min-dalle-test/min_dalle/min_dalle.py

236 lines
9.1 KiB
Python
Raw Normal View History

2022-06-29 02:22:54 +00:00
import os
from PIL import Image
import numpy
2022-07-04 20:06:49 +00:00
from torch import LongTensor, FloatTensor
2022-06-27 15:57:56 +00:00
import torch
import json
2022-07-01 22:16:55 +00:00
import requests
2022-07-04 20:06:49 +00:00
from typing import Callable, Tuple
2022-06-29 01:28:36 +00:00
torch.set_grad_enabled(False)
2022-06-29 02:22:54 +00:00
torch.set_num_threads(os.cpu_count())
2022-06-27 15:57:56 +00:00
from .text_tokenizer import TextTokenizer
from .models import DalleBartEncoder, DalleBartDecoder, VQGanDetokenizer
2022-07-01 22:16:55 +00:00
MIN_DALLE_REPO = 'https://huggingface.co/kuprel/min-dalle/resolve/main/'
2022-06-27 15:57:56 +00:00
class MinDalle:
def __init__(
2022-07-01 22:16:55 +00:00
self,
is_mega: bool,
2022-06-30 15:25:24 +00:00
is_reusable: bool = True,
2022-07-01 22:16:55 +00:00
models_root: str = 'pretrained',
2022-07-02 00:17:20 +00:00
is_verbose = True
):
2022-07-01 22:16:55 +00:00
self.is_mega = is_mega
2022-07-01 19:53:39 +00:00
self.is_reusable = is_reusable
2022-07-02 00:17:20 +00:00
self.is_verbose = is_verbose
2022-07-01 19:53:39 +00:00
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
2022-07-02 00:17:20 +00:00
if self.is_verbose: print("initializing MinDalle")
model_name = 'dalle_bart_{}'.format('mega' if is_mega else 'mini')
2022-07-01 22:16:55 +00:00
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')
2022-07-01 19:53:39 +00:00
self.init_tokenizer()
if is_reusable:
self.init_encoder()
self.init_decoder()
self.init_detokenizer()
2022-07-01 22:16:55 +00:00
def download_tokenizer(self):
2022-07-02 00:17:20 +00:00
if self.is_verbose: print("downloading tokenizer params")
2022-07-01 22:16:55 +00:00
suffix = '' if self.is_mega else '_mini'
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):
2022-07-02 00:17:20 +00:00
if self.is_verbose: print("downloading encoder params")
2022-07-01 22:16:55 +00:00
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):
2022-07-02 00:17:20 +00:00
if self.is_verbose: print("downloading decoder params")
2022-07-01 22:16:55 +00:00
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):
2022-07-02 00:17:20 +00:00
if self.is_verbose: print("downloading detokenizer params")
2022-07-01 22:16:55 +00:00
params = requests.get(MIN_DALLE_REPO + 'detoker.pt')
with open(self.detoker_params_path, 'wb') as f: f.write(params.content)
2022-07-01 19:53:39 +00:00
def init_tokenizer(self):
2022-07-01 22:16:55 +00:00
is_downloaded = os.path.exists(self.vocab_path)
is_downloaded &= os.path.exists(self.merges_path)
if not is_downloaded: self.download_tokenizer()
2022-07-02 00:17:20 +00:00
if self.is_verbose: print("intializing TextTokenizer")
2022-07-01 22:16:55 +00:00
with open(self.vocab_path, 'r', encoding='utf8') as f:
vocab = json.load(f)
2022-07-01 22:16:55 +00:00
with open(self.merges_path, 'r', encoding='utf8') as f:
merges = f.read().split("\n")[1:-1]
2022-07-04 20:06:49 +00:00
self.tokenizer = TextTokenizer(vocab, merges)
2022-06-27 15:57:56 +00:00
def init_encoder(self):
2022-07-01 22:16:55 +00:00
is_downloaded = os.path.exists(self.encoder_params_path)
if not is_downloaded: self.download_encoder()
2022-07-02 00:17:20 +00:00
if self.is_verbose: print("initializing DalleBartEncoder")
self.encoder = DalleBartEncoder(
2022-07-01 19:53:39 +00:00
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
)
params = torch.load(self.encoder_params_path)
self.encoder.load_state_dict(params, strict=False)
del params
2022-06-30 15:44:36 +00:00
if torch.cuda.is_available(): self.encoder = self.encoder.cuda()
2022-06-27 15:57:56 +00:00
def init_decoder(self):
2022-07-01 22:16:55 +00:00
is_downloaded = os.path.exists(self.decoder_params_path)
if not is_downloaded: self.download_decoder()
2022-07-02 00:17:20 +00:00
if self.is_verbose: print("initializing DalleBartDecoder")
self.decoder = DalleBartDecoder(
2022-07-01 19:53:39 +00:00
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,
start_token = self.image_vocab_count
)
params = torch.load(self.decoder_params_path)
self.decoder.load_state_dict(params, strict=False)
del params
2022-06-30 15:44:36 +00:00
if torch.cuda.is_available(): self.decoder = self.decoder.cuda()
2022-06-27 15:57:56 +00:00
2022-07-01 14:17:29 +00:00
def init_detokenizer(self):
2022-07-01 22:16:55 +00:00
is_downloaded = os.path.exists(self.detoker_params_path)
if not is_downloaded: self.download_detokenizer()
2022-07-02 00:17:20 +00:00
if self.is_verbose: print("initializing VQGanDetokenizer")
2022-07-01 14:17:29 +00:00
self.detokenizer = VQGanDetokenizer()
params = torch.load(self.detoker_params_path)
self.detokenizer.load_state_dict(params)
del params
if torch.cuda.is_available(): self.detokenizer = self.detokenizer.cuda()
2022-07-04 20:06:49 +00:00
def image_from_tokens(
self,
grid_size: int,
image_tokens: LongTensor,
is_verbose: bool = False
) -> Image.Image:
if not self.is_reusable: del self.decoder
if torch.cuda.is_available(): torch.cuda.empty_cache()
if not self.is_reusable: self.init_detokenizer()
if is_verbose: print("detokenizing image")
images = self.detokenizer.forward(image_tokens).to(torch.uint8)
if not self.is_reusable: del self.detokenizer
images = images.reshape([grid_size] * 2 + list(images.shape[1:]))
image = images.flatten(1, 2).transpose(0, 1).flatten(1, 2)
image = Image.fromarray(image.to('cpu').detach().numpy())
return image
def generate_image_tokens(
self,
text: str,
seed: int,
2022-07-04 20:06:49 +00:00
grid_size: int,
row_count: int,
mid_count: int = None,
handle_intermediate_image: Callable[[int, Image.Image], None] = None,
is_verbose: bool = False
) -> LongTensor:
2022-07-04 20:06:49 +00:00
if is_verbose: print("tokenizing text")
tokens = self.tokenizer.tokenize(text, is_verbose=is_verbose)
if is_verbose: print("text 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).to(torch.long)
if torch.cuda.is_available(): text_tokens = text_tokens.cuda()
2022-06-27 15:57:56 +00:00
2022-06-30 15:25:24 +00:00
if not self.is_reusable: self.init_encoder()
2022-07-04 20:06:49 +00:00
if is_verbose: print("encoding text tokens")
encoder_state = self.encoder.forward(text_tokens)
2022-06-30 15:25:24 +00:00
if not self.is_reusable: del self.encoder
2022-07-04 12:05:55 +00:00
if torch.cuda.is_available(): torch.cuda.empty_cache()
2022-06-27 15:57:56 +00:00
2022-06-30 15:25:24 +00:00
if not self.is_reusable: self.init_decoder()
2022-07-04 20:06:49 +00:00
encoder_state, attention_mask, attention_state, image_tokens = (
self.decoder.decode_initial(
seed,
grid_size ** 2,
text_tokens,
encoder_state
)
)
2022-07-04 20:06:49 +00:00
for row_index in range(row_count):
if is_verbose:
print('sampling row {} of {}'.format(row_index + 1, row_count))
attention_state, image_tokens = self.decoder.decode_row(
row_index,
encoder_state,
attention_mask,
attention_state,
image_tokens
)
if mid_count is not None:
if ((row_index + 1) * mid_count) % row_count == 0:
tokens = image_tokens[:, 1:]
image = self.image_from_tokens(grid_size, tokens, is_verbose)
handle_intermediate_image(row_index, image)
return image_tokens[:, 1:]
2022-06-27 15:57:56 +00:00
def generate_image(
self,
2022-07-04 20:06:49 +00:00
text: str,
seed: int = -1,
2022-07-04 20:06:49 +00:00
grid_size: int = 1,
mid_count: int = None,
handle_intermediate_image: Callable[[Image.Image], None] = None,
is_verbose: bool = False
) -> Image.Image:
2022-07-04 20:06:49 +00:00
image_tokens = self.generate_image_tokens(
text,
seed,
grid_size,
row_count = 16,
mid_count = mid_count,
handle_intermediate_image = handle_intermediate_image,
is_verbose = is_verbose
)
return self.image_from_tokens(grid_size, image_tokens, is_verbose)