Rework loading of 'fast-recurrent' sampler.

sample/shape_sampler.py

@@ -12,15 +12,15 @@ class ShapeSampler:
         Args:
             checkpoint_path: Relative or absolute path to a checkpoint file ("*.ckpt") containing a trained model.
             fast_recurrent: Changes the 'fast' architecture of the Transformer into an equivalent, but recurrent
-                formulation durring inference time.
+                formulation durring inference time, otherwise uses the standard full pass technique.
             device: Selects the device on which the sampling should be performed. Either "cpu" or "cuda" (gpu-support)
                 available.
         """
         # load and restore model from checkpoint
-        if fast_recurrent is True:
-            pl_module = ShapeTransformer.load_from_checkpoint(checkpoint_path, architecture='fast-recurrent')
-        else:
         pl_module = ShapeTransformer.load_from_checkpoint(checkpoint_path)
+        if fast_recurrent is True and pl_module.hparams['architecture'] == 'fast':
+            print("Reload model as a recurrent implementation for a major improvement of inference time.")
+            pl_module = ShapeTransformer.load_from_checkpoint(checkpoint_path, architecture='fast-recurrent')
         pl_module.freeze()
 
         # extract hyperparameters from the model