Dataset too large

[asahi417]

Dataset too large

[asahi417]

Would it possible to share the sentence that causes the error? I can run it locally to see if I can reproduce the error.

[NHendrickson9616]

It is not a sentence, it is a a list of lines from .txt files, so it is essentially a lot of sentences in a list and I want to evaluate them all

[NHendrickson9616]

And unfortunately I cannot share the dataset anywhere because it has PHI in it

[asahi417]

Ok. Probably, something to do with the sequence length. Which model are you using to compute perplexity? Also, could you tell me the number of the maximum character in a single line in your file?

[NHendrickson9616]

I am fairly sure that it is 512, I think I need help on how to batch my dataset

[asahi417]

If you run it on a single instance, instead of passing a list, you should be able to find the line that causes the error.

text =  dsMap['test']['text']
for t in text:
    ppl = scorer.get_perplexity([t])

[NHendrickson9616]

I am running a version of BERT, but the line that caused issues has only 129 characters in it

[asahi417]

Does the sentence contain only roman characters?

[NHendrickson9616]

It has dashes, commas, a period, and a colon, I don't know if those count, but other than that, yes

[NHendrickson9616]

I have pretrained BERT on this same dataset before with the same tokenizer and model.

[asahi417]

Could you try to compute perplexity on the same files but with roberta-base? It might be BERT specific issue.

[NHendrickson9616]

Same error, same spot

[NHendrickson9616]

But it did take it a little longer to get there although that may be because it was loading a new model in

[asahi417]

So RoBERTa and BERT raises the same error in the same line, correct?

[NHendrickson9616]

Correct.

[asahi417]

If you could mask the alphabet in the sentence with a random character, would it be possible to share here? For instance, if the sentence is @lmppl I have, to say this √√ is &&private 1234---, then you could convert it into @aaaa a aaaa, aa aaa aaaa √√ aa &&aaaaaaa aaaa---. Here, I replaced all the alphabet by a. This way, there's no way I can restore the original sentence, but there's a high chance that you would see same issue there.

[asahi417]

I guess alphabet is not the root cause of the error, and that's why to debug the issue, they are not imporatn.

[NHendrickson9616]

Here:
Aaaaaaaaa: Aaaaaaaaa aaaaaaa aaaa aaaaaaaa - AAA10-AA A19.94, Aaaaaaaaa aaaaa - AAA10-AA A19.10, Aaaaaaa aaaaaaaaaa - AAA10-

[NHendrickson9616]

I can confirm that in my code, running just that sentence, does reproduce the same error, so if that doesn't work for you, it may be that I have accidentally edited the code aside from replacing the transformer.

[asahi417]

It's working fine with the latest lmppl (0.3.1) indeed as below.

In [1] from lmppl import LM
In [2] lm = LM('roberta-base')
In [3] text = "Aaaaaaaaa: Aaaaaaaaa aaaaaaa aaaa aaaaaaaa - AAA10-AA A19.94, Aaaaaaaaa aaaaa - AAA10-AA A19.10, Aaaaaaa aaaaaaaaaa - AAA10-"
In [4] lm.get_perplexity(text)
Out[4]: 115701099.01106478

[NHendrickson9616]

So recopied the run_mlm.py and then ran it as is with roberta-base and here is what I got:

RuntimeError                              Traceback (most recent call last)
Cell In[115], line 5
      3 scorer = MaskedLM('roberta-base')
      4 text = dsMap['test']['text']
----> 5 ppl = scorer.get_perplexity("Aaaaaaaaa:  Aaaaaaaaa aaaaaaa aaaa aaaaaaaa - AAA10-AA A19.94, Aaaaaaaaa aaaaa - AAA10-AA A19.10, Aaaaaaa aaaaaaaaaa - AAA10-")
      6 # for t in text:
      7 #     ppl = scorer.get_perplexity([t])
      8 #     print(t)
      9 #ppl = scorer.get_perplexity(text, batch=32)
     10 print(ppl)

Cell In[113], line 155, in MaskedLM.get_perplexity(self, input_texts, batch)
    153 for s, e in tqdm(batch_id):
    154     _encode = data[s:e]
--> 155     _encode = {k: torch.cat([o[k] for o in _encode], dim=0).to(self.device) for k in _encode[0].keys()}
    156     labels = _encode.pop('labels')
    157     output = self.model(**_encode, return_dict=True)

Cell In[113], line 155, in <dictcomp>(.0)
    153 for s, e in tqdm(batch_id):
    154     _encode = data[s:e]
--> 155     _encode = {k: torch.cat([o[k] for o in _encode], dim=0).to(self.device) for k in _encode[0].keys()}
    156     labels = _encode.pop('labels')
    157     output = self.model(**_encode, return_dict=True)

RuntimeError: Sizes of tensors must match except in dimension 0. Expected size 54 but got size 53 for tensor number 5 in the list.

[NHendrickson9616]

I just ran your exact code and that worked, can I send you my version of the code? (run_mlm.py I mean)

[NHendrickson9616]

Ok so with a little bit of modification, I can run my whole dataset on the code that you sent, now I just need to figure out how to be able to modify that to have a different transformer without breaking it I think.

[NHendrickson9616]

Although I do wonder why you chose to use the version for GPT variants rather than the one for BERT variants in your example

[NHendrickson9616]

If I import and use MaskedLM rather than LM, it breaks, not sure why though.

[asahi417]

Could you try following instead?

scorer.get_perplexity(["Aaaaaaaaa: Aaaaaaaaa aaaaaaa aaaa aaaaaaaa - AAA10-AA A19.94, Aaaaaaaaa aaaaa - AAA10-AA A19.10, Aaaaaaa aaaaaaaaaa - AAA10-"])

[asahi417]

Ah wait, you're right. I should have use MaskedLM, but not LM.

[asahi417]

Yeah, it's working without any issue.

In [1]: from lmppl import MaskedLM

In [2]: text = "Aaaaaaaaa: Aaaaaaaaa aaaaaaa aaaa aaaaaaaa - AAA10-AA A19.94, Aaaaaaaaa aaaaa - AAA10-AA A19.10, Aaaaaaa aaaaaaaaaa - AAA10-"

In [3]: lm = MaskedLM('roberta-base')

In [4]: lm.get_perplexity(text)
100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1/1 [00:02<00:00,  2.13s/it]
Out[4]: 2.676919185346931

[NHendrickson9616]

Ok, so that simple example works, but then when I tried to loop through it like so:

from lmppl import MaskedLM
lm = MaskedLM('/path/to/local/model/')
text = dsMap['test']['text']
num = 0
sum1 = 0
for t in text:
    sum1 = lm.get_perplexity(t) + sum1
    num = num + 1
print(sum1/num)

it returns another tensor error on the very same text that we were just testing. It gets past the first two loops but then breaks on the third which is what we have been working on:

RuntimeError                              Traceback (most recent call last)
Cell In[14], line 8
      6 sum1 = 0
      7 for t in text:
----> 8     sum1 = lm.get_perplexity(t) + sum1
      9     num = num + 1
     10 print(sum1/num)

File ~/.conda/envs/BioBERTUAB/lib/python3.10/site-packages/lmppl/ppl_mlm.py:154, in MaskedLM.get_perplexity(self, input_texts, batch)
    152 for s, e in tqdm(batch_id):
    153     _encode = data[s:e]
--> 154     _encode = {k: torch.cat([o[k] for o in _encode], dim=0).to(self.device) for k in _encode[0].keys()}
    155     labels = _encode.pop('labels')
    156     output = self.model(**_encode, return_dict=True)

File ~/.conda/envs/BioBERTUAB/lib/python3.10/site-packages/lmppl/ppl_mlm.py:154, in <dictcomp>(.0)
    152 for s, e in tqdm(batch_id):
    153     _encode = data[s:e]
--> 154     _encode = {k: torch.cat([o[k] for o in _encode], dim=0).to(self.device) for k in _encode[0].keys()}
    155     labels = _encode.pop('labels')
    156     output = self.model(**_encode, return_dict=True)

RuntimeError: Sizes of tensors must match except in dimension 0. Expected size 13 but got size 12 for tensor number 1 in the list.

[NHendrickson9616]

It does not return any errors on the same exact code but with LM, but it also returns a perplexity number that is wildly incorrect because it isn't the right type of eval for the model