
self.KEYS.CONFIG.POI_LOSS_WEIGHT)
loss = loss + loss_poi
result = {
self.KEYS.TENSOR.LOSS: loss,
SRKeys.MSE_LOSS: loss_mse,
SRKeys.ALIGNED_LABEL: align_label
}
return result
else:
return {}
def _inference(self, represents, upsampled):
with tf.variable_scope('inference'):
upsampled = boundary_crop(input_=upsampled,
offset=self.config(
self.KEYS.CONFIG.BOUNDARY_CROP))
if self.config(self.KEYS.CONFIG.INTERP):
return {self.KEYS.TENSOR.INFERENCE: upsampled}
residual = tf.layers.conv2d(inputs=represents,
filters=1,
kernel_size=3,
padding='same',
reuse=tf.AUTO_REUSE)
residual = align_crop(input_=residual, target=upsampled)
inference = residual + upsampled
return {
self.KEYS.TENSOR.INFERENCE: inference,
SRKeys.RESIDUAL: residual,
SRKeys.INTERP: upsampled
}