StreamlitModel.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Mon Feb 19 13:56:57 2024

@author: selbl
"""

#This code runs the lyric generation model
#Because it is a script, I do not spend time writing about what each part does
#For more info please refer to the Github script

import torch
from transformers import GPT2Tokenizer, GPT2Config, GPT2Model, GPT2PreTrainedModel
from torch.nn import functional as F
from better_profanity import profanity
import re
import torch.hub

profanity.load_censor_words()    

#It seems the streamlit space does not allow mps
#device = 'cuda' if torch.cuda.is_available() else 'mps' if torch.backends.mps.is_built() else 'cpu'
device = 'cuda' if torch.cuda.is_available() else 'cpu'
    
class GPT2_Model(GPT2PreTrainedModel):

    def __init__(self, config):

        super().__init__(config)

        self.transformer = GPT2Model.from_pretrained('gpt2')
        tokenizer = GPT2Tokenizer.from_pretrained('gpt2', pad_token='<|pad|>')

        # this is necessary since we add a new unique token for pad_token
        self.transformer.resize_token_embeddings(len(tokenizer))

        self.lm_head = torch.nn.Linear(config.n_embd, len(tokenizer), bias=False)

    def forward(self, input_ids, attention_mask=None, token_type_ids=None):

        x = self.transformer(input_ids, attention_mask=attention_mask, token_type_ids=token_type_ids)[0]
        x = self.lm_head(x)

        return x
        

#Dedfine generation function
def generate(idx, max_new_tokens, context_size, tokenizer, model, top_k=10, top_p=0.95):

        for _ in range(max_new_tokens):
            if idx[:,-1].item() != tokenizer.encode(tokenizer.eos_token)[0]:
                # crop idx to the last block_size tokens
                idx_cond = idx[:, -context_size:]
                # get the predictions
                logits = model(idx_cond)
                # focus only on the last time step
                logits = logits[:, -1, :]
                # apply softmax to get probabilities
                probs = F.softmax(logits, dim=-1)
                # sort probabilities in descending order
                sorted_probs, indices = torch.sort(probs, descending=True)
                # compute cumsum of probabilities
                probs_cumsum = torch.cumsum(sorted_probs, dim=1)
                # choose only top_p tokens
                sorted_probs, indices = sorted_probs[:, :probs_cumsum[[probs_cumsum < top_p]].size()[0] + 1], indices[:, :probs_cumsum[[probs_cumsum < top_p]].size()[0] +1]
                # choose only top_k tokens
                sorted_probs, indices = sorted_probs[:,:top_k], indices[:,:top_k]
                # sample from the distribution
                sorted_probs = F.softmax(sorted_probs, dim=-1)
                idx_next = indices[:, torch.multinomial(sorted_probs, num_samples=1)].squeeze(0)
                # append new token ids
                idx = torch.cat((idx, idx_next), dim=1)
            else:
                break

        return idx

#Define capitalization functions
def custom_capitalize(match):
    return match.group(1).capitalize()

def capitalize_string(input_string):
    # Capitalize every first letter after "\n\n "
    result = re.sub(r'\n\n\s*([a-zA-Z])', lambda x: '\n\n' + x.group(1).upper(), input_string)
    # Capitalize every first letter in every word inside brackets ([ ])
    result = re.sub(r'\[([^\]]*)\]', lambda x: '[' + ' '.join(word.capitalize() for word in x.group(1).split()) + ']', result)
    # Capitalize every instance of the letter i by itself
    result = re.sub(r'\bi\b', 'I', result)
    return result

#Format the string
def format_string(input_string):
    result = ''.join([char + ('\n\n' if char == ']' else '') for char in input_string.replace('\n', '')])
    result = result.replace('[', '\n\n[')
    #Capitalize for good measure
    result = capitalize_string(result)
    return result


def TextGeneration(prompt,prof=False,parts=True):
    #Load everything
    #Load model
    configuration = GPT2Config()
    gpt_model = GPT2_Model(configuration).to(device)
    gpt_model.load_state_dict(torch.load('GPT-Trained-Model.pt',map_location=device))
    #state_dict = torch.hub.load_state_dict_from_url(r'https://github.com/Selbl/LyricGeneration/raw/main/GPT-Trained-Model-Prod.pt?download=',map_location=device)
    #state_dict = torch.hub.load_state_dict_from_url(r'https://huggingface.co/spaces/selbl/LyricGeneration/resolve/main/GPT-Trained-Model',map_location=device)
    #gpt_model.load_state_dict(state_dict)
    #Load tokenizer
    tokenizer = GPT2Tokenizer.from_pretrained('gpt2', pad_token='<|pad|>') 
    #Call loader function
    gpt_model.eval()
    #Set model parameters
    prefix = '[verse]'

    #Input prompt
    prompt = prompt.lower()

    #Pre-process prompt
    if parts:
        prompt = prefix + ' ' + prompt
    #Generate
    generated = torch.tensor(tokenizer.encode(prompt)).unsqueeze(0)
    generated = generated.to(device)
    
    sample_outputs = generate(generated,
                             max_new_tokens=200,
                             context_size=450,
                             tokenizer=tokenizer,
                             model=gpt_model,
                             top_k=10,
                             top_p=0.95)
    
    #Store the lyrics and decode
    lyric = tokenizer.decode(sample_outputs[0], skip_special_tokens=True)
    #Format if it has parts
    if parts:
        lyric = format_string(lyric)
    #Remove profanity
    if not prof:
        lyric = profanity.censor(lyric)
    return lyric