devinjake / ns-cqa Goto Github PK

When reimplement baseline CIPITR on webqeustionSP, if setting the question type all, the outer index error for GPU occur:
/pytorch/aten/src/THC/THCTensorIndex.cu:361: void indexSelectLargeIndex(TensorInfo<T, IndexType>, TensorInfo<T, IndexType>, TensorInfo<long, IndexType>, int, int, IndexType, IndexType, long) [with T = float, IndexType = unsigned int, DstDim = 2, SrcDim = 2, IdxDim = -2, IndexIsMajor = true]: block: [7,0,0], thread: [0,0,0] Assertion srcIndex < srcSelectDimSize failed.

Have you ever meet this problem?
Could you please help me out with it?

I read the implementation of NSM in https://github.com/DevinJake/MRL-CQA/blob/master/S2SRL/train_scst_nsm.py. I find that the code is different from the original paper.
In the original paper, the decoder generates tokens using structures like Pointer Network with variable vocabulary provided by Computer module. In your code the decoding vocabulary is fixed. The decoder dose not query KB during the decoding process whereas in the original paper, the decoder executes a query once it is generated and saves results in memory.
It seems that your implementation is a hybrid with NS-CQA generative model structure and NSM training method.
I am not pretty sure. There might be some misunderstandings.
I wonder if the code is different from the NSM paper.

I found the actions (A1~A10 for example) are different between CSQA and WebqustionsSP experiment.
And I am puzzled by this. Can the actions be generally defined?

Thanks for publishing such outstanding work.
I train your model on google colab(gpu Tesla P100 16Gmemory, Xeon [email protected],1core) with the KG server run on local computer (64g memory, i9 12core, flask multi thread).
The training is too slow, about over 24hrs an epoch.
I do not know why this happends.
It will help a lot if you can tell me youre machine configuration

I am interested with your pretraining process but I can not find the code for it.
Could you please provide the code for it?
Since CSQA is a very big dataset, running your work on it take a lot of time.
It will help a lot if you can provide the work on Webquestions SP

I find the some provided datasets have INT at end while some are not.
CSQA_ANNOTATIONS_test.json and CSQA_ANNOTATIONS_test_INT.json for example
What is the difference between these two datasets