Train T5 from scratch

Hi all!

I want to train T5 in a new language from scratch an I think the best way to do this is through the unsupervised denoising task. I’ve found that there is no function in huggingface to create the train data (masked data) as the T5 documentation indicates so I’ve tried to create by my own.

According with the T5 original paper, if you have two consecutive tokens to masking you must mask them using only one sentinel token so I need a function that searches the consecutive tokens (“rachas” in my language) in the random indices choosed. Here you have my code:

def racha_detection(lista):
    # It returns a list of lists where each sub-list contains the consecutive tokens in the list
    rachas = []
    racha = []
    for i, element in enumerate(lista):
        if (i<len(lista)-1) and (lista[i+1] == element+1):
            if len(racha)>0:
                rachas.append(racha + [element])          
            else:# (i!=len(lista)-1):
            racha = []
    return rachas

def masking(tokenized_sentence, rachas):
    # Function to mask a tokenized_sentence (token ids) following the rachas described in rachas
    # Only one sentinel_token per racha
    sent_token_id = 0
    enmascared = tokenized_sentence.copy()
    for racha in rachas:
         sent_token = f'<extra_id_{sent_token_id}>'
         sent_id = tokenizer.encode(sent_token)[0]
        for i, idx in enumerate(racha):
            if i==0:
                enmascared[idx] = sent_id
                enmascared[idx] = -100
        sent_token_id += 1
    enmascared = [t for t in enmascared if t!=-100] 

    return enmascared

def add_noise(sentence, tokenizer, percent=0.15):
    # Function that takes a sentence, tokenizer and a noise percentage and returns
    # the masked input_ids and masked target_ids accordling with the T5 paper and HuggingFace docs
    # To see the process working uncomment all the prints ;)
    tokenized_sentence = tokenizer.encode(sentence)
    #print('INPUT: {}'.format(tokenizer.convert_ids_to_tokens(tokenized_sentence)))

    idxs_2_mask = sorted(random.sample(range(len(tokenized_sentence)), 
    rachas = racha_detection(idxs_2_mask)
    enmascared_input = masking(tokenized_sentence, rachas)
    #print('RACHAS INPUT: {}'.format(rachas))
    idxs_2_mask = [idx for idx in range(len(tokenized_sentence)) if idx not in idxs_2_mask]
    rachas = racha_detection(idxs_2_mask)
    enmascared_target = masking(tokenized_sentence, rachas)
    #print('RACHAS TARGET: {}'.format(rachas))

    #print('INPUT: {}'.format(tokenizer.convert_ids_to_tokens(enmascared_input)))
    #print('TARGET: {}'.format(tokenizer.convert_ids_to_tokens(enmascared_target)))

    return enmascared_input, enmascared_target

I dont know if it is correct but it generates sequences like the sequences in the examples :slight_smile: What do you think?


@amlarraz - how did you get on with this? Did it work ok with T5 training? Any tips gratefully appreciated! :slight_smile:

Hi @TheLongSentance , sorry for my late response, I’ve this muted. Yep, I tried it with my own corpus and it seems like it worked, however I can’t make all the test I would like to do, so much work in other areas…

1 Like

Thanks @amlarraz - always too much to do. Thank you for the reply though and all the best.