salesforce / glad Goto Github PK

I noticed that they had the same input

@vzhong Thank you!

On my local machine, when I run python preprocess_data.py and the script computes word embeddings, it dies with a MemoryError. I've killed all other processes and have about 6GB free RAM, but that doesn't seem to be enough. Is this expected? Anything I can do against it? Perhaps download and use precomputed embeddings from somewhere?

Same issue with stanfordnlp/stanza#16. Any solution?

I have downloaded the clean dstc-2 dataset and trained the system but unable to reproduce the results. Please share the hyper-parameters for the dstc-2 dataset.

It seems that one batch contains batch_size of one-turn-utterance.
@vzhong Thank you!!

I didn't find any code of dstc2.0

I am trying to run the train.py and the embedding I am using is Wikipedia embedding with 50 dimension size in python 3. I am getting this following error.

Namespace(batch_size=50, demb=400, dexp='exp', dhid=200, dout='exp/glad/default', dropout={'emb': 0.2, 'local': 0.2, 'global': 0.2}, epoch=50, gpu=0, lr=0.001, model='glad', nick='default', resume=None, seed=42, stop='joint_goal', test=False)

WARNING:root:loading split train
WARNING:root:loading split dev
WARNING:root:loading split test

INFO:root:dataset sizes: {'dev': 200, 'test': 400, 'train': 600}
INFO:root:loaded model <class 'models.glad.Model'>
INFO:root:saving config to exp/glad/default/config.json

Traceback (most recent call last):
File "train.py", line 66, in
run(args)
File "train.py", line 24, in run
model.load_emb(Eword)
File "/mount/studenten/SpokenLanguageProcessing/2019/WokeSpoke/SLU/glad_master/models/glad.py", line 151, in load_emb
self.emb_fixed.weight.data.copy_(new(Eword))

RuntimeError: The expanded size of the tensor (400) must match the existing size (150) at non-singleton dimension 1. Target sizes: [950, 400]. Tensor sizes: [950, 150]

Hi.

Since the local RNN for a slot type is a bidirectional RNN, should the local self-attention component be initialized as

SelfAttention(2 * dhid, dropout=self.dropout.get('selfattn', 0.))

instead of

SelfAttention(din, dropout=self.dropout.get('selfattn', 0.))

as in the line 101 of glad.py?

Could you please provide the cleaned DSTC2 data? I found this link (mi.eng.cam.ac.uk/~nm480/dstc2-clean.zip) is invalid.