min_dalle/min_dalle.py

import os
from PIL import Image
import numpy
from torch import LongTensor, FloatTensor
import torch
import torch.backends.cudnn, torch.backends.cuda
import json
import requests
from typing import Iterator
from .text_tokenizer import TextTokenizer
from .models import DalleBartEncoder, DalleBartDecoder, VQGanDetokenizer
import streamlit as st

torch.set_grad_enabled(False)
torch.set_num_threads(os.cpu_count())
torch.backends.cudnn.enabled = True
torch.backends.cudnn.allow_tf32 = True

MIN_DALLE_REPO = 'https://huggingface.co/kuprel/min-dalle/resolve/main/'
IMAGE_TOKEN_COUNT = 256


class MinDalle:
    def __init__(
        self,
        models_root: str = 'pretrained',
        dtype: torch.dtype = torch.float32,
        device: str = None,
        is_mega: bool = True, 
        is_reusable: bool = True,
        is_verbose = True
    ):
        if device == None:
            device = 'cuda' if torch.cuda.is_available() else 'cpu'
        if is_verbose: print("using device", device)
        self.device = device
        self.is_mega = is_mega
        self.is_reusable = is_reusable
        self.dtype = dtype
        self.is_verbose = is_verbose
        self.text_token_count = 64
        self.layer_count = 24 if is_mega else 12
        self.attention_head_count = 32 if is_mega else 16
        self.embed_count = 2048 if is_mega else 1024
        self.glu_embed_count = 4096 if is_mega else 2730
        self.text_vocab_count = 50272 if is_mega else 50264
        self.image_vocab_count = 16415 if is_mega else 16384

        model_name = 'dalle_bart_{}'.format('mega' if is_mega else 'mini')
        dalle_path = os.path.join(models_root, model_name)
        vqgan_path = os.path.join(models_root, 'vqgan')
        if not os.path.exists(dalle_path): os.makedirs(dalle_path)
        if not os.path.exists(vqgan_path): os.makedirs(vqgan_path)
        self.vocab_path = os.path.join(dalle_path, 'vocab.json')
        self.merges_path = os.path.join(dalle_path, 'merges.txt')
        self.encoder_params_path = os.path.join(dalle_path, 'encoder.pt')
        self.decoder_params_path = os.path.join(dalle_path, 'decoder.pt')
        self.detoker_params_path = os.path.join(vqgan_path, 'detoker.pt')

        self.init_tokenizer()
        if is_reusable:
            self.init_encoder()
            self.init_decoder()
            self.init_detokenizer()


    def download_tokenizer(self):
        if self.is_verbose: print("downloading tokenizer params")
        suffix = '' if self.is_mega else '_mini'
        _ = requests.get(MIN_DALLE_REPO + 'config.json') # trigger HF download
        vocab = requests.get(MIN_DALLE_REPO + 'vocab{}.json'.format(suffix))
        merges = requests.get(MIN_DALLE_REPO + 'merges{}.txt'.format(suffix))
        with open(self.vocab_path, 'wb') as f: f.write(vocab.content)
        with open(self.merges_path, 'wb') as f: f.write(merges.content)


    def download_encoder(self):
        if self.is_verbose: print("downloading encoder params")
        suffix = '' if self.is_mega else '_mini'
        params = requests.get(MIN_DALLE_REPO + 'encoder{}.pt'.format(suffix))
        with open(self.encoder_params_path, 'wb') as f: f.write(params.content)


    def download_decoder(self):
        if self.is_verbose: print("downloading decoder params")
        suffix = '' if self.is_mega else '_mini'
        params = requests.get(MIN_DALLE_REPO + 'decoder{}.pt'.format(suffix))
        with open(self.decoder_params_path, 'wb') as f: f.write(params.content)
    

    def download_detokenizer(self):
        if self.is_verbose: print("downloading detokenizer params")
        params = requests.get(MIN_DALLE_REPO + 'detoker.pt')
        with open(self.detoker_params_path, 'wb') as f: f.write(params.content)


    def init_tokenizer(self):
        is_downloaded = os.path.exists(self.vocab_path)
        is_downloaded &= os.path.exists(self.merges_path)
        if not is_downloaded: self.download_tokenizer()
        if self.is_verbose: print("intializing TextTokenizer")
        with open(self.vocab_path, 'r', encoding='utf8') as f:
            vocab = json.load(f)
        with open(self.merges_path, 'r', encoding='utf8') as f:
            merges = f.read().split("\n")[1:-1]
        self.tokenizer = TextTokenizer(vocab, merges)


    def init_encoder(self):
        is_downloaded = os.path.exists(self.encoder_params_path)
        if not is_downloaded: self.download_encoder()
        if self.is_verbose: print("initializing DalleBartEncoder")
        self.encoder = DalleBartEncoder(
            attention_head_count = self.attention_head_count,
            embed_count = self.embed_count,
            glu_embed_count = self.glu_embed_count,
            text_token_count = self.text_token_count,
            text_vocab_count = self.text_vocab_count,
            layer_count = self.layer_count,
            device=self.device
        ).to(self.dtype).eval()
        params = torch.load(self.encoder_params_path)
        self.encoder.load_state_dict(params, strict=False)
        del params
        self.encoder = self.encoder.to(device=self.device)


    def init_decoder(self):
        is_downloaded = os.path.exists(self.decoder_params_path)
        if not is_downloaded: self.download_decoder()
        if self.is_verbose: print("initializing DalleBartDecoder")
        self.decoder = DalleBartDecoder(
            image_vocab_count = self.image_vocab_count,
            attention_head_count = self.attention_head_count,
            embed_count = self.embed_count,
            glu_embed_count = self.glu_embed_count,
            layer_count = self.layer_count,
            device=self.device
        ).to(self.dtype).eval()
        params = torch.load(self.decoder_params_path)
        self.decoder.load_state_dict(params, strict=False)
        del params
        self.decoder = self.decoder.to(device=self.device)


    def init_detokenizer(self):
        is_downloaded = os.path.exists(self.detoker_params_path)
        if not is_downloaded: self.download_detokenizer()
        if self.is_verbose: print("initializing VQGanDetokenizer")
        self.detokenizer = VQGanDetokenizer().eval()
        params = torch.load(self.detoker_params_path)
        self.detokenizer.load_state_dict(params)
        del params
        self.detokenizer = self.detokenizer.to(device=self.device)


    def image_grid_from_tokens(
        self,
        image_tokens: LongTensor,
        is_seamless: bool,
        is_verbose: bool = False
    ) -> FloatTensor:
        if not self.is_reusable: del self.decoder
        torch.cuda.empty_cache()
        if not self.is_reusable: self.init_detokenizer()
        if is_verbose: print("detokenizing image")
        images = self.detokenizer.forward(is_seamless, image_tokens)
        if not self.is_reusable: del self.detokenizer
        return images


    def generate_raw_image_stream(
        self, 
        text: str, 
        seed: int,
        grid_size: int,
        progressive_outputs: bool = False,
        is_seamless: bool = False,
        temperature: float = 1,
        top_k: int = 256,
        supercondition_factor: int = 16,
        is_verbose: bool = False
    ) -> Iterator[FloatTensor]:
        image_count = grid_size ** 2
        if is_verbose: print("tokenizing text")
        tokens = self.tokenizer.tokenize(text, is_verbose=is_verbose)
        if len(tokens) > self.text_token_count: 
            tokens = tokens[:self.text_token_count]
        if is_verbose: print("{} text tokens".format(len(tokens)), tokens)
        text_tokens = numpy.ones((2, 64), dtype=numpy.int32)
        text_tokens[0, :2] = [tokens[0], tokens[-1]]
        text_tokens[1, :len(tokens)] = tokens
        text_tokens = torch.tensor(
            text_tokens, 
            dtype=torch.long, 
            device=self.device
        )

        if not self.is_reusable: self.init_encoder()
        if is_verbose: print("encoding text tokens")
        with torch.cuda.amp.autocast(dtype=self.dtype):
            encoder_state = self.encoder.forward(text_tokens)
        if not self.is_reusable: del self.encoder
        torch.cuda.empty_cache()

        if not self.is_reusable: self.init_decoder()

        with torch.cuda.amp.autocast(dtype=self.dtype):
            expanded_indices = [0] * image_count + [1] * image_count
            text_tokens = text_tokens[expanded_indices]
            encoder_state = encoder_state[expanded_indices]
            attention_mask = text_tokens.not_equal(1)
            attention_state = torch.zeros(
                size=(
                    self.layer_count,
                    image_count * 4,
                    IMAGE_TOKEN_COUNT,
                    self.embed_count
                ), 
                device=self.device
            )
            image_tokens = torch.full(
                (IMAGE_TOKEN_COUNT + 1, image_count), 
                self.image_vocab_count,
                dtype=torch.long,
                device=self.device
            )
            
            if seed > 0: torch.manual_seed(seed)

        token_indices = torch.arange(IMAGE_TOKEN_COUNT, device=self.device)
        settings = torch.tensor(
            [temperature, top_k, supercondition_factor], 
            dtype=torch.float32,
            device=self.device
        )
        for i in range(IMAGE_TOKEN_COUNT):
            if(st.session_state.page != 0):
                break
            st.session_state.bar.progress(i/IMAGE_TOKEN_COUNT)

            torch.cuda.empty_cache()                
            with torch.cuda.amp.autocast(dtype=self.dtype):
                image_tokens[i + 1], attention_state = self.decoder.forward(
                    settings=settings,
                    attention_mask=attention_mask,
                    encoder_state=encoder_state,
                    attention_state=attention_state,
                    prev_tokens=image_tokens[i],
                    token_index=token_indices[[i]]
                )

            with torch.cuda.amp.autocast(dtype=torch.float32):
                if ((i + 1) % 32 == 0 and progressive_outputs) or i + 1 == 256:
                    yield self.image_grid_from_tokens(
                        image_tokens=image_tokens[1:].T,
                        is_seamless=is_seamless,
                        is_verbose=is_verbose
                    )

    def generate_image_stream(self, *args, **kwargs) -> Iterator[Image.Image]:
        image_stream = self.generate_raw_image_stream(*args, **kwargs)
        for image in image_stream:
            image = image.to(torch.uint8).to('cpu').numpy()
            yield Image.fromarray(image)


    def generate_images_stream(self, *args, **kwargs) -> Iterator[FloatTensor]:
        image_stream = self.generate_raw_image_stream(*args, **kwargs)
        for image in image_stream:
            grid_size = kwargs['grid_size']
            image = image.view([grid_size * 256, grid_size, 256, 3])
            image = image.transpose(1, 0)
            image = image.reshape([grid_size ** 2, 2 ** 8, 2 ** 8, 3])
            yield image


    def generate_image(self, *args, **kwargs) -> Image.Image:
        image_stream = self.generate_image_stream(
            *args, **kwargs, 
            progressive_outputs=False
        )
        return next(image_stream)


    def generate_images(self, *args, **kwargs) -> Image.Image:
        images_stream = self.generate_images_stream(
            *args, **kwargs, 
            progressive_outputs=False
        )
        return next(images_stream)