Don't assign weights to changed topos.

src/analyzers/executables/cloudplot.py

@@ -71,8 +71,10 @@ def post_status_closure(status):
     if error_bars:
       means = predict[:,j,:].mean(axis=1)
       stds = predict[:,j,:].std(axis=1)
-      plt.errorbar(inputs, means, stds, uplimes=False, lolimes=False, capsize=3, fmt="none")
-      plt.scatter(inputs, means)
+      plt.errorbar(outputs[:,j], means, stds, uplims=False, lolims=False, capsize=3, fmt="none")
+      plt.scatter(outputs[:,j], means)
+      ll = np.amin(means - stds)
+      ul = np.amax(means + stds)
     else:
       for i in range(n_rows):
         known = outputs[i,j].repeat(n_samples)

src/common/mko/mko.py

@@ -59,9 +59,11 @@ def from_dict(data: dict) -> 'MKO':
   def parameterize_model(self):
     if self._compiled and self._has_weights:
       weights = encodings.b64decode_datatype(self._weights)
-      for i, layer in enumerate(self._model.layers):
-        layer.set_weights(weights[i])
-      self._parameterized = True
+      n_layers = len(self._model.layers)
+      if len(weights) == n_layers:
+        for i, layer in enumerate(self._model.layers):
+          layer.set_weights(weights[i])
+        self._parameterized = True
 
   @staticmethod
   def from_json(j_str: str) -> 'MKO':

src/common/ml/model_ops.py

@@ -10,7 +10,9 @@ def compile_model(model, hypers, optimizer_state=None):
     _ = model.optimizer.iterations
     model.optimizer._create_hypers()
     model.optimizer._create_slots(model.trainable_weights)
-    model.optimizer.set_weights(optimizer_state)
+    n_weights = len(model.optimizer.get_weights())
+    if n_weights == len(optimizer_state):
+      model.optimizer.set_weights(optimizer_state)
   else:
     K.set_value(model.optimizer.learning_rate, hypers['learning_rate'])
   return model