zyh190507 / knncontrastiveforood

Source code for 'KNN-Contrastive Learning for Out-of-Domain Intent Classofication'.

Dependencies

Use anaconda to create python environemnt:

[conda create --name python=3.6.12]

Install all required libraries:

[pip install -r requirements.txt]

Usage:

Run the experiments (for example banking_25):

[python run_main.py json/banking_25.json]

Results:

[Results stored in model_output, e.g. banaking_0.25.csv]

knncontrastiveforood's Issues

Version not found

torchvision==0.8.2+cu110

This version of torchvision is not available and by using another version I'm not able to train on GPU, getting no CUDA available at the time of torch._C._cuda_init() , even if I have cuda 11.0 installed. Can you help me with this?

ValueError: Sample larger than population or is negative

Thanks for releasing your code. I use another dataset to run your code. I met this error:

trainer.py", line 289, in generate_positive_sample
positive_sample.extend(random.sample(self.negative_data[input_label], positive_num)
raise ValueError("Sample larger than population or is negative")

Can i modify positive_num to one? In your paper, default positive_num is three. Does this modification have a big impact on the results? Looking forward to your reply.

Is the result much worse because of a missing parameter file?

After running, I found that there was a big difference between the results I ran and those provided by the author. Is it because of the lack of these two corresponding hyperparameter files?

sim_embed: str = field(default="/remote-home/dmsong/train_data/adv_learn/counter-fitted-vectors.txt")
cosine_npy: str = field(default="/remote-home/dmsong/train_data/adv_learn/cos_sim_counter_fitting.npy")

question about dataset

Hi, yunhua.

I see that you write ''After datasets are split into train, validation, and test respectively, we randomly sample 25%, 50%, and 75% of the intent classes and discard the remaining classes in the training and validation sets.'' in your paper.
I can see that BANKING and stackoverflow dataset do not have OOD class in validation.

For CLINC-FULL and CLINC-SMALL dataset, do their validation set have OOD class?