tanyuqian / learning-data-manipulation Goto Github PK

I want to know that the method update DM grad every batch by running on validation set once. If the size of validation set is big, the time cost so much, any method could deal with it?

Thank you very much

what the function do? why define a magic_model?
classifier.finetune_generator(example, aug_probs, finetune_batch_size=len(examples))

Hi, I think here in augmentation_main.py

Hi !
The loop in run_augmentation_sst5_low.sh
https://github.com/tanyuqian/learning-data-manipulation/blob/master/scripts/run_augmentation_sst5_low.sh#L1-L14
keeps args.data_seed=159, which means the train/valid/test examples are all the same during this loop, right? I want to know whether the mean/variance of the results reported in your paper is only for model training, not data under different random seeds?

Hello, there
Thanks a lot for your opening source your code and your impressive work.
I met up with a little problem when I was trying to reproduce the data augmentation part for text in scripts folder providered by you. (specifically, say, I want to run the script : run_augmentation_sst5_low.sh)
It seems that you set the data augmentation experience running on a single GPU So I just run your code without changing anything on a RTX2080 GPU(around 10GiB storage space available) and here is a 'CUDA out of memory' error occurring when I finished the phases of 'Classifier Pre-training' and 'Generator Pre-training' and just started the phase of 'Training'.

Could you plz imform me of the rough memory requirement of gpu towards running this part? And whether you were making use of multi-gpus to run this part or not? (I noticed that you metioned that '4*RTX1080Ti GPUs and 64GBRAM' in your paper, does 'RAM' here mean storage of GPU?)

I'd appreciate it if receiving a reply!

I configured the same environment and ran the run_augmentation_sst5_low.sh file and got 15 identical results. The result is "Final Test Acc: 28.6425". Is there some error in the code?

Thanks for your attention.
I have reproduced the results on SST-5(36.71), IMDB(65.13) and TREC(85.13/82.50). The results are the average of final test acc on 15 runs.
The SST-5 and IMDB results are fine, but the results on TREC is significantly lower than 89.15 on your paper.
Are the hyperparameters on TREC different from SST-5 and IMDB?

Thanks for your great jobs on data augmentation.
I am confused by the code in augmentation/classifier.py def finetune_generator(self, example, aug_probs, finetune_batch_size):
What's the correlation between the reward (Eq.9 in the paper) and the aforementioned code?

And. what is the purpose of the magic_module.py

how to understand magic_model

why augmentation.classifier line 158 dev_loss.backward() can update the Generator weight

When I run weighting_imb01.sh this error occurs. Is there anyone know how to fix this?

I find the magicmodule can's process RNNBased Module,
such as :meta_model = MagicModule(nn.RNN(10,10))
The errors "AttributeError: can't set attribute " will appear, could you help me find out what's wrong with it?