f8e1c6af00330faeaa386e951d82a206ba735bfa,model.py,Artgan,_build_model,#Artgan#,72

Before Change

                              tf.add_n(self.output_photo_discr_loss.values())

            // Compute discriminator accuracies.
            self.input_painting_discr_acc = {key: tf.reduce_mean(tf.cast(x=(pred > tf.ones_like(pred)*0.5),
                                                                         dtype=tf.float32)) * scale_weight[key]
                                             for key, pred in zip(self.input_painting_discr_predictions.keys(),
                                                                  self.input_painting_discr_predictions.values())}
            self.input_photo_discr_acc = {key: tf.reduce_mean(tf.cast(x=(pred < tf.ones_like(pred)*0.5),
                                                                      dtype=tf.float32)) * scale_weight[key]
                                          for key, pred in zip(self.input_photo_discr_predictions.keys(),
                                                               self.input_photo_discr_predictions.values())}
            self.output_photo_discr_acc = {key: tf.reduce_mean(tf.cast(x=(pred < tf.ones_like(pred)*0.5),
                                                                       dtype=tf.float32)) * scale_weight[key]
                                           for key, pred in zip(self.output_photo_discr_predictions.keys(),
                                                                self.output_photo_discr_predictions.values())}
            self.discr_acc = (tf.add_n(self.input_painting_discr_acc.values()) + \
                              tf.add_n(self.input_photo_discr_acc.values()) + \
                              tf.add_n(self.output_photo_discr_acc.values())) / float(len(scale_weight.keys())*3)


            // Generator.
            // Predicts ones for both output images.
            self.output_photo_gener_loss = {key: self.loss(pred, tf.ones_like(pred)) * scale_weight[key]
                                            for key, pred in zip(self.output_photo_discr_predictions.keys(),
                                                                 self.output_photo_discr_predictions.values())}

            self.gener_loss = tf.add_n(self.output_photo_gener_loss.values())

            // Compute generator accuracies.
            self.output_photo_gener_acc = {key: tf.reduce_mean(tf.cast(x=(pred > tf.ones_like(pred)*0.5),
                                                                       dtype=tf.float32)) * scale_weight[key]
                                           for key, pred in zip(self.output_photo_discr_predictions.keys(),
                                                                self.output_photo_discr_predictions.values())}

            self.gener_acc = tf.add_n(self.output_photo_gener_acc.values()) / float(len(scale_weight.keys()))


            // Image loss.
            self.img_loss_photo = mse_criterion(transformer_block(self.output_photo),
                                                transformer_block(self.input_photo))
            self.img_loss = self.img_loss_photo

            // Features loss.
            self.feature_loss_photo = abs_criterion(self.output_photo_features, self.input_photo_features)
            self.feature_loss = self.feature_loss_photo

            // ================== Define optimization steps. =============== //
            t_vars = tf.trainable_variables()
            self.discr_vars = [var for var in t_vars if "discriminator" in var.name]
            self.encoder_vars = [var for var in t_vars if "encoder" in var.name]
            self.decoder_vars = [var for var in t_vars if "decoder" in var.name]

            // Discriminator and generator steps.
            update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)

            with tf.control_dependencies(update_ops):
                self.d_optim_step = tf.train.AdamOptimizer(self.lr).minimize(
                    loss=self.options.discr_loss_weight * self.discr_loss,
                    var_list=[self.discr_vars])
                self.g_optim_step = tf.train.AdamOptimizer(self.lr).minimize(
                    loss=self.options.discr_loss_weight * self.gener_loss +
                         self.options.transformer_loss_weight * self.img_loss +
                         self.options.feature_loss_weight * self.feature_loss,
                    var_list=[self.encoder_vars + self.decoder_vars])

            // ============= Write statistics to tensorboard. ================ //

            // Discriminator loss summary.
            s_d1 = [tf.summary.scalar("discriminator/input_painting_discr_loss/"+key, val)
                    for key, val in zip(self.input_painting_discr_loss.keys(), self.input_painting_discr_loss.values())]
            s_d2 = [tf.summary.scalar("discriminator/input_photo_discr_loss/"+key, val)
                    for key, val in zip(self.input_photo_discr_loss.keys(), self.input_photo_discr_loss.values())]
            s_d3 = [tf.summary.scalar("discriminator/output_photo_discr_loss/" + key, val)
                    for key, val in zip(self.output_photo_discr_loss.keys(), self.output_photo_discr_loss.values())]
            s_d = tf.summary.scalar("discriminator/discr_loss", self.discr_loss)
            self.summary_discriminator_loss = tf.summary.merge(s_d1+s_d2+s_d3+[s_d])

            // Discriminator acc summary.
            s_d1_acc = [tf.summary.scalar("discriminator/input_painting_discr_acc/"+key, val)
                    for key, val in zip(self.input_painting_discr_acc.keys(), self.input_painting_discr_acc.values())]
            s_d2_acc = [tf.summary.scalar("discriminator/input_photo_discr_acc/"+key, val)
                    for key, val in zip(self.input_photo_discr_acc.keys(), self.input_photo_discr_acc.values())]
            s_d3_acc = [tf.summary.scalar("discriminator/output_photo_discr_acc/" + key, val)
                    for key, val in zip(self.output_photo_discr_acc.keys(), self.output_photo_discr_acc.values())]
            s_d_acc = tf.summary.scalar("discriminator/discr_acc", self.discr_acc)
            s_d_acc_g = tf.summary.scalar("discriminator/discr_acc", self.gener_acc)
            self.summary_discriminator_acc = tf.summary.merge(s_d1_acc+s_d2_acc+s_d3_acc+[s_d_acc])

After Change

                                                                         dtype=tf.float32)) * scale_weight[key]
                                             for key, pred in zip(self.input_painting_discr_predictions.keys(),
                                                                  self.input_painting_discr_predictions.values())}
            self.input_photo_discr_acc = {key: tf.reduce_mean(tf.cast(x=(pred < tf.zeros_like(pred)),
                                                                      dtype=tf.float32)) * scale_weight[key]
                                          for key, pred in zip(self.input_photo_discr_predictions.keys(),
                                                               self.input_photo_discr_predictions.values())}
            self.output_photo_discr_acc = {key: tf.reduce_mean(tf.cast(x=(pred < tf.zeros_like(pred)),
                                                                       dtype=tf.float32)) * scale_weight[key]
                                           for key, pred in zip(self.output_photo_discr_predictions.keys(),
                                                                self.output_photo_discr_predictions.values())}
            self.discr_acc = (tf.add_n(self.input_painting_discr_acc.values()) + \
                              tf.add_n(self.input_photo_discr_acc.values()) + \
                              tf.add_n(self.output_photo_discr_acc.values())) / float(len(scale_weight.keys())*3)


            // Generator.
            // Predicts ones for both output images.
            self.output_photo_gener_loss = {key: self.loss(pred, tf.ones_like(pred)) * scale_weight[key]
                                            for key, pred in zip(self.output_photo_discr_predictions.keys(),
                                                                 self.output_photo_discr_predictions.values())}

            self.gener_loss = tf.add_n(self.output_photo_gener_loss.values())

            // Compute generator accuracies.
            self.output_photo_gener_acc = {key: tf.reduce_mean(tf.cast(x=(pred > tf.zeros_like(pred)),
                                                                       dtype=tf.float32)) * scale_weight[key]
                                           for key, pred in zip(self.output_photo_discr_predictions.keys(),
                                                                self.output_photo_discr_predictions.values())}

            self.gener_acc = tf.add_n(self.output_photo_gener_acc.values()) / float(len(scale_weight.keys()))


            // Image loss.
            self.img_loss_photo = mse_criterion(transformer_block(self.output_photo),
                                                transformer_block(self.input_photo))
            self.img_loss = self.img_loss_photo

            // Features loss.
            self.feature_loss_photo = abs_criterion(self.output_photo_features, self.input_photo_features)
            self.feature_loss = self.feature_loss_photo

            // ================== Define optimization steps. =============== //
            t_vars = tf.trainable_variables()
            self.discr_vars = [var for var in t_vars if "discriminator" in var.name]
            self.encoder_vars = [var for var in t_vars if "encoder" in var.name]
            self.decoder_vars = [var for var in t_vars if "decoder" in var.name]

            // Discriminator and generator steps.
            update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)

            with tf.control_dependencies(update_ops):
                self.d_optim_step = tf.train.AdamOptimizer(self.lr).minimize(
                    loss=self.options.discr_loss_weight * self.discr_loss,
                    var_list=[self.discr_vars])
                self.g_optim_step = tf.train.AdamOptimizer(self.lr).minimize(
                    loss=self.options.discr_loss_weight * self.gener_loss +
                         self.options.transformer_loss_weight * self.img_loss +
                         self.options.feature_loss_weight * self.feature_loss,
                    var_list=[self.encoder_vars + self.decoder_vars])

            // ============= Write statistics to tensorboard. ================ //

            // Discriminator loss summary.
            s_d1 = [tf.summary.scalar("discriminator/input_painting_discr_loss/"+key, val)
                    for key, val in zip(self.input_painting_discr_loss.keys(), self.input_painting_discr_loss.values())]
            s_d2 = [tf.summary.scalar("discriminator/input_photo_discr_loss/"+key, val)
                    for key, val in zip(self.input_photo_discr_loss.keys(), self.input_photo_discr_loss.values())]
            s_d3 = [tf.summary.scalar("discriminator/output_photo_discr_loss/" + key, val)
                    for key, val in zip(self.output_photo_discr_loss.keys(), self.output_photo_discr_loss.values())]
            s_d = tf.summary.scalar("discriminator/discr_loss", self.discr_loss)
            self.summary_discriminator_loss = tf.summary.merge(s_d1+s_d2+s_d3+[s_d])

            // Discriminator acc summary.
            s_d1_acc = [tf.summary.scalar("discriminator/input_painting_discr_acc/"+key, val)
                    for key, val in zip(self.input_painting_discr_acc.keys(), self.input_painting_discr_acc.values())]
            s_d2_acc = [tf.summary.scalar("discriminator/input_photo_discr_acc/"+key, val)
                    for key, val in zip(self.input_photo_discr_acc.keys(), self.input_photo_discr_acc.values())]
            s_d3_acc = [tf.summary.scalar("discriminator/output_photo_discr_acc/" + key, val)
                    for key, val in zip(self.output_photo_discr_acc.keys(), self.output_photo_discr_acc.values())]
            s_d_acc = tf.summary.scalar("discriminator/discr_acc", self.discr_acc)
            s_d_acc_g = tf.summary.scalar("discriminator/discr_acc", self.gener_acc)
            self.summary_discriminator_acc = tf.summary.merge(s_d1_acc+s_d2_acc+s_d3_acc+[s_d_acc])

Instances