speed-reproducibility fix #17

train.py

@@ -63,7 +63,7 @@ def train(hyp):
     weights = opt.weights  # initial training weights
 
     # Configure
-    init_seeds()
+    init_seeds(1)
     with open(opt.data) as f:
         data_dict = yaml.load(f, Loader=yaml.FullLoader)  # model dict
     train_path = data_dict['train']

utils/torch_utils.py

@@ -12,8 +12,11 @@ import torch.nn.functional as F
 def init_seeds(seed=0):
     torch.manual_seed(seed)
 
-    # Reduce randomness (may be slower on Tesla GPUs) # https://pytorch.org/docs/stable/notes/randomness.html
-    if seed == 0:
+    # Speed-reproducibility tradeoff https://pytorch.org/docs/stable/notes/randomness.html
+    if seed == 0:  # slower, more reproducible
+        cudnn.deterministic = True
+        cudnn.benchmark = False
+    else:  # faster, less reproducible
         cudnn.deterministic = False
         cudnn.benchmark = True