6da0d3cd3b28bb85c2766a90fc205ab80b932eb0,niftynet/network/simulator_gan.py,ImageGenerator,layer_op,#ImageGenerator#,35

Before Change

        keep_prob_ph = 1  // not passed in as a placeholder
        for i in range(4):
            ch.append(round((ch[-1] + conditioning_channels * self.generator_shortcuts[i]) / 2))
            sz = [[round(i / 2) for i in sz[0]]] + sz
        if spatial_rank == 3:
            def resize_func(x, sz):
                sz_x = x.get_shape().as_list()
                r1 = tf.image.resize_images(tf.reshape(x, sz_x[:3] + [-1]), sz[0:2])
                r2 = tf.image.resize_images(tf.reshape(r1, [sz_x[0], sz[0] * sz[1], sz_x[3], -1]),
                                            [sz[0] * sz[1], sz[2]])
                return tf.reshape(r2, [sz_x[0]] + sz + [sz_x[-1]])
        elif spatial_rank == 2:
            resize_func = tf.image.resize_bilinear

        def concat_cond(x, i):
            if add_noise:
                noise = [tf.random_normal(x.get_shape().as_list()[0:-1] + [add_noise], 0, .1)]
            else:
                noise = []
            if conditioning is not None and self.generator_shortcuts[i]:
                with tf.name_scope("concat_conditioning"):
                    resized_cond = resize_func(conditioning, x.get_shape().as_list()[1:-1])
                    return tf.concat([x, resized_cond] + noise, axis=3)
            else:
                return x

        def conv(ch, x):
            with tf.name_scope("conv"):
                conv_layer = ConvolutionalLayer(ch, 3, w_initializer=w_init)
                return tf.nn.relu(conv_layer(x, is_training=is_training))

        def up(ch, x):
            with tf.name_scope("up"):
                deconv_layer = DeconvolutionalLayer(ch, 3, stride=2, w_initializer=w_init)
                return tf.nn.relu(deconv_layer(x, is_training=is_training))

        def up_block(ch, x, i):
            with tf.name_scope("up_block"):
                u = up(ch, x)
                cond = concat_cond(u, i)
                return conv(cond.get_shape().as_list()[-1], cond)

        def noise_to_image(sz, ch, random_source):
            noise_size = random_source.get_shape().as_list()[1]
            batch_size = random_source.get_shape().as_list()[0]
            with tf.name_scope("noise_to_image"):
                g_no_0 = np.prod(sz) * ch
                w1p = tf.get_variable("G_W1p", shape=[noise_size, g_no_0], initializer=w_init)
                b1p = tf.get_variable("G_b1p", shape=[g_no_0], initializer=b_init)
                g_h1p = tf.nn.dropout(tf.nn.relu(tf.matmul(random_source, w1p) + b1p), keep_prob_ph)
                g_h1p = tf.reshape(g_h1p, [batch_size] + sz + [ch])
                g_h1p = concat_cond(g_h1p, 0)
                return conv(ch + conditioning_channels, g_h1p)

        g_h1 = noise_to_image(sz[0], ch[0], random_source)
        g_h2 = up_block(ch[1], g_h1, 1)
        g_h3 = up_block(ch[2], g_h2, 2)
        g_h4 = up_block(ch[3], g_h3, 3)
        g_h5 = up_block(ch[4], g_h4, 4)  // did not implement different epsilon
        with tf.name_scope("final_image"):
            if add_noise:
                noise = tf.random_normal(g_h5.get_shape().as_list()[0:-1] + [add_noise], 0, .1)
                g_h5 = tf.concat([g_h5, noise], axis=3)
            x_sample = ConvolutionalLayer(1, 3, with_bn=False, with_bias=True,
                                          w_initializer=w_init,
                                          b_initializer=b_init)(g_h5, is_training=is_training)
            x_sample = tf.nn.dropout(tf.nn.tanh(x_sample), keep_prob_ph)
        with tf.name_scope("summaries_verbose"):
            tf.summary.histogram("hist_g_h2", g_h2, [logging.LOG])
            tf.summary.histogram("hist_g_h3", g_h3, [logging.LOG])

After Change

        keep_prob_ph = 1  // not passed in as a placeholder
        for i in range(4):
            ch.append(round((ch[-1] + conditioning_channels * self.generator_shortcuts[i]) / 2))
            sz = [[int(round(i / 2)) for i in sz[0]]] + sz
        if spatial_rank == 3:
            def resize_func(x, sz):
                sz_x = x.get_shape().as_list()
                r1 = tf.image.resize_images(tf.reshape(x, sz_x[:3] + [-1]), sz[0:2])
                r2 = tf.image.resize_images(tf.reshape(r1, [sz_x[0], sz[0] * sz[1], sz_x[3], -1]),
                                            [sz[0] * sz[1], sz[2]])
                resized_3d = tf.reshape(r2, [sz_x[0]] + sz + [sz_x[-1]])
                return resized_3d
        elif spatial_rank == 2:
            resize_func = tf.image.resize_bilinear

        def concat_cond(x, i):
            if add_noise:
                noise = [tf.random_normal(x.get_shape().as_list()[0:-1] + [add_noise], 0, .1)]
            else:
                noise = []
            if conditioning is not None and self.generator_shortcuts[i]:
                with tf.name_scope("concat_conditioning"):
                    resized_cond = resize_func(conditioning, x.get_shape().as_list()[1:-1])
                    return tf.concat([x, resized_cond] + noise, axis=-1)
            else:
                return x

        def conv(ch, x):
            with tf.name_scope("conv"):
                conv_layer = ConvolutionalLayer(ch, 3, w_initializer=w_init)
                return tf.nn.relu(conv_layer(x, is_training=is_training))

        def up(ch, x):
            with tf.name_scope("up"):
                deconv_layer = DeconvolutionalLayer(ch, 3, stride=2, w_initializer=w_init)
                return tf.nn.relu(deconv_layer(x, is_training=is_training))

        def up_block(ch, x, i):
            with tf.name_scope("up_block"):
                u = up(ch, x)
                cond = concat_cond(u, i)
                return conv(cond.get_shape().as_list()[-1], cond)

        def noise_to_image(sz, ch, random_source):
            noise_size = random_source.get_shape().as_list()[1]
            batch_size = random_source.get_shape().as_list()[0]
            with tf.name_scope("noise_to_image"):
                g_no_0 = np.prod(sz) * ch
                w1p = tf.get_variable("G_W1p", shape=[noise_size, g_no_0], initializer=w_init)
                b1p = tf.get_variable("G_b1p", shape=[g_no_0], initializer=b_init)
                g_h1p = tf.nn.dropout(tf.nn.relu(tf.matmul(random_source, w1p) + b1p), keep_prob_ph)
                g_h1p = tf.reshape(g_h1p, [batch_size] + sz + [ch])
                g_h1p = concat_cond(g_h1p, 0)
                return conv(ch + conditioning_channels, g_h1p)

        g_h1 = noise_to_image(sz[0], ch[0], random_source)
        g_h2 = up_block(ch[1], g_h1, 1)
        g_h3 = up_block(ch[2], g_h2, 2)
        g_h4 = up_block(ch[3], g_h3, 3)
        g_h5 = up_block(ch[4], g_h4, 4)  // did not implement different epsilon
        with tf.name_scope("final_image"):
            if add_noise:
                noise = tf.random_normal(g_h5.get_shape().as_list()[0:-1] + [add_noise], 0, .1)
                g_h5 = tf.concat([g_h5, noise], axis=3)
            x_sample = ConvolutionalLayer(1, 3, with_bn=False, with_bias=True,
                                          w_initializer=w_init,
                                          b_initializer=b_init)(g_h5, is_training=is_training)
            x_sample = tf.nn.dropout(tf.nn.tanh(x_sample), keep_prob_ph)
        //with tf.name_scope("summaries_verbose"):
        //    tf.summary.histogram("hist_g_h2", g_h2, [logging.LOG])
        //    tf.summary.histogram("hist_g_h3", g_h3, [logging.LOG])

Instances