search

localminimum / R-net

A Tensorflow Implementation of R-net: Machine reading comprehension with self matching networks
MIT License
323 stars 122 forks source link

what were your performance results mate? the EM and F1? #2

Closed Nick7hill closed 7 years ago

ghost commented 7 years ago

Currently I am struggling to reduce the gpu memory usage in tensorflow to fit a minibatch of 48 and higher. Until then I can't train with full dataset, as the dataset shows high variance and it won't converge with small mini-batches.

ghost commented 7 years ago

Please take a look at ec238d6f7ac57890423298acaf71befbceddf012 and 0523897bbf61db0df75b60dd09ff56cdff7d8310. Now you can check your training set F1 and EM from Tensorboard while training, as well as checking your dev set F1 and EM by changing test = True from params.py file and running the model.py file. If you obtain any results please report the hyperparameters you used with results here so we can share the results and work towards improving the implementation.

Nick7hill commented 7 years ago

yeah, sure, thanks a lot! your have done tremendous work in trying to emulate R-net

Nick7hill commented 7 years ago

can you please make a blog post about your implementation, it is kind of difficult to correlate the code with the architecture at times, and if you don't mind comment the code at certain places, i suggest jupyter notebooks are better

ghost commented 7 years ago

Thanks for your input! I'll look into jupyter notebooks. In the meantime, I suggest you look into the tensorboard graph of the R-Net implementation. It helps you understand what's going on behind each node better than the lines of code. (https://github.com/minsangkim142/R-net/blob/master/screenshots/graph.png) Just run this command on bash once you have a saved log directory tensorboard --logdir=./YOUR_LOGDIR

ghost commented 7 years ago

Let's keep all performance related discussions in this issue #12 thanks!