fixed the bug raised by "tmp_eval_loss += tmp_eval_loss.item()" when parallelly using multi-gpu.

2019-10-23 20:27:13 +08:00
parent ef1b8b2ae5
commit bd847ce7d7
1 changed files with 3 additions and 0 deletions
--- a/examples/run_ner.py
+++ b/examples/run_ner.py
@@ -210,6 +210,9 @@ def evaluate(args, model, tokenizer, labels, pad_token_label_id, mode, prefix=""
            outputs = model(**inputs)
            tmp_eval_loss, logits = outputs[:2]

+            if args.n_gpu > 1:
+                tmp_eval_loss = tmp_eval_loss.mean()  # mean() to average on multi-gpu parallel evaluating
+
            eval_loss += tmp_eval_loss.item()
        nb_eval_steps += 1
        if preds is None: