Unconditional Summarization Evaluation

[Shashi456]

I have the interaction summarisation setup done, but could you guide me towards how I can replicate the unconditional summarisation results in the paper for cnn-dailymail especially the evaluation part?

[jxhe]

Hi, you need to first train the keyword tagger and generate unconditional summaries. You can follow the README on "train the keyword tagger" and "evaluate CTRLsum" to replicate the results. Please let us know if you encounter any further issues.

[Shashi456]

Thanks for the response, I'll try to follow those steps, please allow me to keep the issue alive for a little longer

[Shashi456]

when you say run this command for evaluation, bash scripts/test_bart.sh -g [GPUs] -s [source file name, NOT full path] -d [dataset] -p [ctrlsum checkpoint directory] could you mention what exactly you mean by the source file, would it be test.source in the following case?

bash scripts/test_bart.sh -g 1 -s test.source -d cnndm -p ../cnndm_ctrlsum

[jxhe]

The source file is the actual input to CTRLsum, for example, it would be test.predwordsource for unconditional summarization, and test.oraclenssource would produce the oracle performance.

[Shashi456]

I'm off by a couple points when I ran the numbers,

1 ROUGE-1 Average_P: 0.36141 (95%-conf.int. 0.35914 - 0.36381)
1 ROUGE-1 Average_F: 0.43591 (95%-conf.int. 0.43398 - 0.43800)
---------------------------------------------
1 ROUGE-2 Average_R: 0.29378 (95%-conf.int. 0.29083 - 0.29666)
1 ROUGE-2 Average_P: 0.17329 (95%-conf.int. 0.17135 - 0.17529)
1 ROUGE-2 Average_F: 0.21002 (95%-conf.int. 0.20798 - 0.21212)
---------------------------------------------
1 ROUGE-L Average_R: 0.56304 (95%-conf.int. 0.56008 - 0.56615)
1 ROUGE-L Average_P: 0.33580 (95%-conf.int. 0.33365 - 0.33806)
1 ROUGE-L Average_F: 0.40557 (95%-conf.int. 0.40362 - 0.40768)

I'll recheck my process, but I did most of the stuff correctly.

[jxhe]

Hi, can you check this thread to see if there is any helpful information there? One important point is that we used the tagger checkpoint with the best validation loss instead of the last checkpoint (because of overfitting).

I can try to help debug if you post your tagger training log here. I am also glad to share our pretrained tagger if you contact me through email: [email protected]

[Shashi456]

Hello @jxhe, Could you tell me what was the compute you used to train the model? Im trying to replicate the actual training too (but Im afraid I might have a smaller GPU) so i just want to know what i would need

[jxhe]

Hi, we used 8 16G v100 GPUs to train, which takes 1-2 days on the CNNDM dataset

[Shashi456]

@jxhe Do you have any suggestions if I'm trying to make this work on a GPU with 12GB vRAM

[jxhe]

You can play with the max_tokens and update_freq variables in the training script to match our effective batch size:

https://github.com/salesforce/ctrl-sum/blob/b9afc42be504f55795b0c3b3606163d77a7a852c/scripts/train_bart.sh#L17

If you want to train this on one GPU, then you may need to set update_freq 8x larger like 64 to match 8-gpu batch size; if max_tokens=1024 results in an out-of-memory error in your 12GB VRAM, you may need to set that smaller like 512 and further increase the update_freq, for example, max_tokens=512, update_freq=128, but this would take a long time to train

[Shashi456]

thanks a lot @jxhe for the tips :)

[Shashi456]

hello, @jxhe

2021-07-24 12:36:18 | WARNING | fairseq.data.data_utils | 232243 samples have invalid sizes and will be skipped, max_positions=(512, 512), first few sample ids=[99630, 150074, 103313, 240036, 226747, 108275, 29995, 138361, 64376, 130301]

I keep getting this error, I understand that if i set the max_tokens=512, then i'll probably have to decrease the max_position=512 in the preprocessing script as well, but that didnt exactly solve the problem, do you have any idea?

and since a tonne of these examples are being skipped, the data loader is emptier than what is expected which results in

2021-07-24 13:17:12 | INFO | fairseq.data.iterators | Data loading buffer is empty or nearly empty. This may indicate a data loading bottleneck, and increasing the number of workers (--num-workers) may help.

Sorry to fall back on you for every issue.

[jxhe]

Hi, I am not sure why this happens, have you turned on --truncate-source when training the model? Can you share your training log which would be better to debug?