MathT5.py

import torch
from transformers import T5Tokenizer, T5ForConditionalGeneration

def pretty_print(text, prompt=True):
    s = ""
    if prompt:
        for section in text.split(', '):
            premises = section.split(" and ")
            if len(premises) > 1:
                for premise in premises[:-1]:
                    s += premise + "\n\n\n" + "and" + "\n\n\n"
                s += premises[-1] + "\n\n\n"
            else:
                s += section + "\n\n\n"
    else:
        for equation in text.split("and"):
            s += equation + "\n\n\n"
    return print(s[:-3])


def load_model(model_id):
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
    tokenizer = T5Tokenizer.from_pretrained(model_id)
    model = T5ForConditionalGeneration.from_pretrained(model_id).to(device)
    return tokenizer, model


def inference(prompt, tokenizer, model):
    device = 'cuda' if torch.cuda.is_available() else 'cpu'
    input_ids = tokenizer.encode(prompt, return_tensors='pt', max_length=512, truncation=True).to(device)
    output = model.generate(input_ids=input_ids, max_length=512, early_stopping=True)
    generated_text = tokenizer.decode(output[0], skip_special_tokens=True)
    
    # post-processing
    derivation = generated_text.replace("\\ ","\\")
    partial_symbols = derivation.split(" ")
    backslash_syms = set([i for i in partial_symbols if "\\" in i])
    for i in range(len(partial_symbols)):
        sym = partial_symbols[i]
        for b_sym in backslash_syms:
            if b_sym.replace("\\","") == sym:
                partial_symbols[i] = b_sym
    return " ".join(partial_symbols)