Tweak accuracy topk safety. Fix #807

Author: rwightman

timm/utils/metrics.py

@@ -2,7 +2,6 @@
 
 Hacked together by / Copyright 2020 Ross Wightman
 """
-import torch
 
 
 class AverageMeter:
@@ -30,7 +29,4 @@ def accuracy(output, target, topk=(1,)):
     _, pred = output.topk(maxk, 1, True, True)
     pred = pred.t()
     correct = pred.eq(target.reshape(1, -1).expand_as(pred))
-    return [
-        correct[:k].reshape(-1).float().sum(0) * 100. / batch_size
-        if k <= maxk else torch.tensor(100.) for k in topk
-    ]
+    return [correct[:min(k, maxk)].reshape(-1).float().sum(0) * 100. / batch_size for k in topk]