ndcg_array = dcg_array / dcg_max_array

[SaidaSaad]

Hello
i would like to know after we use the command for computing the evaluation scores on prediction files there is any explanation for that ?
I got that erros
RuntimeWarning: invalid value encountered in true_divide
ndcg_array = dcg_array / dcg_max_array
RuntimeWarning: invalid value encountered in true_divide
alpha_ndcg_array = alpha_dcg_array / alpha_dcg_max_array
henlo
henlo
henlo

Thanks

[kenchan0226]

It sounds like some of the values in dcg_max_array are zero or NaN. Can you please wrap the ndcg_array = dcg_array / dcg_max_array line into a try execpt block and print out the values of dcg_max_array? If some of the values are zero, maybe you try to do the division in an element-wise way and set the value to zero if dcg_max_array[i]=0.

[SaidaSaad]

Yes
I printed dcg_max_array and alpha_dcg_max_array and some of the valuse are zeros.
so I edited that that line alpha_ndcg_array = alpha_dcg_array/ alpha_dcg_max_array
to be

alpha_ndcg_array = np.zeros(shape=(2,))
for i in range(2):
if alpha_dcg_max_array[i] == 0 and np.isnan(alpha_dcg_max_array[i]):
alpha_ndcg_array[i] = 0
else:
alpha_ndcg_array[i] = alpha_dcg_array[i] / alpha_dcg_max_array[i]
alpha_ndcg_array = np.nan_to_num(alpha_ndcg_array)

I did also the same for ndcg_max_array , I got no error
but it just print

henlo
henlo
henlo

and nothing else. Did i do something wrong ?

Thanks

[SaidaSaad]

This what i got

[kenchan0226]

Hi, I double-checked the code. The script works even some of the values are zeros, it will just print out a warning, but not an error. The np.nan_to_num(ndcg_array) in the original code have already handled the issue of division by zero. I think you can simply use the original evaluate_prediction.py in this github and try it. It should not print out "henlo" since I cannot find any "henlo" in the entire project.

[SaidaSaad]

Yes , the original code give me the first error i mentioned in the first question as you can see here

[SaidaSaad]

I fixed that by what i told you in the last comment
Please let me know if there is something wrong i did

I also would like to know where i can find the scores for the evaluation ?

Thanks

[kenchan0226]

Yes, I know, but my point is the division by zero is just a warning, but not an error, it will not cause the script to terminate, and the NaN value will be corrected by np.nan_to_num(ndcg_array).
So you don't need to fix the issue of division by zero.
Can you use the original code and use the -exp_path option to specify a path. Run the script, then it will create a results_log_*.txt file in the directory specified by the -exp_path. I am sorry that I did not make it clear in the readme.
If you simply use the default arguments in our readme, it will create a results_log_*.txt file in exp/kp20k.[timestamp] folder.
Actually, the value of the -exp_path argument printed out in your screen will show you path that stores the results file.

[SaidaSaad]

Thank you very much , yes , you are right , those were just warning
Now Yes I got the result file , One more last question :)

I am computing the scores in for word_inspec test set prediction

I used this comman to compute the prediction -pred_file_path pred/predict.kp20k.one2many.cat.copy.bi-directional.20191212-151234/predictions.txt -trg_file_path data/cross_domain_sorted/word_inspec_testing_allkeywords.txt -src_file_path data/cross_domain_sorted/word_inspec_testing_context.txt -exp kp20k -export_filtered_pred -disable_extra_one_word_filter -invalidate_unk -all_ks 5 M -present_ks 5 M -absent_ks 5 M

In this results file,
results_log_5_M_5_M_5_M.txt

Can you please explain it. I saw it dived to all, present, absent and MAE stat. Can you explain for example the one for Present and where exactly the final F1@5, F1@M and alph-nDCG@5 for present keyphrase. And what is Micro and Macro mean.

Thanks I appreciated your help

[kenchan0226]

Hi, the results under ====all==== are the F1 scores for the prediction includes both present and absent keyphrases. The results under ====present==== are the F1 scores for present predicted keyphrases only. The results under ====absent==== are the F1 scores for absent predicted keyphreases only. The results under ====MAE==== are the results about the ability of the model to predict the correct number of keyphreases.

[SaidaSaad]

Yes I understood that but what I did not understand is that for example for present keyphrases I found two F1@5
F1@5=0.16388 (Micro) and F1@5=0.18448(Macro)
what is the difference and which one you used to evaluate

==================================present====================================
#predictions after filtering: 2375 #predictions after filtering per src:4.750
#unique targets: 3602 #unique targets per src:7.204
Begin===============classification metrics present@5===============Begin
#target: 2500, #predictions: 3602, #corrects: 500
Micro: P@5=0.2 R@5=0.13881 F1@5=0.16388
Macro: P@5=0.2 R@5=0.1712 F1@5=0.18448
Begin===============classification metrics present@M===============Begin
#target: 2375, #predictions: 3602, #corrects: 511
Micro: P@M=0.21516 R@M=0.14187 F1@M=0.17099
Macro: P@M=0.27203 R@M=0.17369 F1@M=0.21201
Begin==================Ranking metrics present@5==================Begin
MAP@5=0.13014 NDCG@5=0.49478 AlphaNDCG@5=0.7968
Begin==================Ranking metrics present@M==================Begin
MAP@M=0.13103 NDCG@M=0.49963 AlphaNDCG@M=0.82678

[kenchan0226]

Hi, we report the macro F1 scores in our paper since they are used by previous keyphrase generation literature. Macro F1 scores and micro F1 scores are two different ways to aggregate the F1 score of each individual test sample into one score. You can check the following two urls for the details.

http://rushdishams.blogspot.com/2011/08/micro-and-macro-average-of-precision.html
https://datascience.stackexchange.com/questions/15989/micro-average-vs-macro-average-performance-in-a-multiclass-classification-settin

[SaidaSaad]

Hello Kenchan

I have another question,I would to ask when we are the prediction using this command 👍

catSeq on inspec dataset: python3 interactive_predict.py -vocab data/kp20k_sorted/ -src_file data/cross_domain_sorted/word_inspec_testing_context.txt -pred_path pred/%s.%s -copy_attention -one2many -one2many_mode 1 -model [path_to_model] -max_length 60 -remove_title_eos -n_best 1 -max_eos_per_output_seq 1 -beam_size 1 -batch_size 20 -replace_unk

why we shall pass the dateset itself -vocab data/kp20k_sorted/ in that command . I think it should be enough to pass the model and the Test data set only , So Could you please explain me more why it is necessary?

[kenchan0226]

Our source code separately save the word2idx and idx2word dictionary in the vocab.pt, they are not inside our saved model, so we still need to load it.

[SaidaSaad]

Thanks for your reply , I would like to know if your source code is saving word2idx and idx2word during the training of the model?
Another question
In case if i got a model which trained on part of the dataset , Can I still be able use the same command to get the predictions?
Thank you :)

[kenchan0226]

word2idx and idx2word are saved by the preprocessing scripts.
I think you can still use the same command to get the predictions.

[Struggle-lsl]

predict.txt can you run