Note: this code is no longer actively maintained. However, feel free to use the Issues section to discuss the code with other users. Some users have updated this code for newer versions of Tensorflow and Python - see information below and Issues section.
This repository contains code for the ACL 2017 paper Get To The Point: Summarization with Pointer-Generator Networks. For an intuitive overview of the paper, read the blog post.
Looking for test set output?
The test set output of the models described in the paper can be found here.
Looking for pretrained model?
A pretrained model is available here:
(The only difference between these two is the naming of some of the variables in the checkpoint. Tensorflow 1.0 uses lstm_cell/biases and lstm_cell/weights whereas Tensorflow 1.2.1 uses lstm_cell/bias and lstm_cell/kernel).
Note: This pretrained model is not the exact same model that is reported in the paper. That is, it is the same architecture, trained with the same settings, but resulting from a different training run. Consequently this pretrained model has slightly lower ROUGE scores than those reported in the paper. This is probably due to us slightly overfitting to the randomness in our original experiments (in the original experiments we tried various hyperparameter settings and selected the model that performed best). Repeating the experiment once with the same settings did not perform quite as well. Better results might be obtained from further hyperparameter tuning.
Why can't you release the trained model reported in the paper? Due to changes to the code between the original experiments and the time of releasing the code (e.g. TensorFlow version changes, lots of code cleanup), it is not possible to release the original trained model files.
Looking for CNN / Daily Mail data?
Instructions are here.
About this code
This code is based on the TextSum code from Google Brain.
This code was developed for Tensorflow 0.12, but has been updated to run with Tensorflow 1.0. In particular, the code in attention_decoder.py is based on tf.contrib.legacy_seq2seq_attention_decoder, which is now outdated. Tensorflow 1.0's new seq2seq library probably provides a way to do this (as well as beam search) more elegantly and efficiently in the future.
Python 3 version: This code is in Python 2. If you want a Python 3 version, see @becxer's fork.
How to run
Get the dataset
To obtain the CNN / Daily Mail dataset, follow the instructions here. Once finished, you should have chunked datafiles train_000.bin, ..., train_287.bin, val_000.bin, ..., val_013.bin, test_000.bin, ..., test_011.bin (each contains 1000 examples) and a vocabulary file vocab.
Note: If you did this before 7th May 2017, follow the instructions here to correct a bug in the process.
Run training
To train your model, run:
python run_summarization.py --mode=train --data_path=/path/to/chunked/train_* --vocab_path=/path/to/vocab --log_root=/path/to/a/log/directory --exp_name=myexperiment
This will create a subdirectory of your specified log_root called myexperiment where all checkpoints and other data will be saved. Then the model will start training using the train_*.bin files as training data.
Warning: Using default settings as in the above command, both initializing the model and running training iterations will probably be quite slow. To make things faster, try setting the following flags (especially max_enc_steps and max_dec_steps) to something smaller than the defaults specified in run_summarization.py: hidden_dim, emb_dim, batch_size, max_enc_steps, max_dec_steps, vocab_size.
Increasing sequence length during training: Note that to obtain the results described in the paper, we increase the values of max_enc_steps and max_dec_steps in stages throughout training (mostly so we can perform quicker iterations during early stages of training). If you wish to do the same, start with small values of max_enc_steps and max_dec_steps, then interrupt and restart the job with larger values when you want to increase them.
Run (concurrent) eval
You may want to run a concurrent evaluation job, that runs your model on the validation set and logs the loss. To do this, run:
python run_summarization.py --mode=eval --data_path=/path/to/chunked/val_* --vocab_path=/path/to/vocab --log_root=/path/to/a/log/directory --exp_name=myexperiment
Note: you want to run the above command using the same settings you entered for your training job.
Restoring snapshots: The eval job saves a snapshot of the model that scored the lowest loss on the validation data so far. You may want to restore one of these "best models", e.g. if your training job has overfit, or if the training checkpoint has become corrupted by NaN values. To do this, run your train command plus the --restore_best_model=1 flag. This will copy the best model in the eval directory to the train directory. Then run the usual train command again.
Run beam search decoding
To run beam search decoding:
python run_summarization.py --mode=decode --data_path=/path/to/chunked/val_* --vocab_path=/path/to/vocab --log_root=/path/to/a/log/directory --exp_name=myexperiment
Note: you want to run the above command using the same settings you entered for your training job (plus any decode mode specific flags like beam_size).
This will repeatedly load random examples from your specified datafile and generate a summary using beam search. The results will be printed to screen.
Visualize your output: Additionally, the decode job produces a file called attn_vis_data.json. This file provides the data necessary for an in-browser visualization tool that allows you to view the attention distributions projected onto the text. To use the visualizer, follow the instructions here.
If you want to run evaluation on the entire validation or test set and get ROUGE scores, set the flag single_pass=1. This will go through the entire dataset in order, writing the generated summaries to file, and then run evaluation using pyrouge. (Note this will not produce the attn_vis_data.json files for the attention visualizer).
Evaluate with ROUGE
decode.py uses the Python package pyrouge to run ROUGE evaluation. pyrouge provides an easier-to-use interface for the official Perl ROUGE package, which you must install for pyrouge to work. Here are some useful instructions on how to do this:
Note: As of 18th May 2017 the website for the official Perl package appears to be down. Unfortunately you need to download a directory called ROUGE-1.5.5 from there. As an alternative, it seems that you can get that directory from here (however, the version of pyrouge in that repo appears to be outdated, so best to install pyrouge from the official source).
Tensorboard
Run Tensorboard from the experiment directory (in the example above, myexperiment). You should be able to see data from the train and eval runs. If you select "embeddings", you should also see your word embeddings visualized.
Help, I've got NaNs!
For reasons that are difficult to diagnose, NaNs sometimes occur during training, making the loss=NaN and sometimes also corrupting the model checkpoint with NaN values, making it unusable. Here are some suggestions:
- If training stopped with the
Loss is not finite. Stopping.exception, you can just try restarting. It may be that the checkpoint is not corrupted. - You can check if your checkpoint is corrupted by using the
inspect_checkpoint.pyscript. If it says that all values are finite, then your checkpoint is OK and you can try resuming training with it. - The training job is set to keep 3 checkpoints at any one time (see the
max_to_keepvariable inrun_summarization.py). If your newer checkpoint is corrupted, it may be that one of the older ones is not. You can switch to that checkpoint by editing thecheckpointfile inside thetraindirectory. - Alternatively, you can restore a "best model" from the
evaldirectory. See the note Restoring snapshots above. - If you want to try to diagnose the cause of the NaNs, you can run with the
--debug=1flag turned on. This will run Tensorflow Debugger, which checks for NaNs and diagnoses their causes during training.
is not fully defined."">
![[Potential] error in computing attention and final word probability](https://avatars.githubusercontent.com/u/16652882?v=4)
190 Jan 3, 2023
7 Oct 27, 2022
101 Dec 29, 2022
1 Dec 25, 2021
2.1k Jan 1, 2023
119 Jan 2, 2023
541 Nov 27, 2022
753 Dec 23, 2022
217 Dec 12, 2022
1.4k Dec 23, 2022
43 Dec 5, 2022
3.4k Jan 2, 2023
475 Dec 24, 2022
105 Nov 25, 2022
146 Nov 25, 2022
95 Nov 25, 2022
43 Nov 7, 2022