DocDex | JSON API & Discord Bot for Javadocs | REST library

I have written tensorflow code based on:

http://www.wildml.com/2015/12/implementing-a-cnn-for-text-classification-in-tensorflow/

but using precomputed word embeddings from the GoogleNews word2vec 300 dimension model.

I created my own data from the UCML News Aggregator Dataset in which I parsed the content of the news articles and have created my own labels.

Due to the size of the articles I use TF-IDF to filter out the top 120 words per article and embed those into 300 dimensions.

When I run the CNN I created regardless of the hyper parameters it converges to a small general accuracy, around 38%.

Hyper parameters changed:

Various filter sizes:

I've tried a single filter of 1,2,3 Combinations of filters [3,4,5], [1,3,4]

Learning Rate:

I've varied this from very low to very high, very low doesn't converge to 38% but anything between 0.0001 and 0.4 does.

Batch Size:

Tried many ranges between 5 and 100.

Weight and Bias Initialization:

Set stddev of weights between 0.4 and 0.01. Set bias initial values between 0 and 0.1. Tried using the xavier initializer for the conv2d weights.

Dataset Size:

I have only tried on two partial data sets, one with 15 000 training data, and the other on the 5000 test data. In total I have 263 000 data to train on. There is no accuracy difference whether trained and evaluated on the 15 000 training data or by using the 5000 test data as the training data (to save testing time).

I've run successful classifications on the 15 000 / 5000 split using a feed forward network with a BoW input (93% accurate), TF-IDF with SVM (92%), and TF-IDF with Native Bayes (91.5%). So I don't think it is the data.

What does this imply? Is the model just a poor model for this task? Is there an error in my work?

I feel like my do_eval function is incorrect to evaluate the accuracy / loss over an epoch of the data: