--- title: modeling.seq2seq.summarization keywords: fastai sidebar: home_sidebar summary: "This module contains custom models, custom splitters, etc... summarization tasks." description: "This module contains custom models, custom splitters, etc... summarization tasks." nb_path: "nbs/02zc_modeling-seq2seq-summarization.ipynb" ---
{% raw %}
{% endraw %} {% raw %}
 
{% endraw %} {% raw %}
{% endraw %} {% raw %}
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
{% endraw %}

Summarization

The objective of summarization is to generate a concise and accurate representation of a much larger body of text. For example, we may want to summarize an article in a single sentence.

Prepare the data

{% raw %}
path = Path('./')
cnndm_df = pd.read_csv(path/'cnndm_sample.csv'); len(cnndm_df)
1000
{% endraw %} {% raw %}
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
{% endraw %} {% raw %}
pretrained_model_name = "facebook/bart-large-cnn"
hf_arch, hf_config, hf_tokenizer, hf_model = BLURR.get_hf_objects(pretrained_model_name, 
                                                                  model_cls=BartForConditionalGeneration)

hf_arch, type(hf_config), type(hf_tokenizer), type(hf_model)
('bart',
 transformers.models.bart.configuration_bart.BartConfig,
 transformers.models.bart.tokenization_bart_fast.BartTokenizerFast,
 transformers.models.bart.modeling_bart.BartForConditionalGeneration)
{% endraw %} {% raw %}
text_gen_kwargs = {}
if (hf_arch in ['bart', 't5']):
    text_gen_kwargs = {**hf_config.task_specific_params['summarization'], **{'max_length': 30, 'min_length': 10}}

# not all "summarization" parameters are for the model.generate method ... remove them here
generate_func_args = list(inspect.signature(hf_model.generate).parameters.keys())
for k in text_gen_kwargs.copy():
    if k not in generate_func_args: del text_gen_kwargs[k]

if (hf_arch == 'mbart'):
    text_gen_kwargs['decoder_start_token_id'] = hf_tokenizer.get_vocab()["en_XX"]
{% endraw %} {% raw %}
tok_kwargs = {}
if (hf_arch == 'mbart'):
    tok_kwargs['src_lang'], tok_kwargs['tgt_lang'] = "en_XX", "en_XX"
{% endraw %} {% raw %}
before_batch_tfm = HF_Seq2SeqBeforeBatchTransform(hf_arch, hf_config, hf_tokenizer, hf_model, 
                                                  max_length=256, max_target_length=130,
                                                  tok_kwargs=tok_kwargs, text_gen_kwargs=text_gen_kwargs)

blocks = (HF_Seq2SeqBlock(before_batch_tfm=before_batch_tfm), noop)

dblock = DataBlock(blocks=blocks, 
                   get_x=ColReader('article'), 
                   get_y=ColReader('highlights'), 
                   splitter=RandomSplitter())
{% endraw %} {% raw %}
dls = dblock.dataloaders(cnndm_df, bs=2)
{% endraw %} {% raw %}
b = dls.one_batch()
{% endraw %} {% raw %}
len(b), b[0]['input_ids'].shape, b[1].shape
(2, torch.Size([2, 256]), torch.Size([2, 79]))
{% endraw %} {% raw %}
dls.show_batch(dataloaders=dls, max_n=2)
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 varieties of birds and 26,000 classifications of plants. Pronatura, a non-profit organization that works to promote conservation and sustainable development in Mexico, has selected six species which it says symbolize the problems faced by the 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 under threat in spite of being the country's national symbol.
1 (CNN) -- Five former Blackwater Worldwide security guards indicted on voluntary manslaughter and other charges in connection with killings in Iraq were released on their own recognizance Monday after a court hearing. Dustin Heard, an ex-security guard from Blackwater Worldwide, surrenders Monday in Salt Lake City, Utah. The charges, which the Justice Department announced Monday, stem from a September 16, 2007, shooting that killed 17 Iraqis in a Baghdad square. A judge ordered the ex-guards to appear for another court hearing in Washington on January 6. The 35-count indictment charges each of the former guards with 14 counts of manslaughter, 20 counts of attempted manslaughter and one count of using a firearm in the commission of a violent crime. Prosecutors do not believe they have enough evidence to bring manslaughter charges on three of the 17 deaths. If convicted, the defendants would face a maximum sentence of 10 years in prison for each manslaughter count, seven years in prison for each count of attempted manslaughter and a 30-year mandatory minimum sentence for the firearms charge. Watch victims speak of the shooting ยป. The men named in the indictment are Donald Ball, 26, of West Valley City, Utah; Dustin Heard, 27, of Maryville, Tennessee; Evan Liberty, 26, of Rochester A 35-count indictment charges each of five former guards.\nEx-guards from Blackwater Worldwide surrender to authorities in Utah.\nA sixth Blackwater guard has pleaded guilty to voluntary manslaughter.\nEx-guards are charged in 2007 shootings that killed 17 Iraqis.
{% endraw %}

Train model

{% raw %}
seq2seq_metrics = {
        'rouge': {
            'compute_kwargs': { 'rouge_types': ["rouge1", "rouge2", "rougeL"], 'use_stemmer': True },
            'returns': ["rouge1", "rouge2", "rougeL"]
        },
        'bertscore': {
            'compute_kwargs': { 'lang': 'en' },
            'returns': ["precision", "recall", "f1"]
        }
    }
{% endraw %} {% raw %}
model = HF_BaseModelWrapper(hf_model)
learn_cbs = [HF_BaseModelCallback]
fit_cbs = [HF_Seq2SeqMetricsCallback(custom_metrics=seq2seq_metrics)]

learn = Learner(dls, 
                model,
                opt_func=partial(Adam),
                loss_func=CrossEntropyLossFlat(), #HF_PreCalculatedLoss()
                cbs=learn_cbs,
                splitter=partial(seq2seq_splitter, arch=hf_arch)) #.to_native_fp16() #.to_fp16()

learn.freeze()
{% endraw %} {% raw %}
 
{% endraw %} {% raw %}
b = dls.one_batch()
preds = learn.model(b[0])

len(preds),preds['loss'].shape, preds['logits'].shape
(4, torch.Size([]), torch.Size([2, 71, 50264]))
{% endraw %} {% raw %}
len(b), len(b[0]), b[0]['input_ids'].shape, len(b[1]), b[1].shape
(2, 3, torch.Size([2, 256]), 2, torch.Size([2, 71]))
{% endraw %} {% raw %}
print(len(learn.opt.param_groups))
3
{% endraw %} {% raw %}
learn.lr_find(suggestions=True)
SuggestedLRs(lr_min=8.317637839354575e-05, lr_steep=5.754399353463668e-06)
{% endraw %} {% raw %}
learn.fit_one_cycle(1, lr_max=4e-5, cbs=fit_cbs)
epoch train_loss valid_loss rouge1 rouge2 rougeL bertscore_precision bertscore_recall bertscore_f1 time
0 1.705197 1.744309 0.292073 0.124522 0.227580 0.887974 0.861465 0.874397 02:02
{% endraw %} {% raw %}
learn.show_results(learner=learn, input_trunc_at=500, target_trunc_at=250)
text target prediction
0 Dan Condon believes in recycling. Just not when it comes to his hotel towels. Condon composts when he's at home in Boulder, Colorado. He eats local, organic and fair-trade food and drives a Honda CR-Z hybrid sports car. You might call him green. Except he's not so green when he travels for his work at an education nonprofit and stays in a hotel, which happens about 10 weeks per year. There, he uses a new towel every day. And don't try to bribe him with a drink or dessert coupon to get him to re Hotel guests who "go green" are happier with their stay.\nIncreasing water and energy costs are pushing hotels to cut costs wherever they can.\nMany hotels find that guests don't mind using the same towels and sheets every night.\nTripAdvisor will be a Dan Condon composts when he's at home in Boulder, Colorado, but not when he stays in a hotel .\n
1 London (CNN) -- In 1948, a hospital outside London witnessed the birth of the Paralympic movement, as a Jewish doctor who had fled Nazi Germany sought to change the lives of patients with spinal injuries -- and inspire new hope in them through sport. The first "Stoke Mandeville Games" were organized in 1948 to coincide with the London Olympics, the second to be held in Britain. Named for the hospital in Buckinghamshire where Prof. Ludwig Guttmann's pioneering spinal injuries unit was based, the Paralympic movement was born in Stoke Mandeville, outside London, in 1948.\n2012 Games will be the biggest yet, with 4,200 competitors from 165 countries.\nIn an echo of the first, post-World War II Games, injured veterans are among the athletes.\nThey "Stoke Mandeville Games" were organized in 1948 to coincide with the London Olympics .\nThe first Paralympic
{% endraw %} {% raw %}
test_article = """
About 10 men armed with pistols and small machine guns raided a casino in Switzerland and made off 
into France with several hundred thousand Swiss francs in the early hours of Sunday morning, police said. 
The men, dressed in black clothes and black ski masks, split into two groups during the raid on the Grand Casino 
Basel, Chief Inspector Peter Gill told CNN. One group tried to break into the casino's vault on the lower level 
but could not get in, but they did rob the cashier of the money that was not secured, he said. The second group 
of armed robbers entered the upper level where the roulette and blackjack tables are located and robbed the 
cashier there, he said. As the thieves were leaving the casino, a woman driving by and unaware of what was 
occurring unknowingly blocked the armed robbers' vehicles. A gunman pulled the woman from her vehicle, beat 
her, and took off for the French border. The other gunmen followed into France, which is only about 100 
meters (yards) from the casino, Gill said. There were about 600 people in the casino at the time of the robbery. 
There were no serious injuries, although one guest on the Casino floor was kicked in the head by one of the 
robbers when he moved, the police officer said. Swiss authorities are working closely with French authorities, 
Gill said. The robbers spoke French and drove vehicles with French lRicense plates. CNN's Andreena Narayan 
contributed to this report.
"""
{% endraw %} {% raw %}
outputs = learn.blurr_generate(test_article, num_return_sequences=3)

for idx, o in enumerate(outputs):
    print(f'=== Prediction {idx+1} ===\n{o}\n')
=== Prediction 1 ===
 Police: About 10 men armed with pistols and machine guns raided a casino in Switzerland .
They made off with several hundred thousand Swiss francs in the early hours of Sunday morning .
There were no serious injuries, although one guest was kicked in the head by one of the robbers .
The robbers spoke French and drove vehicles with French lRicense plates, police say .

=== Prediction 2 ===
 Police: About 10 men armed with pistols and machine guns raided a casino in Switzerland .
They made off with several hundred thousand Swiss francs in the early hours of Sunday morning .
There were no serious injuries, although one guest was kicked in the head by one of the robbers .
The robbers spoke French and drove vehicles with French lRicense plates .

=== Prediction 3 ===
 Police: About 10 men armed with pistols and machine guns raided a casino in Switzerland .
They made off with several hundred thousand Swiss francs in the early hours of Sunday morning .
There were no serious injuries, although one guest was kicked in the head by one of the robbers .
The robbers spoke French and drove vehicles with French lRicense plates, police said .

{% endraw %}

Inference

{% raw %}
export_fname = 'summarize_export'
{% endraw %} {% raw %}
learn.metrics = None
learn.export(fname=f'{export_fname}.pkl')
{% endraw %} {% raw %}
inf_learn = load_learner(fname=f'{export_fname}.pkl')
inf_learn.blurr_generate(test_article)
[' Police: About 10 men armed with pistols and machine guns raided a casino in Switzerland .\nThey made off with several hundred thousand']
{% endraw %}

Tests

The purpose of the following tests is to ensure as much as possible, that the core training code works for the pretrained summarization models below. 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)

{% raw %}
try: del learn; torch.cuda.empty_cache()
except: pass
{% endraw %} {% raw %}
[ model_type for model_type in BLURR.get_models(task='ConditionalGeneration') 
 if (not model_type.__name__.startswith('TF')) ]
[transformers.models.bart.modeling_bart.BartForConditionalGeneration,
 transformers.models.blenderbot.modeling_blenderbot.BlenderbotForConditionalGeneration,
 transformers.models.blenderbot_small.modeling_blenderbot_small.BlenderbotSmallForConditionalGeneration,
 transformers.models.fsmt.modeling_fsmt.FSMTForConditionalGeneration,
 transformers.models.led.modeling_led.LEDForConditionalGeneration,
 transformers.models.m2m_100.modeling_m2m_100.M2M100ForConditionalGeneration,
 transformers.models.mbart.modeling_mbart.MBartForConditionalGeneration,
 transformers.models.mt5.modeling_mt5.MT5ForConditionalGeneration,
 transformers.models.pegasus.modeling_pegasus.PegasusForConditionalGeneration,
 transformers.models.prophetnet.modeling_prophetnet.ProphetNetForConditionalGeneration,
 transformers.models.speech_to_text.modeling_speech_to_text.Speech2TextForConditionalGeneration,
 transformers.models.t5.modeling_t5.T5ForConditionalGeneration,
 transformers.models.xlm_prophetnet.modeling_xlm_prophetnet.XLMProphetNetForConditionalGeneration]
{% endraw %} {% raw %}
pretrained_model_names = [
    'facebook/bart-base',
    #'facebook/blenderbot_small-90M',
    'allenai/led-base-16384',
    'sshleifer/tiny-mbart',
    'google/mt5-small',
    'sshleifer/distill-pegasus-cnn-16-4',
    't5-small', 
    #'microsoft/prophetnet-large-uncased',
    #'microsoft/xprophetnet-large-wiki100-cased', # XLMProphetNet
]
{% endraw %} {% raw %}
path = Path('./')
cnndm_df = pd.read_csv(path/'cnndm_sample.csv')
{% endraw %} {% raw %}
#hide_output
model_cls = AutoModelForSeq2SeqLM
bsz = 2
inp_seq_sz = 64; trg_seq_sz = 40

test_results = []
for model_name in pretrained_model_names:
    error=None
    
    print(f'=== {model_name} ===\n')
    
    hf_tok_kwargs = {}
    if (model_name == 'sshleifer/tiny-mbart'):
        hf_tok_kwargs['src_lang'], hf_tok_kwargs['tgt_lang'] = "en_XX", "en_XX"
    
    hf_arch, hf_config, hf_tokenizer, hf_model = BLURR.get_hf_objects(model_name, 
                                                                      model_cls=model_cls, 
                                                                      tokenizer_kwargs=hf_tok_kwargs)
    
    print(f'architecture:\t{hf_arch}\ntokenizer:\t{type(hf_tokenizer).__name__}\nmodel:\t\t{type(hf_model).__name__}\n')

    # 1. build your DataBlock
    text_gen_kwargs = {}
    if (hf_arch in ['bart', 't5']):
        text_gen_kwargs = {**hf_config.task_specific_params['summarization'], **{'max_length': 30, 'min_length': 10}}
    
    # not all "summarization" parameters are for the model.generate method ... remove them here
    generate_func_args = list(inspect.signature(hf_model.generate).parameters.keys())
    for k in text_gen_kwargs.copy():
        if k not in generate_func_args: del text_gen_kwargs[k]
            
    if (hf_arch == 'mbart'):
        text_gen_kwargs['decoder_start_token_id'] = hf_tokenizer.get_vocab()["en_XX"]
            
            
    def add_t5_prefix(inp): return f'summarize: {inp}' if (hf_arch == 't5') else inp
    
    before_batch_tfm = HF_Seq2SeqBeforeBatchTransform(hf_arch, hf_config, hf_tokenizer, hf_model,
                                                      padding='max_length', 
                                                      max_length=inp_seq_sz, 
                                                      max_target_length=trg_seq_sz, 
                                                      text_gen_kwargs=text_gen_kwargs)
    
    blocks = (HF_Seq2SeqBlock(before_batch_tfm=before_batch_tfm), noop)
    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()

    # 2. build your Learner
    seq2seq_metrics = {
        'rouge': {
            'compute_kwargs': { 'rouge_types': ["rouge1", "rouge2", "rougeL"], 'use_stemmer': True },
            'returns': ["rouge1", "rouge2", "rougeL"]
        }
    }
    
    model = HF_BaseModelWrapper(hf_model)
    learn_cbs = [HF_BaseModelCallback]
    fit_cbs = [
        ShortEpochCallback(0.05, short_valid=True), 
        HF_Seq2SeqMetricsCallback(custom_metrics=seq2seq_metrics)
    ]
 
    learn = Learner(dls, 
                    model,
                    opt_func=ranger,
                    loss_func=HF_PreCalculatedLoss(),
                    cbs=learn_cbs,
                    splitter=partial(seq2seq_splitter, arch=hf_arch)).to_fp16()

    learn.create_opt() 
    learn.freeze()
    
    # 3. Run your tests
    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, inp_seq_sz]))
        test_eq(len(b[1]), bsz)

#         print('*** TESTING One pass through the model ***')
#         preds = learn.model(b[0])
#         test_eq(preds[1].shape[0], bsz)
#         test_eq(preds[1].shape[2], hf_config.vocab_size)

        print('*** TESTING Training/Results ***')
        learn.fit_one_cycle(1, lr_max=1e-3, cbs=fit_cbs)

        test_results.append((hf_arch, type(hf_tokenizer).__name__, type(hf_model).__name__, 'PASSED', ''))
        learn.show_results(learner=learn, max_n=2, input_trunc_at=500, target_trunc_at=250)
    except Exception as err:
        test_results.append((hf_arch, type(hf_tokenizer).__name__, type(hf_model).__name__, 'FAILED', err))
    finally:
        # cleanup
        del learn; torch.cuda.empty_cache()
{% endraw %} {% raw %}
arch tokenizer model_name result error
0 bart BartTokenizerFast BartForConditionalGeneration PASSED
1 led LEDTokenizerFast LEDForConditionalGeneration PASSED
2 mbart MBartTokenizerFast MBartForConditionalGeneration PASSED
3 mt5 T5TokenizerFast MT5ForConditionalGeneration PASSED
4 pegasus PegasusTokenizerFast PegasusForConditionalGeneration PASSED
5 t5 T5TokenizerFast T5ForConditionalGeneration PASSED
{% endraw %}

Cleanup