gan_model.py

from __future__ import absolute_import

import tensorflow as tf
import collections

a = None
EPS = 1e-12
Model = collections.namedtuple("Model", "outputs, predict_real, predict_fake, discrim_loss, discrim_grads_and_vars, gen_loss_GAN, gen_loss_L1, gen_grads_and_vars, train")


def conv(batch_input, out_channels, stride):
    with tf.variable_scope("conv"):
        in_channels = batch_input.get_shape()[3]
        filter = tf.get_variable("filter", [4, 4, in_channels, out_channels], dtype=tf.float32, initializer=tf.random_normal_initializer(0, 0.02))
        # [batch, in_height, in_width, in_channels], [filter_width, filter_height, in_channels, out_channels]
        #     => [batch, out_height, out_width, out_channels]
        padded_input = tf.pad(batch_input, [[0, 0], [1, 1], [1, 1], [0, 0]], mode="CONSTANT")
        conv = tf.nn.conv2d(padded_input, filter, [1, stride, stride, 1], padding="VALID")
        return conv


def lrelu(x, a):
    with tf.name_scope("lrelu"):
        # adding these together creates the leak part and linear part
        # then cancels them out by subtracting/adding an absolute value term
        # leak: a*x/2 - a*abs(x)/2
        # linear: x/2 + abs(x)/2

        # this block looks like it has 2 inputs on the graph unless we do this
        x = tf.identity(x)
        return (0.5 * (1 + a)) * x + (0.5 * (1 - a)) * tf.abs(x)


def batchnorm(input):
    with tf.variable_scope("batchnorm"):
        # this block looks like it has 3 inputs on the graph unless we do this
        input = tf.identity(input)

        channels = input.get_shape()[3]
        offset = tf.get_variable("offset", [channels], dtype=tf.float32, initializer=tf.zeros_initializer())
        scale = tf.get_variable("scale", [channels], dtype=tf.float32, initializer=tf.random_normal_initializer(1.0, 0.02))
        mean, variance = tf.nn.moments(input, axes=[0, 1, 2], keep_dims=False)
        variance_epsilon = 1e-5
        normalized = tf.nn.batch_normalization(input, mean, variance, offset, scale, variance_epsilon=variance_epsilon)
        return normalized


def deconv(batch_input, out_channels):
    with tf.variable_scope("deconv"):
        batch, in_height, in_width, in_channels = [int(d) for d in batch_input.get_shape()]
        filter = tf.get_variable("filter", [4, 4, out_channels, in_channels], dtype=tf.float32, initializer=tf.random_normal_initializer(0, 0.02))
        # [batch, in_height, in_width, in_channels], [filter_width, filter_height, out_channels, in_channels]
        #     => [batch, out_height, out_width, out_channels]
        conv = tf.nn.conv2d_transpose(batch_input, filter, [batch, in_height * 2, in_width * 2, out_channels], [1, 2, 2, 1], padding="SAME")
        return conv


def create_generator(generator_scenes, generator_inputs, generator_outputs_channels):
    layers = []

    # encoder_1: [batch, 256, 256, in_channels] => [batch, 128, 128, ngf]
    with tf.variable_scope("encoder_1"):
        output = conv(generator_inputs, a.ngf, stride=2)
        layers.append(output)

    layer_specs = [
        a.ngf * 2, # encoder_2: [batch, 128, 128, ngf] => [batch, 64, 64, ngf * 2]
        a.ngf * 4, # encoder_3: [batch, 64, 64, ngf * 2] => [batch, 32, 32, ngf * 4]
        a.ngf * 8, # encoder_4: [batch, 32, 32, ngf * 4] => [batch, 16, 16, ngf * 8]
        a.ngf * 8, # encoder_5: [batch, 16, 16, ngf * 8] => [batch, 8, 8, ngf * 8]
        a.ngf * 8, # encoder_6: [batch, 8, 8, ngf * 8] => [batch, 4, 4, ngf * 8]
        a.ngf * 8, # encoder_7: [batch, 4, 4, ngf * 8] => [batch, 2, 2, ngf * 8]
        a.ngf * 8, # encoder_8: [batch, 2, 2, ngf * 8] => [batch, 1, 1, ngf * 8]
    ]

    for out_channels in layer_specs:
        with tf.variable_scope("encoder_%d" % (len(layers) + 1)):
            rectified = lrelu(layers[-1], 0.2)
            # [batch, in_height, in_width, in_channels] => [batch, in_height/2, in_width/2, out_channels]
            convolved = conv(rectified, out_channels, stride=2)
            output = batchnorm(convolved)
            layers.append(output)
    num_encoder_layers = len(layers)

    # fusion: [batch, 1, 1, ngf * 8] + [batch, # scenes]=> [batch, 1, 1, ngf * 8]
    with tf.variable_scope("fusion"):
        encoder_flat = tf.reshape(layers[-1], [-1, a.ngf * 8])
        scenes_flat = tf.reshape(generator_scenes, [-1, 365])
        fusion_layer = tf.concat([encoder_flat, scenes_flat], axis=1)
        fusion_output = tf.layers.dense(inputs=fusion_layer, units=(a.ngf * 8), activation=tf.nn.relu)
        fusion_output = tf.reshape(fusion_output, layers[-1].shape)

    layer_specs = [
        (a.ngf * 8, 0.5),   # decoder_8: [batch, 1, 1, ngf * 8] => [batch, 2, 2, ngf * 8 * 2]
        (a.ngf * 8, 0.5),   # decoder_7: [batch, 2, 2, ngf * 8 * 2] => [batch, 4, 4, ngf * 8 * 2]
        (a.ngf * 8, 0.5),   # decoder_6: [batch, 4, 4, ngf * 8 * 2] => [batch, 8, 8, ngf * 8 * 2]
        (a.ngf * 8, 0.0),   # decoder_5: [batch, 8, 8, ngf * 8 * 2] => [batch, 16, 16, ngf * 8 * 2]
        (a.ngf * 4, 0.0),   # decoder_4: [batch, 16, 16, ngf * 8 * 2] => [batch, 32, 32, ngf * 4 * 2]
        (a.ngf * 2, 0.0),   # decoder_3: [batch, 32, 32, ngf * 4 * 2] => [batch, 64, 64, ngf * 2 * 2]
        (a.ngf, 0.0),       # decoder_2: [batch, 64, 64, ngf * 2 * 2] => [batch, 128, 128, ngf * 2]
    ]

    for decoder_layer, (out_channels, dropout) in enumerate(layer_specs):
        skip_layer = num_encoder_layers - decoder_layer - 1
        with tf.variable_scope("decoder_%d" % (skip_layer + 1)):
            if decoder_layer == 0:
                # first decoder layer doesn't have skip connections
                # since it is directly connected to the skip_layer
                input = fusion_output
            else:
                input = tf.concat([layers[-1], layers[skip_layer]], axis=3)

            rectified = tf.nn.relu(input)
            # [batch, in_height, in_width, in_channels] => [batch, in_height*2, in_width*2, out_channels]
            output = deconv(rectified, out_channels)
            output = batchnorm(output)

            if dropout > 0.0:
                output = tf.nn.dropout(output, keep_prob=1 - dropout)

            layers.append(output)

    # decoder_1: [batch, 128, 128, ngf * 2] => [batch, 256, 256, generator_outputs_channels]
    with tf.variable_scope("decoder_1"):
        input = tf.concat([layers[-1], layers[0]], axis=3)
        rectified = tf.nn.relu(input)
        output = deconv(rectified, generator_outputs_channels)
        output = tf.tanh(output)
        layers.append(output)

    return layers[-1]


def create_model(args, scenes, inputs, targets):
    global a
    a = args

    def create_discriminator(discrim_inputs, discrim_targets):
        n_layers = 3
        layers = []

        # 2x [batch, height, width, in_channels] => [batch, height, width, in_channels * 2]
        input = tf.concat([discrim_inputs, discrim_targets], axis=3)

        # layer_1: [batch, 256, 256, in_channels * 2] => [batch, 128, 128, ndf]
        with tf.variable_scope("layer_1"):
            convolved = conv(input, a.ndf, stride=2)
            rectified = lrelu(convolved, 0.2)
            layers.append(rectified)

        # layer_2: [batch, 128, 128, ndf] => [batch, 64, 64, ndf * 2]
        # layer_3: [batch, 64, 64, ndf * 2] => [batch, 32, 32, ndf * 4]
        # layer_4: [batch, 32, 32, ndf * 4] => [batch, 31, 31, ndf * 8]
        for i in range(n_layers):
            with tf.variable_scope("layer_%d" % (len(layers) + 1)):
                out_channels = a.ndf * min(2**(i+1), 8)
                stride = 1 if i == n_layers - 1 else 2  # last layer here has stride 1
                convolved = conv(layers[-1], out_channels, stride=stride)
                normalized = batchnorm(convolved)
                rectified = lrelu(normalized, 0.2)
                layers.append(rectified)

        # layer_5: [batch, 31, 31, ndf * 8] => [batch, 30, 30, 1]
        with tf.variable_scope("layer_%d" % (len(layers) + 1)):
            convolved = conv(rectified, out_channels=1, stride=1)
            output = tf.sigmoid(convolved)
            layers.append(output)

        return layers[-1]

    with tf.variable_scope("generator") as scope:
        out_channels = int(targets.get_shape()[-1])
        outputs = create_generator(scenes, inputs, out_channels)

    # create two copies of discriminator, one for real pairs and one for fake pairs
    # they share the same underlying variables
    with tf.name_scope("real_discriminator"):
        with tf.variable_scope("discriminator"):
            # 2x [batch, height, width, channels] => [batch, 30, 30, 1]
            predict_real = create_discriminator(inputs, targets)

    with tf.name_scope("fake_discriminator"):
        with tf.variable_scope("discriminator", reuse=True):
            # 2x [batch, height, width, channels] => [batch, 30, 30, 1]
            predict_fake = create_discriminator(inputs, outputs)

    with tf.name_scope("discriminator_loss"):
        # minimizing -tf.log will try to get inputs to 1
        # predict_real => 1
        # predict_fake => 0
        discrim_loss = tf.reduce_mean(-(tf.log(predict_real + EPS) + tf.log(1 - predict_fake + EPS)))

    with tf.name_scope("generator_loss"):
        # predict_fake => 1
        # abs(targets - outputs) => 0
        gen_loss_GAN = tf.reduce_mean(-tf.log(predict_fake + EPS))
        gen_loss_L1 = tf.reduce_mean(tf.abs(targets - outputs))
        gen_loss = gen_loss_GAN * a.gan_weight + gen_loss_L1 * a.l1_weight

    with tf.name_scope("discriminator_train"):
        discrim_tvars = [var for var in tf.trainable_variables() if var.name.startswith("discriminator")]
        discrim_optim = tf.train.AdamOptimizer(a.lr, a.beta1)
        discrim_grads_and_vars = discrim_optim.compute_gradients(discrim_loss, var_list=discrim_tvars)
        discrim_train = discrim_optim.apply_gradients(discrim_grads_and_vars)

    with tf.name_scope("generator_train"):
        with tf.control_dependencies([discrim_train]):
            gen_tvars = [var for var in tf.trainable_variables() if var.name.startswith("generator")]
            gen_optim = tf.train.AdamOptimizer(a.lr, a.beta1)
            gen_grads_and_vars = gen_optim.compute_gradients(gen_loss, var_list=gen_tvars)
            gen_train = gen_optim.apply_gradients(gen_grads_and_vars)

    ema = tf.train.ExponentialMovingAverage(decay=0.99)
    update_losses = ema.apply([discrim_loss, gen_loss_GAN, gen_loss_L1])

    global_step = tf.contrib.framework.get_or_create_global_step()
    incr_global_step = tf.assign(global_step, global_step+1)

    return Model(
        predict_real=predict_real,
        predict_fake=predict_fake,
        discrim_loss=ema.average(discrim_loss),
        discrim_grads_and_vars=discrim_grads_and_vars,
        gen_loss_GAN=ema.average(gen_loss_GAN),
        gen_loss_L1=ema.average(gen_loss_L1),
        gen_grads_and_vars=gen_grads_and_vars,
        outputs=outputs,
        train=tf.group(update_losses, incr_global_step, gen_train),
    )