shen-lab / tale Goto Github PK

In the attention_layer.py file, where the attention mechanism is implemented, scaling the attention weights is done by q *= depth**0.5 in line 126 (mentioned below).

Shouldn't it be done by division instead of multiplication i.e. q /= depth**0.5 according to the paper?

+q /= depth**0.5
-q *= depth**0.5

Hey guys, this is a very meaningful work and attracts me a lot. Thanks you for sharing your work!

But I met a problem when I tried to reproduce your work. Which is
----No module named 'deepgo_tf'----

I tried "pip install deepgoplus". It works but cannot solve this problem.
Can you tell me how to get "deepgo_tf"? By the way, a more detailed dependence is needed.

Thans a lot again!

Hi, thanks for making your code accessible. I'm trying to adapt my dataset to be able to run with your code. It is not clear to me some files in our data, like how to generate {on}_go_1.pickle, {on}_label_regular_1.npy and {on}_label_matrix_1_sparse.npy. Also, It seems that some files (../Swiss_Prot/seq_{on}.pkl and ../ontology/{on}_go_trunc.pickle) are not available in the repository. Could you help me, please?

Hi,
Thank you for making your code publicly available. I would like to reproduce your results on CAFA3 challenge dataset, but I cannot run the predict.py script because some files are missing there. For example, bp_go_1.pickle, mf_go_1.pickle, cc_go_1.pickle and also some _ic files. Could you please kindly provide them or explain how to generate.
Thanks.

iteration 1/1197 totaltrain_loss:0.695 regular loss 0.017
iteration 2/1197 totaltrain_loss:0.675 regular loss 0.017 
iteration 3/1197 totaltrain_loss:0.651 regular loss 0.018
iteration 4/1197 totaltrain_loss:0.626 regular loss 0.019
iteration 5/1197 totaltrain_loss:0.601 regular loss 0.019
iteration 6/1197 totaltrain_loss:0.575 regular loss 0.019
iteration 7/1197 totaltrain_loss:0.550 regular loss 0.019
iteration 8/1197 totaltrain_loss:0.525 regular loss 0.019
Traceback(most recent call last):
  File"train.py"，line 317，in <module>
    model1.train()
  File"train.py"，line 205，in train
    x= cut(train_x[ite*batch_size: (ite+1)*batch_size]，hparams['MAXLEN'])
  File"train.py"，line 32，in cut
    start = np.random.randint(i-maxlen+1)
  File"mtrand.pyx"，line 743，in numpy.random.mtrand.RandomState.randint
  File"_bounded_integers.pyx"，line 1275，in numpy.random.bounded_integers._rand_int64
  File"_bounded_integers.pyx"，line 687，in numpy.random._bounded_integers._rand_int64_broadcast
ValueError:low >= high

It seems that the "maxlen" in train.py is setted to a wrong value, causing "i-maxlen+1" to be a negative number.
I just tried to run the code in the default way.
My run command is:
python train.py --batch_size 32 --epochs 100 --lr 1e-3 --save_path ./log/ --ontology mf --data_path ../data/ours/ --regular_lambda 0
(By the way, it seems that you need to renew the "Readme.md" because the commend of it is too old to work normally. The folder "../data/" has changed.)
Could you help me solve the BUG? It's quite important for my work.
Thanks, and best wishes for you.