Open renganxu opened 6 years ago
I think you are looking at "training accuracy" which can be as high as 100% if overfitting https://github.com/caffe2/caffe2/blob/master/caffe2/python/examples/resnet50_trainer.py#L201
while top-1 80% accuracy typically refers to the test_accuracy or validation accuracy https://github.com/caffe2/caffe2/blob/master/caffe2/python/examples/resnet50_trainer.py#L234
@0wu Thanks for the clarification. But since my test data is not None, why the log does not output the test accuracy? How to make it also output the test accuracy? I haven't tried python gdb yet, I will try it to run this python program step by step.
I was training resnet50 from scratch with the example provided in Caffe2. The training is on one server with 4 V100 GPUs. The following is the command I used:
I trained with 100 epochs and the acuracy is as high as 95%. But based on the accuracy operator definition on https://caffe2.ai/docs/operators-catalogue.html#accuracy, the default accuracy should be top-1 accuracy and it should be less than 80% for Imagenet dataset. I also didn't find any code that changed the default accuracy level. So is there something wrong with the Resnet50 implementation? Or I did anything wrong?
The following is the last few lines in the training log.