search

domerin0 / neural-sentiment

A sentiment analyzer using deep rnns, built with TensorFlow
62 stars 22 forks source link

Training is not working? #35

Open EvGe22 opened 7 years ago

EvGe22 commented 7 years ago

I've just cloned your repo, downloaded the nltk tokenizer dataset and just tried to train the model using the default config. And that's what I've got after training it for 3 hours on CPU image The accuracy is just 50% all the time. Am I doing something wrong or maybe something has changed? Output of the python --version is Python 2.7.13 :: Anaconda custom (64-bit) The TensorFlow version is 1.2.1 Ubuntu 16.04

domerin0 commented 7 years ago

Thanks for bringing this to my attention.

I had this before as well, and what I found was changing the size of the hidden layer caused it to begin learning. It was pretty finicky though, working at specifically hidden_size=55 units, but not 57 or 53. I made some code changes to update this project to work with the latest tensorflow version.

There is one of two things going on:

  1. I need to check more closely to make sure I didn't introduce a bug or error
  2. I need to perform another hyper parameter search to find the new combination of hyper parameters that will work.

I intend to upload a model once I have one trained with this latest version.

Either way, I'm currently focused on my other project [inikdom/neural-chatbot]. I will be switching gears back to this project sometime later this week to look at the above points.