malllabiisc / hyte Goto Github PK

can you upload the dataset file, thanks!

in time_project.py line 40:
num_batches = len(data) // self.p.batch_size
error: unresolved attribute reference 'p' for class

Thanking for sharing your code!
I'm having a hard time figuring out how you implement the "time-dependent negative sampling(TDNS)" in your paper, can you help me on this? I think the negative sampling process is implemented in line 243-270 of "time_proj.py", isn't this the normal sampling(or "TANS" in your paper)? Or did I miss something?

I couldn't find any code to calculate hit@10 metric for the tail and head prediction in the result_eval file.
So, I wrote a piece of code after the line number 68. Let me know if it is correct or not?

tail_array = np.array(ranks_tail)
head_array = np.array(ranks_head)

hit_at_10_tail = tail_array[np.where( tail_array < 10 ) ]
hit_at_10_head = head_array[np.where( head_array < 10 ) ]

print('Epoch {} : test_tail HIT@10 {}\t test_head HIT@!) {}'.format(k ,len(hit_at_10_tail)/float(len(tail_array))*100, len(hit_at_10_head)/float(len(head_array))*100))

Thanks for your outstanding work, and I want to know if you have the plan to release pytorch version code?

I run the training shell. And my loss down to 0.0. But when it will be stoped? Or Maybe Stop It by myself ???

I have been researched static knowledge graph completion. Now I am going to research TKGC, and I've never seen stat.txt，I'd be grateful if someone could give me a explanation.

Hi,
I see that you are not filtering entities in result_eval.py. Shouldn't you filter out entities already seen?

where can find file processed_data.pkl ?

Hi thanks for the code.
But I found some problematic part in your code

in
time_proj.py
in def load_data()
there is this code snippet

with open(self.p.dataset,'r') as filein:
	for line in filein:
		train_triples.append([int(x.strip()) for x in line.split()[0:3]])
		triple_time[count] = [x.split('-')[0] for x in line.split()[3:5]] 
		count+=1

self.p.dataset means the trainset of data.
and the line

triple_time[count] = [x.split('-')[0] for x in line.split()[3:5]] 

splits time interval part with '-'
but inside YAGO's dataset, there is fact like this

And this fact's start time will be parsed like this

It seems problematic to me...
Isn't it???

hi I want to run this code on my system but the data is too large, how can I execute code on local system?

  python time_proj.py -data_type yago -margin 10 -model MODEL_NAME -test_freq 25 

hello，I cannot find the parameter MODEL_NAME.

While testing, it looks like the model only looks at the start time of an interval. Is that correct?

Hi
I'm curious about this code snippet

for i in range (len(train_triples)-1,-1,-1): # range(start, stop, step)
			if i not in keep_idx:
				del train_triples[i] 
                # if start / end time is like -405 or like 100, delete                
                # But why?

Why do you delete those train facts?