Understanding true_classes in log_uniform_candidate_sampler - tensorflow

Question

https://www.tensorflow.org/tutorials/text/word2vec uses tf.random.log_uniform_candidate_sampler for negative sampling.
The tutorial sets true_classes to context_class.
My experiment shows no matter what I set for true_classes, the function always yields good results.
> tf.random.log_uniform_candidate_sampler( true_classes=[[1]],
num_true=1, num_sampled=num_ns,
unique=True, range_max=vocab_size)
[0, 1, 7, 5]
> tf.random.log_uniform_candidate_sampler( true_classes=[[2]],
num_true=1, num_sampled=num_ns,
unique=True, range_max=vocab_size)
[0, 6, 2, 5]
What does true_classes mean in this function?

Answer 1

The line in the tutorial:
You can call the function on one skip-grams's target word and pass the
context word as a true class to exclude it from being sampled
That's misleading.
What does true_classes mean in this function?
Function returns true_expected_count which is defined in this line of the source code..
true_classes seems only used to calculate true_expected_count. So this function does not exclude negative classes. Every label has a probability to get sampled.
I copy an example code that can be experimented on (in case something happens to the link), taken from this GitHub issue:
# Do sampling 1000 times using true_classes [0, 8]
sample_func = lambda ii: tf.random.log_uniform_candidate_sampler(true_classes=[[ii]], num_true=1, num_sampled=4, unique=True, range_max=8, seed=42)
dd = {ii : np.stack([sample_func(ii)[0].numpy() for jj in range(1000)]) for ii in range(8)}
# Calculate the distribution in each true_class
for ii in dd:
print("true_class:", ii, ", negative value_counts:", pd.value_counts(dd[ii].flatten()).to_dict())
# true_class: 0 , negative value_counts: {0: 871, 1: 722, 2: 584, 3: 466, 4: 402, 5: 329, 7: 319, 6: 307}
# true_class: 1 , negative value_counts: {0: 867, 1: 695, 2: 571, 3: 485, 4: 411, 5: 380, 6: 316, 7: 275}
# true_class: 2 , negative value_counts: {0: 869, 1: 716, 2: 541, 3: 488, 4: 389, 5: 357, 6: 321, 7: 319}
# true_class: 3 , negative value_counts: {0: 877, 1: 715, 2: 582, 3: 482, 4: 394, 5: 355, 6: 318, 7: 277}
# true_class: 4 , negative value_counts: {0: 883, 1: 716, 2: 566, 3: 489, 4: 394, 5: 367, 6: 316, 7: 269}
# true_class: 5 , negative value_counts: {0: 862, 1: 717, 2: 583, 3: 496, 4: 376, 5: 357, 6: 315, 7: 294}
# true_class: 6 , negative value_counts: {0: 859, 1: 725, 2: 575, 3: 482, 4: 413, 5: 356, 6: 302, 7: 288}
# true_class: 7 , negative value_counts: {0: 880, 1: 724, 2: 555, 3: 488, 4: 425, 5: 324, 7: 302, 6: 302}
# Result of `true_expected_count`
print({ii : np.mean([sample_func(ii)[1].numpy() for jj in range(1000)]) for ii in range(8)})
# {0: 0.99967235, 1: 0.7245632, 2: 0.5737029, 3: 0.47004792, 4: 0.3987442, 5: 0.34728608, 6: 0.3084587, 7: 0.27554017}