Adding the state-of-the-art contrastive search decoding methods for the codebase of generation_utils.py #19477

gmftbyGMFTBY · 2022-10-11T01:24:22Z

Adding the state-of-the-art contrastive search decoding method for the `generation_utils` codebase

Fixes #19182

In this PR, I add the source codes of our proposed state-of-the-art decoding methods for the off-the-shelf neural text generation models. The main changes are in the following files: (1) src/transformers/generation_utils.py; (2) examples/pytorch/text-generation/run_generation_contrastive_search.py. To run the test script, please follow these commands:

cd examples/pytorch/text-generation;
CUDA_VISIBLE_DEVICES=0 python run_generation_contrastive_search.py --model_type=gpt2 --model_name_or_path=gpt2-large

Before submitting

Did you read the contributor guideline, Pull Request section?
The PR has been well discussed in 19182
Did you make sure to update the documentation with your changes? Here are the
documentation guidelines, and
here are tips on formatting docstrings.
Did you write any new necessary tests? Yes, I have written the test scripts for the contrastive search

Who can review?

According to the suggestions of @gante, @patrickvonplaten and @sgugger can review this PR.

…ation

HuggingFaceDocBuilderDev · 2022-10-11T01:38:40Z

The documentation is not available anymore as the PR was closed or merged.

…formers into csearch-pr-v2

gante · 2022-10-11T17:13:10Z

@sgugger @patrickvonplaten context: this is the implementation by the authors of this NeurIPS paper, as first proposed in #19182 -- a new generation strategy with very interesting results!

sgugger

Thanks a lot for adding this new method. The added code in generate looks clean to me, I just left a couple of nits.

For the example, it would really be good if you could write a new one leveraging the auto-APIs instead of copying the old run_generation.py which is severly outdated and only works for very few models.

src/transformers/generation_utils.py

sgugger · 2022-10-11T19:17:10Z

src/transformers/generation_utils.py

@@ -3446,3 +3766,152 @@ def top_k_top_p_filtering(
        )

    return logits
+
+
+# ========== utils for contrastive search decoding method ========= #


Those utils would probably be best in a submodule (like we have generation_beam_search or generation_beam_constraint) to avoid generation_utils being too big.

src/transformers/generation_utils.py

sgugger · 2022-10-11T19:18:41Z

examples/pytorch/text-generation/run_generation_contrastive_search.py

+        # kwargs["language"] = tokenizer.lang2id[language]
+
+    # TODO fix mask_token_id setup when configurations will be synchronized between models and tokenizers
+    # XLM masked-language modeling (MLM) models need masked token
+    # is_xlm_mlm = "mlm" in args.model_name_or_path
+    # if is_xlm_mlm:
+    #     kwargs["mask_token_id"] = tokenizer.mask_token_id


To clean up?

gmftbyGMFTBY · 2022-10-12T01:33:24Z

Hi, @sgugger, thank you so much for your suggestions. I will fix these problems quickly!

gmftbyGMFTBY · 2022-10-12T03:13:35Z

Hello @sgugger, is there any document or introduction for auto-APIs?

sgugger · 2022-10-12T12:48:37Z

The documentation would be the place to start. You can also look at all other examples!

…e_search.py to the auto-APIs format

gmftbyGMFTBY · 2022-10-12T14:37:19Z

Hello, @sgugger, I have fixed the problems based on your valuable suggestions! Besides, I have updated the test scripts to the auto-APIs of inference.

The command line to run this test script can be found in its docstring.

sgugger

Mostly have nits on the docstrings. The new example looks great, thanks a lot!

docs/source/en/main_classes/text_generation.mdx

src/transformers/generation_contrastive_search.py

src/transformers/generation_utils.py

gmftbyGMFTBY · 2022-10-18T07:47:00Z

Oh, I am still working on the integration test.

gante · 2022-10-18T08:56:33Z

@gmftbyGMFTBY you probably have to add the @slow decorator to the test, and run it locally with RUN_SLOW=1 py.test (...) to confirm that it is working.

Our CI doesn't run tests with @slow on push (and fails if the test doesn't have the decorator and is actually slow), but we run them every 24h and track them internally :)

gmftbyGMFTBY · 2022-10-18T09:31:14Z

Ok, I got it!

gante

Cool, seems ready to go in (except for the minor comment I added)

@patrickvonplaten can you do a final check, please? :)

gante · 2022-10-18T10:09:31Z

src/transformers/generation_utils.py

-            # 10. prepare logits warper: get the TopKLogitsWarper for contrastive_search
-            logits_warper = self._get_logits_warper(
-                top_k=top_k,
-                top_p=top_p,
-                typical_p=typical_p,
-                temperature=temperature,
-                num_beams=num_beams,
-                renormalize_logits=renormalize_logits,
-            )


This shouldn't have been removed :)

During the test, I found that adding the logits_warper (the TopKLogitsWarper is used) could influence the generations of contrastive search. Because the TopKLogitsWarper filters the logits of other tokens (not Top-k tokens) and calculates the softmax, the model confidence is different from the case that the TopKLogitsWarper is not used.

So, in this case, I think the contrastive search should disable the logits_warper by default.

Is there any solution that the TopKLogitsWarper warper is not activated for contrastive search?
For example, edit the _get_logits_warper function and pass the penalty_alpha parameter to indicate whether the top_k parameter is used?

Interesting. In that case, I think we can do without the logits warper for now :)

gante · 2022-10-18T10:09:43Z

src/transformers/generation_utils.py

            return self.contrastive_search(
                input_ids,
                top_k=top_k,
                penalty_alpha=penalty_alpha,
                logits_processor=logits_processor,
-                logits_warper=logits_warper,


(related to the comment above)

patrickvonplaten · 2022-10-18T11:17:59Z

src/transformers/generation_utils.py

+        logits_processor: Optional[LogitsProcessorList] = None,
+        logits_warper: Optional[LogitsProcessorList] = None,
+        stopping_criteria: Optional[StoppingCriteriaList] = None,
+        max_length: Optional[int] = None,


Suggested change

max_length: Optional[int] = None,

patrickvonplaten · 2022-10-18T11:18:09Z

src/transformers/generation_utils.py

+            stopping_criteria (`StoppingCriteriaList`, *optional*):
+                An instance of [`StoppingCriteriaList`]. List of instances of class derived from [`StoppingCriteria`]
+                used to tell if the generation loop should stop.
+            max_length (`int`, *optional*, defaults to 20):


Suggested change

max_length (`int`, *optional*, defaults to 20):

Let's maybe not add a deprecated argument :-)

patrickvonplaten · 2022-10-18T11:18:50Z

src/transformers/generation_utils.py

+        logits_processor = logits_processor if logits_processor is not None else LogitsProcessorList()
+        logits_warper = logits_warper if logits_warper is not None else LogitsProcessorList()
+        stopping_criteria = stopping_criteria if stopping_criteria is not None else StoppingCriteriaList()
+        if max_length is not None:


I think we should remove this if statement

patrickvonplaten · 2022-10-18T11:23:34Z

src/transformers/generation_utils.py

+            )
+            # compute the candidate tokens by the language model and collects their hidden_states
+            output = self(output_hidden_states=True, **next_model_inputs)
+            past_key_values = output.past_key_values


(nit) - not all language gerenation models can return past_key_values (e.g. TransfoXL or XLNet) these models are still surprisingly used a lot:

https://huggingface.co/xlnet-base-cased (> 200k downloads)

Maybe we could add a better error message here?

Suggested change

past_key_values = output.past_key_values

if "past_key_values" not in output:

raise ValueError(f"self.__class__ cannot return `past_key_values` and can therefore **not** be used for contrastive generation.")

past_key_values = output.past_key_values

patrickvonplaten · 2022-10-18T11:25:26Z

src/transformers/generation_utils.py

+                items = []
+                # item is either the key or the value matrix
+                for item in layer:
+                    bsz_and_beam, num_head, seq_len, esz = item.size()


I'm not sure it holds always true that the past_key_values have this size. Did we test contrastive search on all of the following models:

GPT2

T5

GPT-J

BART
?

It should work at least on those 4 models

Yes, GPT2, T5, BART, GPT-J, and OPT models work fine.

Our implementation is compatible with the encoder-decoder models (the degeneration penalty is calculated on the decoder's hidden states). But we didn't carefully conduct the human evaluation of the encoder-decoder models, such as T5 and BART. Whether contrastive search could significantly boost their performance is still an open problem for us.

patrickvonplaten · 2022-10-18T11:26:42Z

tests/generation/test_generation_utils.py

@@ -1693,6 +1693,25 @@ def test_diverse_beam_search(self):
            ],
        )

+    @slow
+    def test_contrastive_search(self):


If possible I'd be really happy if we could also test this on BART, T5 and GPT-J . Then we should have covered 95% of the model architectures. But ok to do in a follow-up PR . Currently I don't expect the method to work with T5.

patrickvonplaten

Ok for me to merge, but it'd be nice to always make sure the method works for all T5, BART, GPT2 and GPT-J. Also, we currently have only slow tests which is dangerous given that changes in generate can also affect contrastive search.

If it doesn't take too much time, I'd advocate to at least add 7 more tests:

4 fast tests with dummy models that just check constrastive search outputs the correct shape, one for each GPT2, T5, BART, GPT-J
3 more slow tests exactly like test_contrastive_search for T5, BART, GPT-J

I leave it up to you @gante to decide :-)

Overall, great work! Thanks a lot @gmftbyGMFTBY for adapting the code so quickly here!

EDIT: Sorry, actually let's please remove the deprecated max_length before merging - that's actually a "must-do" before merging IMO (so not a fully approval here 😅 )

gmftbyGMFTBY · 2022-10-18T11:33:48Z

Okay, I am working on it! Thanks a lot for your reviews!

patrickvonplaten · 2022-10-18T11:39:44Z

BTW @gmftbyGMFTBY,

Just read a through your extremely nice issue! It seems like you experimented with OPT as well, so maybe let's add a test for OPT as well then ? :-) OPT's past_key_values are slightly different compared to GPT2's past_key_values so maybe instead of adding a test for GPT-J and GPT-2, it would make more sense to add a test for OPT in addition to GPT2?

Also, if the paper is only concerned with open-ended generation (so less with encoder-decoder architectures), I'm also totally fine with not testing for T5 and BART (it's a nice to have, but if it takes too much time and it's not too important - happy to skip it!).

Regarding the fast dummy test, could you maybe make use of those dummy models:

The tests colud look very similar to:

transformers/tests/generation/test_generation_utils.py

Line 2053 in 71ca794

def test_max_new_tokens_decoder_only(self):

just much shorter, i.e. they only need to test for shape equality.

gmftbyGMFTBY · 2022-10-18T11:45:08Z

Yeah, we have already tested the OPT models, and it works fine. I will supply more tests to the pre-trained models that you mentioned.

gmftbyGMFTBY · 2022-10-19T05:34:23Z

@patrickvonplaten more tests about these models are added:

gpt2-large
gpt-j (EleutherAI/gpt-j-6B)
opt (facebook/opt-6.7b)
BART (facebook/bart-large-cnn)
T5 (flax-community/t5-base-cnn-dm)

These tests are passed successfully. Can you do the final check about this PR?

gante · 2022-10-19T09:17:37Z

Thank you for being part of this process @gmftbyGMFTBY 🙌 All queries have been addressed and the PR looks in a good state, merging!

gmftbyGMFTBY · 2022-10-19T09:19:27Z

@gante @patrickvonplaten @sgugger Wow, Thank you very much for your help and support. Love huggingface team!

yxuansu · 2022-10-19T10:08:06Z

@gante @patrickvonplaten @sgugger -- Many thanks for your kind help throughout the process! It means a great deal to me and @gmftbyGMFTBY. Huggingface is the best!

patrickvonplaten · 2022-10-20T16:32:20Z

Great work @gmftbyGMFTBY and @yxuansu, thanks for bearing with us through the PR :-)

…he codebase of generation_utils.py (huggingface#19477) * add: the contrastive search for generaton_utils * add: testing scripts for contrastive search under examples/text-generation * update the quality of codes * revise the docstring; make the generation_contrastive_search.py scripts; * revise the examples/pytorch/text-generation/run_generation_contrastive_search.py to the auto-APIs format * revise the necessary documents * fix: revise the docstring of generation_contrastive_search.py * Fix the code indentation * fix: revise the nits and examples in contrastive_search docstring. * fix the copyright * delete generation_contrastive_search.py * revise the logic in contrastive_search * update the intergration test and the docstring * run the tests over * add the slow decorate to the contrastive_search intergrate test * add more test * do the style, quality, consistency checks

gmftbyGMFTBY added 4 commits October 10, 2022 18:53

add: the contrastive search for generaton_utils

942a7c6

Merge branch 'huggingface:main' into csearch-pr-v2

5909423

add: testing scripts for contrastive search under examples/text-gener…

3e71819

…ation

Merge branch 'huggingface:main' into csearch-pr-v2

47b2b3b

gmftbyGMFTBY added 2 commits October 11, 2022 10:41

update the quality of codes

41e37a5

Merge branch 'csearch-pr-v2' of https://github.com/gmftbyGMFTBY/trans…

e278a46

…formers into csearch-pr-v2

gante requested review from sgugger, gante and patrickvonplaten October 11, 2022 17:10

sgugger reviewed Oct 11, 2022

View reviewed changes

Merge branch 'huggingface:main' into csearch-pr-v2

38b100f

gmftbyGMFTBY added 2 commits October 12, 2022 11:40

Merge branch 'huggingface:main' into csearch-pr-v2

9abd1bb

revise the docstring; make the generation_contrastive_search.py scripts;

32e2a30

gmftbyGMFTBY added 4 commits October 12, 2022 21:43

Merge branch 'huggingface:main' into csearch-pr-v2

e9e2b26

Merge branch 'huggingface:main' into csearch-pr-v2

1f1dac2

Merge branch 'huggingface:main' into csearch-pr-v2

6226b9a

revise the examples/pytorch/text-generation/run_generation_contrastiv…

f3bfd87

…e_search.py to the auto-APIs format

gmftbyGMFTBY added 2 commits October 13, 2022 00:03

Merge branch 'huggingface:main' into csearch-pr-v2

d3a91b8

revise the necessary documents

ce26f9f

sgugger reviewed Oct 12, 2022

View reviewed changes

gmftbyGMFTBY added 4 commits October 13, 2022 09:27

Merge branch 'huggingface:main' into csearch-pr-v2

e801c6f

Merge branch 'huggingface:main' into csearch-pr-v2

c78cf91

Merge branch 'huggingface:main' into csearch-pr-v2

fb4174e

fix: revise the docstring of generation_contrastive_search.py

68429ad

gmftbyGMFTBY added 3 commits October 18, 2022 16:22

Merge branch 'huggingface:main' into csearch-pr-v2

b219a17

update the intergration test and the docstring

183d7cc

run the tests over

65a1ebd

gmftbyGMFTBY added 2 commits October 18, 2022 17:34

Merge branch 'huggingface:main' into csearch-pr-v2

4972bfb

add the slow decorate to the contrastive_search intergrate test

e11d342

gante approved these changes Oct 18, 2022

View reviewed changes

patrickvonplaten reviewed Oct 18, 2022

View reviewed changes

patrickvonplaten approved these changes Oct 18, 2022

View reviewed changes

gmftbyGMFTBY added 2 commits October 19, 2022 13:26

Merge branch 'huggingface:main' into csearch-pr-v2

da014bb

add more test

2aa768c

do the style, quality, consistency checks

ced9f70

gante merged commit 71786b1 into huggingface:main Oct 19, 2022

dhruv2601 mentioned this pull request Nov 17, 2022

Contrastive search decoding | Feature request NVIDIA/FasterTransformer#369

Open

Adding the state-of-the-art contrastive search decoding methods for the codebase of generation_utils.py #19477

Adding the state-of-the-art contrastive search decoding methods for the codebase of generation_utils.py #19477

Conversation

gmftbyGMFTBY commented Oct 11, 2022 • edited Loading

Adding the state-of-the-art contrastive search decoding method for the generation_utils codebase

Before submitting

Who can review?

HuggingFaceDocBuilderDev commented Oct 11, 2022 • edited Loading

gante commented Oct 11, 2022 • edited Loading

sgugger left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

gmftbyGMFTBY commented Oct 12, 2022

gmftbyGMFTBY commented Oct 12, 2022

sgugger commented Oct 12, 2022

gmftbyGMFTBY commented Oct 12, 2022

sgugger left a comment

Choose a reason for hiding this comment

gmftbyGMFTBY commented Oct 18, 2022

gante commented Oct 18, 2022 • edited Loading

gmftbyGMFTBY commented Oct 18, 2022

gante left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

gmftbyGMFTBY Oct 18, 2022 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

patrickvonplaten left a comment • edited Loading

Choose a reason for hiding this comment

gmftbyGMFTBY commented Oct 18, 2022

patrickvonplaten commented Oct 18, 2022

gmftbyGMFTBY commented Oct 18, 2022

gmftbyGMFTBY commented Oct 19, 2022 • edited Loading

gante commented Oct 19, 2022

gmftbyGMFTBY commented Oct 19, 2022 • edited Loading

yxuansu commented Oct 19, 2022

patrickvonplaten commented Oct 20, 2022

gmftbyGMFTBY commented Oct 11, 2022 •

edited

Loading

Adding the state-of-the-art contrastive search decoding method for the `generation_utils` codebase

HuggingFaceDocBuilderDev commented Oct 11, 2022 •

edited

Loading

gante commented Oct 11, 2022 •

edited

Loading

gante commented Oct 18, 2022 •

edited

Loading

gmftbyGMFTBY Oct 18, 2022 •

edited

Loading

patrickvonplaten left a comment •

edited

Loading

gmftbyGMFTBY commented Oct 19, 2022 •

edited

Loading

gmftbyGMFTBY commented Oct 19, 2022 •

edited

Loading