Training resnet50 from examples on Imagnet fails


I am trying to simply train resnet50 on the imagenet dataset, with the given code in the examples of pytorch. No matter what i have tried, the training accuracy remains very very low, over numerous epochs. I have tried with no change to the parameters, and with setting lower learning rates. Both on single and multi gpus. Any idea what needs to be done to get the expected results ? looking at training graphs on the net, it seems that like other trainings, the accuracy rises quickly on the first epochs.

i do get these errors:

/usr/local/lib/python2.7/dist-packages/PIL/ UserWarning: Corrupt EXIF data. Expecting to read 4 bytes but only got 0.

but from what i read it seems these shouldn’t effect the traning.