Hey! I am wondering if you could help me to figure something out, in [1] you mentioned that summing up probabilities for Pr, Pr+ and Pr- leads to better results, I tried it in my implementation based on [2], and it does get better, but my probabilities are getting positive (I am working with log probabilities, so the values should be between (-inf, 0]). Did you experienced this phenomenon while implementing the sum in your algorithm?

[1] Stackexchange CTC [2] CTC implementation github

PS: Sorry if this is not the place to make this question, but I have no other way to reach you.

Thanks for your code. I test my sequence with both ctc decoder in tensorflow and this repo.I always get different result. Tensorflow is always right.This repo sometimes return right and sometimes return wrong. Have you ever compare these two implementation?

Hi, did you add word level language model for beam search?

Currently its easy to add character level bi-gram, but I find it much harder to add word level. I tried CTC token passing algorithm but its just way too slow comparing beam search.

I wanted to experiment with the LMs other than Bigram. Any suggestion on how to approach extending current codebase? or the probability of the last character conditioned on all previous character.

Hi! Could you please tell, why does a beam_search with a beam_width equal to 1 not give the same result as best_path?

For example

import numpy as np
from ctc_decoder import best_path, beam_search

chars = 'ab'
mat = np.array([[0.8, 0, 0.2], [0.4, 0.0, 0.6], [0.8, 0, 0.2]])

print(f'Best path: "{best_path(mat, chars)}"')
print(f'Beam search: "{beam_search(mat, chars, beam_width=1)}"')

Gives: Best path: "aa" Beam search: "a"

Thanks!

It's a great job, thanks to the author.

i have a question in beam search +lm ,why last.norm() only use in the last step ? why not use last.norm() in every time step ? the long the seq the lm is small,so it should be compensate by length norm, i think it should be norm every time step ，is it right ? thanks in advance

I got several problems on rather trivial dictionaries and charsets (see below):

• Without pull-request:

Greedy-Decoder:  # 4 3 5 4 2 3 5 6 7 6 7 5 4 5 4 3 5 . 2 3
Token-Passing:   # 4 3 5 4 2 5 6 7 7 5 4 5 4 3 . 2 3

• With pull-request (as expected):

Greedy-Decoder:  # 4 3 5 4 2 3 5 6 7 6 7 5 4 5 4 3 5 . 2 3
Token-Passing:   # 4 3 5 4 2 3 5 6 7 6 7 5 4 5 4 3 5 . 2 3

• Test code (copy the snipped in TokenPassing.py and use all files in the appendix)

with open("charset.txt") as f:
	classes = f.read().split("\n")
with open("dictionary.txt") as f:
	dictionary = f.read().split("\n")
probabilities = np.load("propabilities.npy")

# greedy
from itertools import groupby
best_path = np.argmax(probabilities, axis=1)
blank_idx = 0
best_chars_collapsed = [classes[k] for k, _ in groupby(best_path) if k != blank_idx]
res = ' '.join(best_chars_collapsed)
print("Greedy-Decoder: ", res)

# token passing
tp = ctcTokenPassing(probabilities, classes, dictionary, blankIdx=blank_idx)
print("Token-Passing:  ", tp)

import matplotlib.pyplot as plt
plt.imshow(np.log(probabilities))
plt.show()

The problem is that the current code allows to skip a single output on the first transition (0->1,2) since toks.get(wIdx, s - 1, t - 1) is chosen with does not include the probability at t. Using toks.get(wIdx, s - 1, t) instead fixes this, because this includes the emission probability for blank.

There might be another crucial issue if a there is no blank between two words. I had no time to check this, though.

Appendix:

charset.txt dictionary.txt propabilities.zip

Hi, I use cnn + lstm + ctc in my license plate recognition system。And I use your python scripts to test the precision of "best path decoding" and "beam search decoding"。The result is "best path" got 92.3% while "beam search" got 91.7%。In my experiments, time step equals to 21，and the total num of labels equals to 77。 Could you give me some advice? Thanks a lot!

Hi, Thanks for the sharing. I have a question in language model initialization function. Why do you set numSamples[c] to be len(classes) instead of 0? https://github.com/githubharald/CTCDecoder/blob/d6a88ea5b3187fe5c81508d377b32eb6c860b868/src/LanguageModel.py#L31

I am having a difficult time using the GPU. it runs ok with out GPU but with I get this error

=====Line example (GPU)===== Traceback (most recent call last): File "main.py", line 147, in testLineExampleGPU() File "main.py", line 122, in testLineExampleGPU resBatch = BestPathCL.ctcBestPathCL(batch, classes, clWrapper) File "/home/ubuntu/handwrite/CTCDecoder/src/BestPathCL.py", line 109, in ctcBestPathCL labelStrBatch = clWrapper.compute(batch) File "/home/ubuntu/handwrite/CTCDecoder/src/BestPathCL.py", line 84, in compute cl.enqueue_write_buffer(self.queue, self.batchBuf, batch.astype(np.float32), is_blocking=False) AttributeError: module 'pyopencl' has no attribute 'enqueue_write_buffer'

Hi, I have a question regarding your beam search implementation.

On your ctcBeamSearch method, you put value on blankIdx equals to length of the classes (in this case, the known letters and symbols). But on some other beam-search implementation, they put zero on it.

I tested this using your example, and indeed it differs both in decoded result and how far is it from the ground truth (i'm using CER and WER)

=====Line example (using blankIdx = len(classes))=====
TARGET                  : "the fake friend of the family, like the"
BEAM SEARCH             : "the fak friend of the fomcly hae tC" CER: CER/WER: 0.25714/0.15000
BEAM SEARCH LM          : "the fake friend of the family, lie th" CER: CER/WER: 0.05405/0.03226
=====Line example (using blankIdx=0)=====
TARGET                  : "the fake friend of the family, like the"
BEAM SEARCH             : "the faetker friend of ther foarmnacly,  harse. tHhC." CER: CER/WER: 0.33333/0.22368
BEAM SEARCH LM          : "the fake friend of the family, like the " CER: CER/WER: 0.00000/0.00000

So is there a different case where the blankIdx is not zero? Which value is suitable for beam search decoding?