yichengong / densely-interactive-inference-network Goto Github PK

I like how other similar projects allow you to auto download datasets simply by running a script, so here you go! GH-4

How can I apply this approach to Quora dataset? I saw the train_quora.py file, but what files are needed for it and how do I generate them?

Does anyone has a pretrain model here? And how to use it?

pa("--filter_heights", type=str, default="5")

why not 3,4,5 ?
Thank you!
@YichenGong

I want to run the script of train_quora.py.But it seem that there is not the module of YF in the floder util

Traceback (most recent call last):
  File "train_quora.py", line 17, in <module>
    from util.YF import YFOptimizer
ImportError: No module named YF

How can I fix this error??

As

have sliced logits already

Everything is OK up to loading shared.jsonl but when data_processing.py tries to load it the following error is raised:

[1] Loading data SNLI
550152it [00:09, 55502.23it/s]
10000it [00:00, 58533.31it/s]
10000it [00:00, 64180.29it/s]
[1] Loading data MNLI
392702it [00:07, 50719.45it/s]
10000it [00:00, 61434.57it/s]
10000it [00:00, 60510.07it/s]
9796it [00:00, 57942.87it/s]
9847it [00:00, 56161.09it/s]
../data/shared.jsonl
Traceback (most recent call last):
  File "train_mnli.py", line 68, in <module>
    shared_content = load_mnli_shared_content()
  File "..../Densely-Interactive-Inference-Network/python/util/data_processing.py", line 173, in load_mnli_shared_content
    assert shared_file_exist
AssertionError

It seems like the data downloader downloads shared.json instead of shared.jsonl while the training script tries to load a .jsonl file.

Everything works fine when downloading shared.jsonl without a script.
Also you could rename shared.json (which redirects to downloading the file) to shared.jsonl in README

Hello,
I am going through your code trying to understand the model.

in get_minibatch function there is a line

premise_vectors = fill_feature_vector_with_cropping_or_padding([dataset[i]['sentence1_binary_parse_index_sequence'][:] for i in indices], premise_pad_crop_pair, 1)

'sentence1_binary_parse_index_sequence' is supposed to be a key in train_snli list of dictionaries.

However, I cannot find where you create this key. Original snli set does not have it.

Regards,

just about the readme section, the result tree data should be

data
├── download.py
├── embeddings
│   └── mnli_emb_snli_embedding.pkl.gz
├── glove.840B.300d.txt
├── glove.840B.300d.zip
├── __MACOSX
│   ├── multinli_0.9
│   └── snli_1.0
├── multinli_0.9
│   ├── Icon\015
│   ├── multinli_0.9_dev_matched.jsonl
│   ├── multinli_0.9_dev_matched.txt
│   ├── multinli_0.9_dev_mismatched.jsonl
│   ├── multinli_0.9_dev_mismatched.txt
│   ├── multinli_0.9_test_matched_unlabeled.jsonl
│   ├── multinli_0.9_test_mismatched_unlabeled.jsonl
│   ├── multinli_0.9_train.jsonl
│   ├── multinli_0.9_train.txt
│   └── paper.pdf
├── multinli_0.9.zip
├── shared.json
├── shared.jsonl
├── snli_1.0
│   ├── Icon\015
│   ├── README.txt
│   ├── snli_1.0_dev.jsonl
│   ├── snli_1.0_dev.txt
│   ├── snli_1.0_test.jsonl
│   ├── snli_1.0_test.txt
│   ├── snli_1.0_train.jsonl
│   └── snli_1.0_train.txt
└── snli_1.0.zip

6 directories, 26 files

There should be a multinli_0.9_test_matched_unlabeled.jsonl and multinli_0.9_test_mismatched_unlabeled.jsonl file to run the entire code.
And also ref to #5 , there should be a shared.jsonl file.

how can I change cpu to gpu to run this script?
thanks.

Can anyone tell me how preprocessing files have been generated?

Thank you!
@YichenGong

At line 32 in the my.tensorflow.nn.py, you used reduce_max. But in your paper, you said using max pooling after CNN to extract character features. Could you please explain it? Thanks!