flrngel / cpc-tensorflow

Tensorflow implementation of "Representation Learning with Contrastive Predictive Coding"
64 stars 14 forks source link

Classification accuracy of the model #4

Closed ylfzr closed 5 years ago

ylfzr commented 6 years ago

Hi, I wonder if there is any hint on the final classification accuracy on the cpc representations. (One linear classifier built on top of the well-extracted cpc-features). I ran the code but only got around 25% accuracy on the FASHION_MNIST data after 30000 iters training in the final evaluation, which is weird.

There might be something wrong with my experiment, is here any thing that I should pay more attention run the code? I just directly run the model, train the cpc feature extractor and then set "mode" to be "validation" and train the final classifier.

flrngel commented 5 years ago

Default code makes 47% accuracy Thinking about method which use to be input for CPC and validation is to see cpc is working (which means freezing the model), performance of 47% seems awful but acceptable