sthalles / PyTorch-BYOL

PyTorch implementation of Bootstrap Your Own Latent: A New Approach to Self-Supervised Learning
480 stars 72 forks source link

Unable to replicate experiments for 80 epochs #4

Open snailrowen1337 opened 4 years ago

snailrowen1337 commented 4 years ago

Thanks a ton for making this open-source. I have a question about replicating the 75% accuracy for 80 epochs. I just changed the config file to read "max_epochs: 80" and then trained the model, but it only reached roughly 70% accuracy, see below. Is there anything else that needs to be changed to reach the 75% accuracy you report? Thanks!!

Testing accuracy: 69.5125 Testing accuracy: 69.5125 Testing accuracy: 69.6875 Testing accuracy: 69.9125 Testing accuracy: 69.7875 Testing accuracy: 69.7 Testing accuracy: 69.625 Testing accuracy: 69.8625 Testing accuracy: 69.65 Testing accuracy: 69.5375 Testing accuracy: 70.025 Testing accuracy: 69.8125

BrettLL commented 4 years ago

@snailrowen1337 my test acc is always: 10. no matter which epoch I used.

soumye commented 3 years ago

@snailrowen1337 I had the same issue. Did you find any fix for this problem?