This module contains the bits required to use the fastai DataBlock API and/or mid-level data processing pipelines to organize your data for summarization tasks using architectures like BART and T5.
torch.cuda.set_device(1)
print(f'Using GPU #{torch.cuda.current_device()}: {torch.cuda.get_device_name()}')
Using GPU #1: GeForce GTX 1080 Ti

Summarization tokenization, batch transform, and DataBlock methods

Summarization tasks attempt to generate a human-understandable and sensible representation of a larger body of text (e.g., capture the meaning of a larger document in 1-3 sentences).

path = Path('./')
cnndm_df = pd.read_csv(path/'cnndm_sample.csv'); len(cnndm_df)
1000
cnndm_df.head(2)
article highlights ds_type
0 (CNN) -- Globalization washes like a flood over the world's cultures and economies. Floods can be destructive; however, they can also bring blessings, as the annual floods of the Nile did for ancient Egypt. The world's great universities can be crucial instruments in shaping, in a positive way, humankind's reaction to globalization and the development of humankind itself. Traditionally, universities have been defined and limited by location, creating an academic community and drawing students and scholars to that place. Eventually, some universities began to encourage students to study el... John Sexton: Traditionally, universities have been defined and limited by location .\nGlobal campuses form a network of thought, innovation, he writes .\nFaculty can teach, Sexton says, students can team up in many cities at once .\nSexton: Research, scholarship can be shared and cultural ties made in "century of knowledge" train
1 (CNN) -- Armenian President Robert Kocharian declared a state of emergency Saturday night after a day of clashes between police and protesters, a spokeswoman for the Armenian Foreign Ministry said. Opposition supporters wave an Armenian flag during a protest rally in Yerevan, Armenia, on Saturday. The protesters claim last month's presidential election was rigged. The state of emergency will "hopefully bring some order" to the capital, Yerevan, said Salpi Ghazarian, assistant to the Armenian foreign minister, who spoke to CNN early Sunday. The state of emergency could last until March 20, ... NEW: Protest moves after crackdown at Freedom Square .\nOrder sought after protests over last month's election turn violent .\nDemonstrators say the election was fraudulent .\nState of emergency could last until March 20, official says . train
pretrained_model_name = "facebook/bart-large-cnn"

hf_arch, hf_config, hf_tokenizer, hf_model = BLURR_MODEL_HELPER.get_hf_objects(pretrained_model_name, 
                                                                               model_cls=BartForConditionalGeneration)

hf_arch, type(hf_tokenizer), type(hf_config), type(hf_model)
('bart',
 transformers.tokenization_bart.BartTokenizer,
 transformers.configuration_bart.BartConfig,
 transformers.modeling_bart.BartForConditionalGeneration)

We create a subclass of HF_BeforeBatchTransform for summarization tasks to add decoder_input_ids and labels to our inputs during training, which will in turn allow the huggingface model to calculate the loss for us. See here and here for more information on these additional inputs used in summarization and conversational training tasks. How they should look for particular architectures can be found by looking at those model's forward function's docs (See here for BART for example)

Note also that labels is simply target_ids shifted to the right by one since the task to is to predict the next token based on the current (and all previous) decoder_input_ids.

And lastly, we also update our targets to just be the input_ids of our target sequence so that fastai's Learner.show_results works (again, almost all the fastai bits require returning a single tensor to work).

class HF_SummarizationInput[source]

HF_SummarizationInput(x, **kwargs) :: HF_BaseInput

class HF_SummarizationBeforeBatchTransform[source]

HF_SummarizationBeforeBatchTransform(hf_arch, hf_tokenizer, max_length=None, padding=True, truncation=True, is_split_into_words=False, n_tok_inps=2, ignore_token_id=-100, tok_kwargs={}, **kwargs) :: HF_BeforeBatchTransform

Handles everything you need to assemble a mini-batch of inputs and targets, as well as decode the dictionary produced as a byproduct of the tokenization process in the encodes method.

before_batch_tfm = HF_SummarizationBeforeBatchTransform(hf_arch, hf_tokenizer)
blocks = (HF_Text2TextBlock(before_batch_tfms=before_batch_tfm, input_return_type=HF_SummarizationInput), noop)

dblock = DataBlock(blocks=blocks, 
                   get_x=ColReader('article'), 
                   get_y=ColReader('highlights'), 
                   splitter=RandomSplitter())

Two lines! Notice we pass in noop for our targets (e.g. our summaries) because the batch transform will take care of both out inputs and targets.

 
dls = dblock.dataloaders(cnndm_df, bs=4)
b = dls.one_batch()
len(b), b[0]['input_ids'].shape, b[1].shape
(2, torch.Size([4, 1024]), torch.Size([4, 68]))
t = torch.randn((3,3));

F.pad(t, pad=(1,0), value=1)[:,:-1]
tensor([[ 1.0000, -1.4273, -0.1617],
        [ 1.0000, -0.0583, -1.5524],
        [ 1.0000,  0.8595,  0.9492]])
dls.show_batch(dataloaders=dls, max_n=2, input_trunc_at=1000, target_trunc_at=250)
text target
0 (CNN) -- Home to up to 10 percent of all known species, Mexico is recognized as one of the most biodiverse regions on the planet. The twin threats of climate change and human encroachment on natural environments are, however, threatening the existence of the country's rich wildlife. And there is a great deal to lose. In the United Nations Environment Program (UNEP) World Conservation Monitoring Centre's list of megadiverse countries Mexico ranks 11th. The list represents a group of 17 countries that harbor the majority of the Earth's species and are therefore considered extremely biodiverse. From its coral reefs in the Caribbean Sea to its tropical jungles in Chiapas and the Yucatan peninsula and its deserts and prairies in the north, Mexico boasts an incredibly rich variety of flora and fauna. Some 574 out of 717 reptile species found in Mexico -- the most in any country -- can only be encountered within its borders. It is home to 502 types of mammals, 290 species of birds, 1,150 var Mexico hosts to up to 10 percent of all known species on Earth.\nIt is home to 502 types of mammals, 290 bird species and 26,000 types of plants.\nHuman development and climate change is placing a big strain on its biodiversity.\nThe Golden Eagle is un
1 I have an uncle who has always been a robust and healthy guy. He drank a glass of skim milk every day, bragged about how many pull-ups he was doing and fit into pants he was wearing 20 years before. He didn't take a single medication and retired early. Given that he had no medical problems and ran his own business, he opted to go several years without health insurance. Eventually, when he turned 65, he picked up Medicare. What happened next was a little strange. He fell off the wagon. He exercised only sporadically, and paid hardly any attention to what he was eating. One day, I saw him eat an entire bag of potato chips. He bemoaned the fact that he was forced to buy new, bigger pants, and he stopped drinking his milk. For him, becoming newly insured had nearly the opposite effect on him of what we doctors hope to achieve. He'd become unhealthier. In many ways, my uncle was demonstrating a concept known as the moral hazard. Two economists wrote about this exact scenario in 2006. They Sanjay Gupta: Moral hazard causes some to neglect health when they get health insurance.\nHe says Obamacare alone won't guarantee good health; personal habits must do that.\nHe says research shows 30 minutes of daily exercise cuts heart attack, stroke

Tests

The tests below to ensure the core DataBlock code above works for all pretrained summarization models available in huggingface. These tests are excluded from the CI workflow because of how long they would take to run and the amount of data that would be required to download.

Note: Feel free to modify the code below to test whatever pretrained summarization models you are working with ... and if any of your pretrained summarization models fail, please submit a github issue (or a PR if you'd like to fix it yourself)

BLURR_MODEL_HELPER.get_models(task='ConditionalGeneration')
[transformers.modeling_bart.BartForConditionalGeneration,
 transformers.modeling_blenderbot.BlenderbotForConditionalGeneration,
 transformers.modeling_fsmt.FSMTForConditionalGeneration,
 transformers.modeling_mbart.MBartForConditionalGeneration,
 transformers.modeling_pegasus.PegasusForConditionalGeneration,
 transformers.modeling_prophetnet.ProphetNetForConditionalGeneration,
 transformers.modeling_t5.T5ForConditionalGeneration,
 transformers.modeling_xlm_prophetnet.XLMProphetNetForConditionalGeneration]
pretrained_model_names = [
    ('facebook/bart-base',BartForConditionalGeneration),
    ('t5-small', T5ForConditionalGeneration),
    ('google/pegasus-cnn_dailymail', PegasusForConditionalGeneration)
]
path = Path('./')
cnndm_df = pd.read_csv(path/'cnndm_sample.csv')
#hide_output
task = HF_TASKS_ALL.ConditionalGeneration
bsz = 2
seq_sz = 256
trg_seq_sz = 40

test_results = []
for model_name, model_cls in pretrained_model_names:
    error=None
    
    print(f'=== {model_name} ===\n')
    
    hf_arch, hf_config, hf_tokenizer, hf_model = BLURR_MODEL_HELPER.get_hf_objects(model_name, 
                                                                                   task=task, 
                                                                                   model_cls=model_cls)
    print(f'architecture:\t{hf_arch}\ntokenizer:\t{type(hf_tokenizer).__name__}\n')
    
    before_batch_tfm = HF_SummarizationBeforeBatchTransform(hf_arch, hf_tokenizer, 
                                                            padding='max_length', 
                                                            max_length=[seq_sz, trg_seq_sz])
    
    blocks = (HF_TextBlock(before_batch_tfms=before_batch_tfm, input_return_type=HF_SummarizationInput), noop)
    
    def add_t5_prefix(inp): return f'summarize: {inp}' if (hf_arch == 't5') else inp

    dblock = DataBlock(blocks=blocks, 
                   get_x=Pipeline([ColReader('article'), add_t5_prefix]), 
                   get_y=ColReader('highlights'), 
                   splitter=RandomSplitter())

    dls = dblock.dataloaders(cnndm_df, bs=bsz) 
    b = dls.one_batch()
    
    try:
        print('*** TESTING DataLoaders ***\n')
        test_eq(len(b), 2)
        test_eq(len(b[0]['input_ids']), bsz)
        test_eq(b[0]['input_ids'].shape, torch.Size([bsz, seq_sz]))
        test_eq(len(b[1]), bsz)
        test_eq(b[1].shape, torch.Size([bsz, trg_seq_sz - 1]))

        if (hasattr(hf_tokenizer, 'add_prefix_space')):
            test_eq(dls.before_batch[0].tok_kwargs['add_prefix_space'], True)
            
        test_results.append((hf_arch, type(hf_tokenizer).__name__, model_name, 'PASSED', ''))
        dls.show_batch(dataloaders=dls, max_n=2, input_trunc_at=1000)
        
    except Exception as err:
        test_results.append((hf_arch, type(hf_tokenizer).__name__, model_name, 'FAILED', err))
arch tokenizer model_name result error
0 bart BartTokenizer facebook/bart-base PASSED
1 t5 T5Tokenizer t5-small PASSED
2 pegasus PegasusTokenizer google/pegasus-cnn_dailymail PASSED

Cleanup