Is there exist performance issue when training or forward model pass?
take UNet for example，
···
def call(self, inputs, training=None, mask=None):
···
when training or prediction， self.backbone(inputs) is calculated for 5 times, but the input and backbone not changed，so can this
code can be changed to
x0, x1, x2, x3, x4 = self.backbone(inputs, training=training)
self.upsample2d_x2_block function can use x0, x1, x2, x3, x4 , in this way, the backbone will calculate only 1 time.
Is there exist performance issue when training or forward model pass? take UNet for example， ··· def call(self, inputs, training=None, mask=None): ···
when training or prediction， self.backbone(inputs) is calculated for 5 times, but the input and backbone not changed，so can this code can be changed to x0, x1, x2, x3, x4 = self.backbone(inputs, training=training) self.upsample2d_x2_block function can use x0, x1, x2, x3, x4 , in this way, the backbone will calculate only 1 time.
thank U.