dinghanshen/swem

The Tensorflow code for this ACL 2018 paper: "Baseline Needs More Love: On Simple Word-Embedding-Based Models and Associated Pooling Mechanisms"

swem's Issues

Difference between the paper and the implementation

Thanks sharing us the code.
I have question regarding DBPedia experiment code.

https://github.com/dinghanshen/SWEM/blob/master/eval_dbpedia_emb.py#L96-L118

It seems to me that you are applying max-pooling (tf.nn.max_pool) then another max pooling while your paper says that you applied average-pooling then max pooling.
Also, pooling window size (max_len option) seems to be very large, it seems to me like you are just applying max pooling rather than hierarchical pooling.

Is this intended implementation? Or am I missing somthing?

Not find SWEM-hier

Hi, seems not find hier encoder as paper mentioned. Very interested to see it :)

aver_emb_encoder

I don't think aver_emb_encoder is actually ever defined in the committed code base--though it's pretty straightforward to understand what it should be doing :-).

executing requirement.txt leaves you with a CPU backed tensorflow

Just a small observation to point out that the requirement.txt when executed on virtual environment installs a CPU backed tensorflow instead of a GPU one.
Better, change it to tensorflow-gpu==1.5 or point out that this give you a CPU tensorflow.

Thanks !!

Package dependency related errors occuring while loading snli_emb.py

While loading the snli_emb.py with current numpy==1.16.* the following error came:

  File "/usr/local/lib/python2.7/dist-packages/numpy/lib/npyio.py", line 451, in load
    raise ValueError("Cannot load file containing pickled data "
ValueError: Cannot load file containing pickled data when allow_pickle=False

When, I downgraded it to numpy==1.14.* the error related to binary compatibility came:

ValueError: numpy.ufunc size changed, may indicate binary incompatibility. Expected 216 from C header, got 192 from PyObject

To me it looks like a package version compatibility related issue with respect to numpy used for creating snli_emb.p pickle file.

It'll be really helpful if you can mention the package versions also that you have used in as your import statements like for gensim, scipy etc packages.

Experiment Setting on Section 4.1.1 in the paper

It seems that the repo do not contain code about the section 4.1.1: Interpreting model predictions, but I am really interested in the part, could you share it if you have time.

Also, I have some questions on the experiment setting of the section.
(1) Do you set the word embedding randomly initialized(as you mention in the section) but not from Glove pretrain.
(2) What is your network architecture, Is Word_embedding + MaxPooling +Classifier or Word_Embedding + MaxPooling + MLP + Classifier.

Why simple average pooling of w2v should be better than LSTM and some CNNs? Unbelievable!

1.18.2 I used the Mob Capture Device by tfarecnim to get my horse out of the water. Now its model is frozen.

I was writing this before I realized that jumping solves the issue. Resets the model or something, I guess. Is it okay if I post this anyway in case other people are having the same problem? If not, where should I post it instead?

how can SWEM-max train a wordEmbedding

i think the SWEM-max is a model that captures the words' features to form a sentence embedding , how can it train a word embedding model?

Why simple average pooling of w2v can

requirement missing gensim matplotlib

As titled

where can i get the Yelp P. and Yelp F. dataset?

can u give me a website to download the dataset this paper has been used?

About NASH

 I am a first-year doctoral student at Beijing University of Posts and Telecommunications,China, and I major in computer science and technology.I have recently read your paper"NASH: Toward End-to-End Neural Architecture for Generative Semantic Hashing" .It would make a really positive contribution to my work. I am wondering if you could kindly send me the source program of your experiments. I promise they will be used only for research purposed.

Floating hitching post and riding horses freezes and crashes the game.