Invalid 'pmf' error in training stage

[herok97]

Hi, i faced some error about 'pmf' in training stage after 0.6M step.

I customized the class 'Cheng2020Anchor' and there is no modification in entropy model.

Is there anyone struggled with this error below?

Traceback (most recent call last):
File "/home/oem/herok97_VML/NNVC/compressai-base/train_mfvc.py", line 547, in
main(sys.argv[1:])
File "/home/oem/herok97_VML/NNVC/compressai-base/train_mfvc.py", line 532, in main
train(
File "/home/oem/herok97_VML/NNVC/compressai-base/train_mfvc.py", line 446, in train
loss = test(global_step, test_dataloader, model, criterion, logger, optimizer.param_groups[0]['lr'],
File "/home/oem/herok97_VML/NNVC/compressai-base/train_mfvc.py", line 352, in test
model.update(force=True)
File "/home/oem/anaconda3/envs/herok97/lib/python3.9/site-packages/compressai/models/google.py", line 299, in update
updated |= super().update(force=force)
File "/home/oem/anaconda3/envs/herok97/lib/python3.9/site-packages/compressai/models/google.py", line 104, in update
rv = m.update(force=force)
File "/home/oem/anaconda3/envs/herok97/lib/python3.9/site-packages/compressai/entropy_models/entropy_models.py", line 426, in update
quantized_cdf = self._pmf_to_cdf(pmf, tail_mass, pmf_length, max_length)
File "/home/oem/anaconda3/envs/herok97/lib/python3.9/site-packages/compressai/entropy_models/entropy_models.py", line 210, in _pmf_to_cdf
_cdf = pmf_to_quantized_cdf(prob, self.entropy_coder_precision)
File "/home/oem/anaconda3/envs/herok97/lib/python3.9/site-packages/compressai/entropy_models/entropy_models.py", line 90, in pmf_to_quantized_cdf
cdf = _pmf_to_quantized_cdf(pmf.tolist(), precision)
ValueError: Invalid pmf: at least one element must have a non-zero probability.

Is it because force=True was set when calling the update() function?

[YodaEmbedding]

What were the changes you made to the code?

[herok97]

I add a CNN module that takes previous decoded latent feature map (this work is for video compression model). This module's output is a additional information for the entropy model's input.

I solved this problem with removing force=True in the update() function. But i don't know what actually happened.

Could you explain what makes this issue?

[liujle]

Hi, I‘m facing the same problem when run update() after training my customized model of 'Cheng2020Anchor', and I tried set force=False but it did not work.
Besides I checked pmf in update()， the upper‘s and lower’s absolute value is large but the difference is small, which caused pmf to be all zeros, is there any way to fix this problem?