meshde / eruditex Goto Github PK

Train and test with the exponential decay function of tensorflow for learning rate in abcnn

Change inputs to a dict and pass inputs as inputs1['word_vectors'] instead of inputs1[0] here

Add feature to shuffle the input data after each epoch in dtrnnTrain.py

There is a bit of redundant code in the functions written in path_utils. We can refactor it to reuse as much code as possible

@mitrjain96 Please change the predictx function in all the SentEmbd models to take as input the actual sentence and not the sequence of word vectors as is being done currently.

Pass word vectors created in IR centroid to subsequent modules to eliminate redundant vectorization.

Calculating gradient of the inner scan loop in DT_RNN gives an error, which I can't seem to fix currently. I'll keep trying to fix it. In the meantime, I guess it'll be better if we start coding the DT-RNN in TensorFlow.

I have written the DT-RNN in numpy, which can be found here

You can refer to that when making the function in TensorFlow

This line uses the python string function split() for tokenization. Using the nltk tokenizer would prove better for this purpose.

@mitrjain96 please correct the cosine_similarity function in Helpers/nn_utils.py to return the following:

T.dot(A,T.transpose(B))/T.sqrt(T.dot(A,T.transpose(A))*T.dot(B,T.transpose(B)))

@virajshastri97 I have made some updates to the file architecture, so please check if our bAbI model and training and testing codes work properly. Also, check that the new state being pickled is pickled under the path states/dmn_basic/file_name. Let me know if you get any errors

Group the whole model in name_scope to use in tensorboard.

calculate and add unknown words' vectors instead of taking random values.