Fix errors when non-adversarial training

segmatch/python/autoencoder/model.py

@@ -409,7 +409,8 @@ def train_on_single_batch(self, batch_input, train_target=None, adversarial=Fals
     cost = [self.cost]
     if adversarial: cost = cost + [self.generator_loss_no_MI, self.discriminator_loss_no_MI, self.mutual_information_est]
     opt = train_target if train_target is not None else self.optimizer
-    if opt is self.discriminator_optimizer or opt is self.generator_optimizer: dict_[self.stop_gradient_placeholder] = True
+    if adversarial:
+      if opt is self.discriminator_optimizer or opt is self.generator_optimizer: dict_[self.stop_gradient_placeholder] = True
     # compute
     cost, _, _, summary = self.sess.run((cost, opt, self.catch_nans, self.merged), feed_dict=dict_)
     if summary_writer is not None: summary_writer.add_summary(summary)

segmatch/python/autoencoder_node.ipynb

@@ -508,11 +508,15 @@
     "  step_times = {'batchmaking': zero, 'training': zero, 'plotting': zero}\n",
     "  avg_step_cost = Average()\n",
     "  training_batchmaker = Batchmaker(train_vox, BATCH_SIZE, MP)\n",
-    "  train_order = 4*[vae.optimizer] + 4*[vae.generator_optimizer] + [vae.discriminator_optimizer]\n",
+    "  if ADVERSARIAL:\n",
+    "    train_order = 4*[vae.optimizer] + 4*[vae.generator_optimizer] + [vae.discriminator_optimizer]\n",
+    "  else:\n",
+    "    train_order = [vae.optimizer]\n",
     "  for train_target in itertools.cycle(train_order):\n",
-    "    if train_target is vae.discriminator_optimizer:\n",
-    "      if cost_value[1] > G_THRESHOLD or cost_value[2] < D_THRESHOLD:\n",
-    "        continue\n",
+    "    if ADVERSARIAL:\n",
+    "      if train_target is vae.discriminator_optimizer:\n",
+    "        if cost_value[1] > G_THRESHOLD or cost_value[2] < D_THRESHOLD:\n",
+    "          continue\n",
     "    if training_batchmaker.is_depleted():\n",
     "      break\n",
     "    else:\n",