def generator_outputs(inputs, sizes):
latent = inputs[0]
input_energy = inputs[1]
h = Lambda(lambda x: x[0] * x[1])([latent, scale(input_energy, 100)])
img_layer0 = build_generator(h, sizes[0], sizes[1])
img_layer1 = build_generator(h, sizes[2], sizes[3])
img_layer2 = build_generator(h, sizes[4], sizes[5])
avgpool = AveragePooling2D(pool_size=(1, 8))
zero2one = avgpool(UpSampling2D(size=(4, 1))(img_layer0))
img_layer1 = inpainting_attention(img_layer1, zero2one)
one2two = AveragePooling2D(pool_size=(1, 2))(img_layer1)
img_layer2 = inpainting_attention(img_layer2, one2two)
outputs = [Activation('relu')(img_layer0),
Activation('relu')(img_layer1),
Activation('relu')(img_layer2)]
return outputs
h = Lambda(lambda x: x[0] * x[1])([latent, scale(input_energy, 100)])
# each of these builds a LAGAN-inspired [arXiv/1701.05927] component with
# linear last layer
img_layer0 = build_generator(h, 3, 96)
img_layer1 = build_generator(h, 12, 12)
img_layer2 = build_generator(h, 12, 6)
if not no_attn:
logger.info('using attentional mechanism')
# resizes from (3, 96) => (12, 12)
zero2one = AveragePooling2D(pool_size=(1, 8))(
UpSampling2D(size=(4, 1))(img_layer0))
img_layer1 = inpainting_attention(img_layer1, zero2one)
# resizes from (12, 12) => (12, 6)
one2two = AveragePooling2D(pool_size=(1, 2))(img_layer1)
img_layer2 = inpainting_attention(img_layer2, one2two)
generator_outputs = [
Activation('relu')(img_layer0),
Activation('relu')(img_layer1),
Activation('relu')(img_layer2)
]
generator = Model(generator_inputs, generator_outputs)
generator.compile(optimizer=Adam(lr=gen_lr, beta_1=adam_beta_1),
loss='binary_crossentropy')
generator_inputs = [latent, input_energy]
# multiply the (scaled) energy into the latent space
h = Lambda(lambda x: x[0] * x[1])([latent, scale(input_energy, 100)])
# build three LAGAN-style generators (checkout out `build_generator` in architectures.py)
img_layer0 = build_generator(h, 3, 96)
img_layer1 = build_generator(h, 12, 12)
img_layer2 = build_generator(h, 12, 6)
# inpainting
# 0 --> 1
zero2one = AveragePooling2D(pool_size=(1,
8))(UpSampling2D(size=(4,
1))(img_layer0))
img_layer1 = inpainting_attention(img_layer1,
zero2one) # this function is in ops.py
# 1 --> 2
one2two = AveragePooling2D(pool_size=(1, 2))(img_layer1)
img_layer2 = inpainting_attention(img_layer2, one2two)
# ^^ pooling and upsampling are needed to reshape images to same dimensions
# outputs
generator_outputs = [
Activation('relu')(img_layer0),
Activation('relu')(img_layer1),
Activation('relu')(img_layer2)
]
# build the actual model
generator = Model(generator_inputs, generator_outputs)
generator_inputs.append(image_class)
else:
# requested energy comes in GeV
h = Lambda(lambda x: x[0] * x[1])([latent, scale(input_energy, 100)])
# each of these builds a LAGAN-inspired [arXiv/1701.05927] component
img_layer = []
for i in range(5):
img_layer.append(build_generator(h, 12, 12))
if not no_attn:
logger.info('using attentional mechanism')
zero2one = AveragePooling2D(pool_size=(1, 1))(
UpSampling2D(size=(1, 1))(img_layer[0]))
img_layer[1] = inpainting_attention(img_layer[1], zero2one)
for j in range(1, 4):
one2N = AveragePooling2D(pool_size=(1, 1))(img_layer[j])
img_layer[j + 1] = inpainting_attention(img_layer[j + 1], one2N)
generator_outputs = []
for k in range(5):
generator_outputs.append(Activation('relu')(img_layer[k]))
generator = Model(generator_inputs, generator_outputs)
generator.compile(optimizer=Adam(lr=gen_lr, beta_1=adam_beta_1),
loss='binary_crossentropy')
discriminator.trainable = False
