inmoonlight/relation-network

Tensorflow implementation of Relation Network (bAbI dataset)

Hello 안녕하세요

I'm Jihyung Moon, Co-founder & CTO at SoftlyAI

We build AI associates for professionals so that they can focus on bigger problems.
💹 For investors: AILookUp | 🏥 For (digital) healthcare providers: AI Clinic Coordinator

⚡ (Some) Interests

• Building a great AI product
• Effective and efficient AI product development
• Scale

🎙 (Some) Talks

• AI and Marginalized Languages, ICML 2023 Panel Discussion
• KLUE and XTREME, Google Internal Seminar Series: XTREME Talks w/ Kyunghyun Cho, Jihyung Moon and Sungjoon Park
• KLUE: Korean Language Understanding Evaluation, BigScience Episode #2 Invited Talk w/ Kyunghyun Cho, Jihyung Moon and Sungjoon Park

📄 (Some) Publications

• Analyzing Norm Violations in Real-Time Live-Streaming Chat, EMNLP 2023
  • Jihyung Moon*, Dong-Ho Lee*, Hyundong J. Cho, Woojeong Jin, Chan Young Park, Minwoo Kim, Jay Pujara and Sungjoon Park
• KOLD: Korean Offensive Language Dataset, EMNLP 2022
  • Younghoon Jeong, Juhyun Oh, Jaimeen Ahn, Jongwon Lee, Jihyung Moon, Sungjoon Park, Alice Oh
• KLUE: Korean Language Understanding Evaluation, NeurIPS 2021 Datasets and Benchmarks
  • Sungjoon Park*, Jihyung Moon*, Sungdong Kim*, Won Ik Cho* et al
• BEEP! Korean Corpus of Online News Comments for Toxic Speech Detection, SocialNLP@ACL 2020
  • Jihyung Moon*, Won Ik Cho*, and Junbum Lee

relation-network's Issues

ValueError: inputs must not be empty

I executed following command:

$ python preprocessing.py --path tasks_1-20_v1-2/en/
It prints "Masking success!" three times. Then I executed:

$python train.py
It gave me ValueError:

Traceback (most recent call last):
File "train.py", line 181, in
main()
File "train.py", line 109, in main
rn = Model(config)
File "/home/teamai/saurabh_do_not_delete/relational_network/Relation-Network/model.py", line 74, in init
self.pred = self.build(is_train=is_train)
File "/home/teamai/saurabh_do_not_delete/relational_network/Relation-Network/model.py", line 228, in build
embedded_c = self.contextLSTM(self.context, self.label, self.context_real_len, reuse = None)
File "/home/teamai/saurabh_do_not_delete/relational_network/Relation-Network/model.py", line 134, in contextLSTM
s_embedded = sentenceLSTM(sentences, real_lens, reuse = reuse)
File "/home/teamai/saurabh_do_not_delete/relational_network/Relation-Network/model.py", line 112, in sentenceLSTM
outputs, _ = rnn.static_rnn(lstm_cell, s_input, dtype = tf.float32, scope = scope)
File "/home/teamai/py2env/local/lib/python2.7/site-packages/tensorflow/python/ops/rnn.py", line 1162, in static_rnn
raise ValueError("inputs must not be empty")

Can somebody help me to figure out why inputs are comming empty ?

2 embedding matrices

Hey,

why do you have a separate embedding matrix for the questions and contexts? Why not just use the same embedding matrix for each?

low accuracy problem

thank you for your code. I am using the latest Tensorflow version to run your code on babi datasets but I couldn't get an accuracy higher than 50% even on the task 2. I am not sure if it is because of tensorflow version. Did you use parameter settings other than ones in train.py?

Answer vectors are order agnostic

Since in preprocessing.py method _index_answer generates the answer vectors by summing up the one-hot vectors of the answer words for multi-answer questions, there is no information about the order preserved. Isn't this a problem especially for the validity on the path-finding task?

ValueError: inputs must not be empty

Hello I have the same problem with #5 .
But I can not solve it with "--path ./"
"path-where-tasks_1-20_v1-2-located' is for "train.py" located or code folder located
thank you!
我的英语很菜，但愿你可以看懂我在说什么。。。

Does it support training on GPU?

What are self.labels and how do you define your context objects?

What is

self.label = tf.placeholder(
     dtype=tf.float32,
     shape=[self.batch_size, self.c_max_len, self.c_max_len],
     name="label"
)

I'm a bit confused as to why the dimensions are context_length by context_length. A bit of context - I don't understand what you're doing in the following lines:

s_embedded = sentenceLSTM(sentences, real_lens, reuse = reuse)
c_embedded = tf.concat([s_embedded, labels], axis=1)
c_embedded = tf.reshape(c_embedded, shape = [self.batch_size, self.c_max_len, self.c_max_len + self.c_word_embed])
tagged_c_objects = tf.unstack(c_embedded, axis=1)

Could you explain this to me?
Cheers