Created by: mrfox321

If I repeatedly call the swig decoder ctc_beam_search_decoder, I get a memory leak of around 10MB per iteration.

def score():

    scorer = Scorer(alpha,beta,language_model_path,vocab)
    best_path = ctc_beam_search_decoder(probs_seq2,vocab,beam_size,
                                            cutoff_prob=cutoff_p,cutoff_top_n=cutoff_n,
                                            ext_scoring_func=scorer)
    print best_path[0][1]

for i in range(num_iter):
    score()

Here is the memory profiler output after the first iteration:

Line Mem usage Increment Line Contents ================================================ 21 35.6 MiB 35.6 MiB @profile 22 def score(): 23
24 878.4 MiB 842.7 MiB scorer = Scorer(alpha,beta,language_model_path,vocab) 25 878.4 MiB 0.0 MiB best_path = ctc_beam_search_decoder(probs_seq2,vocab,beam_size, 26 878.4 MiB 0.0 MiB cutoff_prob=cutoff_p,cutoff_top_n=cutoff_n, 27 878.7 MiB 0.4 MiB ext_scoring_func=scorer) 28 878.7 MiB 0.0 MiB print best_path[0][1]

and after N iterations:

Line # Mem usage Increment Line Contents ================================================ 21 1446.4 MiB 1446.4 MiB @profile 22 def score(): 23
24 2061.3 MiB 614.9 MiB scorer = Scorer(alpha,beta,language_model_path,vocab) 25 2061.3 MiB 0.0 MiB best_path = ctc_beam_search_decoder(probs_seq2,vocab,beam_size, 26 2061.3 MiB 0.0 MiB cutoff_prob=cutoff_p,cutoff_top_n=cutoff_n, 27 2061.3 MiB 0.0 MiB ext_scoring_func=scorer) 28 2061.3 MiB 0.0 MiB print best_path[0][1]

it seems that when I change the loop to:

for j in range(100):
    scorer = Scorer(alpha,beta,language_model_path,vocab)
    for i in range(num_iter):
        score()

    del scorer

The memory leak occurs in the outer loop associated with the construction of the scorer. The memory is roughly constant in each inner loop iteration.

Additionally, if I do not run score(), the construction of scorer does not leak.