def _generator_model(sess, features):
# See Arxiv 1603.05027
model = dm_arch.Model('GENE', 2 * features - 1)
mapsize = 3
# Encoder
layers = [24, 48]
for nunits in layers:
_residual_block(model, nunits, mapsize)
model.add_avg_pool()
# Decoder
layers = [96, 64]
for nunits in layers:
_residual_block(model, nunits, mapsize)
_residual_block(model, nunits, mapsize)
model.add_upscale()
nunits = 48
_residual_block(model, nunits, mapsize)
_residual_block(model, nunits, 1)
model.add_conv2d(3, mapsize=1)
model.add_sigmoid(1.1)
return model
def _discriminator_model(sess, image):
model = dm_arch.Model('DISC', 2 * image - 1.0)
mapsize = 3
layers = [64, 96, 128, 192] #[32, 48, 96, 128]
for nunits in layers:
model.add_batch_norm()
model.add_lrelu()
model.add_conv2d(nunits, mapsize=mapsize)
model.add_avg_pool()
nunits = layers[-1]
model.add_batch_norm()
model.add_lrelu()
model.add_conv2d(nunits, mapsize=mapsize)
#model.add_batch_norm()
model.add_lrelu()
model.add_conv2d(1, mapsize=mapsize)
model.add_mean()
return model
def _discriminator_model(sess, image):
model = dm_arch.Model('DISC', 2 * image - 1.0)
mapsize = 3
layers = [32, 48, 96, 128]
for nunits in layers:
_residual_block(model, nunits, mapsize)
model.add_avg_pool()
nunits = layers[-1]
_residual_block(model, nunits, mapsize)
model.add_conv2d(1, mapsize=1, stride=1)
model.add_mean()
return model