Merge pull request #19 from Helsinki-NLP/fix/undo_forwardhook_sideeff…

…ects

undo forward-hook side-effect

onmt/model_builder.py

@@ -347,7 +347,9 @@ def build_task_specific_model(
     def has_grad_hook(module, input, output) -> None:
         for param in module.parameters(recurse=False):
             if param.requires_grad:
-                param.has_grad = True
+                # NB: we're looking at whether gradient will/has been computed, which is only the
+                # case when the module is training.
+                param.has_grad = module.training
 
     for module in nmt_model.modules():
         module.register_forward_hook(has_grad_hook)

onmt/trainer.py

@@ -447,6 +447,10 @@ def validate(self, valid_iter, moving_average=None, task=None):
         # Set model back to training mode.
         valid_model.train()
 
+        for p in self.model.parameters():
+            if hasattr(p, 'has_grad'):
+                p.has_grad = False
+
         return stats
 
     def _gradient_accumulation_over_lang_pairs(