Brett Kuprel
2 years ago
committed by
GitHub
commit
e3329a7f64
22 changed files with 588 additions and 483 deletions
@ -0,0 +1,2 @@ |
||||
* linguist-vendored |
||||
*.py linguist-vendored=false |
||||
@ -0,0 +1,12 @@ |
||||
build: |
||||
cuda: "11.0" |
||||
gpu: true |
||||
python_version: "3.8" |
||||
system_packages: |
||||
- "libgl1-mesa-glx" |
||||
- "libglib2.0-0" |
||||
python_packages: |
||||
- "torch==1.10.1" |
||||
- "flax==0.5.2" |
||||
|
||||
predict: "predict.py:Predictor" |
||||
|
Before Width: | Height: | Size: 55 KiB |
|
After Width: | Height: | Size: 127 KiB |
|
Before Width: | Height: | Size: 90 KiB After Width: | Height: | Size: 101 KiB |
File diff suppressed because one or more lines are too long
@ -1,77 +0,0 @@ |
||||
import os |
||||
import json |
||||
import numpy |
||||
from PIL import Image |
||||
from typing import Tuple, List |
||||
|
||||
from min_dalle.load_params import load_dalle_bart_flax_params |
||||
from min_dalle.text_tokenizer import TextTokenizer |
||||
from min_dalle.min_dalle_flax import generate_image_tokens_flax |
||||
from min_dalle.min_dalle_torch import ( |
||||
generate_image_tokens_torch, |
||||
detokenize_torch |
||||
) |
||||
|
||||
def load_dalle_bart_metadata(path: str) -> Tuple[dict, dict, List[str]]: |
||||
print("parsing metadata from {}".format(path)) |
||||
for f in ['config.json', 'flax_model.msgpack', 'vocab.json', 'merges.txt']: |
||||
assert(os.path.exists(os.path.join(path, f))) |
||||
with open(path + '/config.json', 'r') as f: |
||||
config = json.load(f) |
||||
with open(path + '/vocab.json') as f: |
||||
vocab = json.load(f) |
||||
with open(path + '/merges.txt') as f: |
||||
merges = f.read().split("\n")[1:-1] |
||||
return config, vocab, merges |
||||
|
||||
|
||||
def tokenize_text( |
||||
text: str, |
||||
config: dict, |
||||
vocab: dict, |
||||
merges: List[str] |
||||
) -> numpy.ndarray: |
||||
print("tokenizing text") |
||||
tokens = TextTokenizer(vocab, merges)(text) |
||||
print("text tokens", tokens) |
||||
text_tokens = numpy.ones((2, config['max_text_length']), dtype=numpy.int32) |
||||
text_tokens[0, :len(tokens)] = tokens |
||||
text_tokens[1, :2] = [tokens[0], tokens[-1]] |
||||
return text_tokens |
||||
|
||||
|
||||
def generate_image_from_text( |
||||
text: str, |
||||
is_mega: bool = False, |
||||
is_torch: bool = False, |
||||
seed: int = 0, |
||||
image_token_count: int = 256 |
||||
) -> Image.Image: |
||||
model_name = 'mega' if is_mega else 'mini' |
||||
model_path = './pretrained/dalle_bart_{}'.format(model_name) |
||||
config, vocab, merges = load_dalle_bart_metadata(model_path) |
||||
text_tokens = tokenize_text(text, config, vocab, merges) |
||||
params_dalle_bart = load_dalle_bart_flax_params(model_path) |
||||
|
||||
image_tokens = numpy.zeros(config['image_length']) |
||||
if is_torch: |
||||
image_tokens[:image_token_count] = generate_image_tokens_torch( |
||||
text_tokens = text_tokens, |
||||
seed = seed, |
||||
config = config, |
||||
params = params_dalle_bart, |
||||
image_token_count = image_token_count |
||||
) |
||||
else: |
||||
image_tokens[...] = generate_image_tokens_flax( |
||||
text_tokens = text_tokens, |
||||
seed = seed, |
||||
config = config, |
||||
params = params_dalle_bart, |
||||
) |
||||
|
||||
if image_token_count == config['image_length']: |
||||
image = detokenize_torch(image_tokens) |
||||
return Image.fromarray(image) |
||||
else: |
||||
return None |
||||
@ -0,0 +1,46 @@ |
||||
import os |
||||
import json |
||||
import numpy |
||||
|
||||
from .text_tokenizer import TextTokenizer |
||||
from .load_params import load_vqgan_torch_params, load_dalle_bart_flax_params |
||||
from .models.vqgan_detokenizer import VQGanDetokenizer |
||||
|
||||
class MinDalleBase: |
||||
def __init__(self, is_mega: bool): |
||||
self.is_mega = is_mega |
||||
model_name = 'dalle_bart_{}'.format('mega' if is_mega else 'mini') |
||||
self.model_path = os.path.join('pretrained', model_name) |
||||
|
||||
print("reading files from {}".format(self.model_path)) |
||||
config_path = os.path.join(self.model_path, 'config.json') |
||||
vocab_path = os.path.join(self.model_path, 'vocab.json') |
||||
merges_path = os.path.join(self.model_path, 'merges.txt') |
||||
|
||||
with open(config_path, 'r', encoding='utf8') as f: |
||||
self.config = json.load(f) |
||||
with open(vocab_path, 'r', encoding='utf8') as f: |
||||
vocab = json.load(f) |
||||
with open(merges_path, 'r', encoding='utf8') as f: |
||||
merges = f.read().split("\n")[1:-1] |
||||
|
||||
self.tokenizer = TextTokenizer(vocab, merges) |
||||
|
||||
|
||||
def init_detokenizer(self): |
||||
print("initializing VQGanDetokenizer") |
||||
params = load_vqgan_torch_params('./pretrained/vqgan') |
||||
self.detokenizer = VQGanDetokenizer() |
||||
self.detokenizer.load_state_dict(params) |
||||
del params |
||||
|
||||
|
||||
def tokenize_text(self, text: str) -> numpy.ndarray: |
||||
print("tokenizing text") |
||||
tokens = self.tokenizer.tokenize(text) |
||||
print("text tokens", tokens) |
||||
text_token_count = self.config['max_text_length'] |
||||
text_tokens = numpy.ones((2, text_token_count), dtype=numpy.int32) |
||||
text_tokens[0, :len(tokens)] = tokens |
||||
text_tokens[1, :2] = [tokens[0], tokens[-1]] |
||||
return text_tokens |
||||
@ -1,79 +1,80 @@ |
||||
import jax |
||||
from jax import numpy as jnp |
||||
import numpy |
||||
from PIL import Image |
||||
import torch |
||||
|
||||
from .min_dalle_base import MinDalleBase |
||||
from .models.dalle_bart_encoder_flax import DalleBartEncoderFlax |
||||
from .models.dalle_bart_decoder_flax import DalleBartDecoderFlax |
||||
|
||||
from .load_params import load_dalle_bart_flax_params |
||||
|
||||
def encode_flax( |
||||
text_tokens: numpy.ndarray, |
||||
config: dict, |
||||
params: dict |
||||
) -> jnp.ndarray: |
||||
print("loading flax encoder") |
||||
encoder: DalleBartEncoderFlax = DalleBartEncoderFlax( |
||||
attention_head_count = config['encoder_attention_heads'], |
||||
embed_count = config['d_model'], |
||||
glu_embed_count = config['encoder_ffn_dim'], |
||||
text_token_count = config['max_text_length'], |
||||
text_vocab_count = config['encoder_vocab_size'], |
||||
layer_count = config['encoder_layers'] |
||||
).bind({'params': params.pop('encoder')}) |
||||
|
||||
print("encoding text tokens") |
||||
encoder_state = encoder(text_tokens) |
||||
del encoder |
||||
return encoder_state |
||||
class MinDalleFlax(MinDalleBase): |
||||
def __init__(self, is_mega: bool, is_reusable: bool = True): |
||||
super().__init__(is_mega) |
||||
self.is_reusable = is_reusable |
||||
print("initializing MinDalleFlax") |
||||
self.model_params = load_dalle_bart_flax_params(self.model_path) |
||||
if is_reusable: |
||||
self.init_encoder() |
||||
self.init_decoder() |
||||
self.init_detokenizer() |
||||
|
||||
|
||||
def decode_flax( |
||||
text_tokens: jnp.ndarray, |
||||
encoder_state: jnp.ndarray, |
||||
config: dict, |
||||
seed: int, |
||||
params: dict |
||||
) -> jnp.ndarray: |
||||
print("loading flax decoder") |
||||
decoder = DalleBartDecoderFlax( |
||||
image_token_count = config['image_length'], |
||||
text_token_count = config['max_text_length'], |
||||
image_vocab_count = config['image_vocab_size'], |
||||
attention_head_count = config['decoder_attention_heads'], |
||||
embed_count = config['d_model'], |
||||
glu_embed_count = config['decoder_ffn_dim'], |
||||
layer_count = config['decoder_layers'], |
||||
start_token = config['decoder_start_token_id'] |
||||
) |
||||
print("sampling image tokens") |
||||
image_tokens = decoder.sample_image_tokens( |
||||
text_tokens, |
||||
encoder_state, |
||||
jax.random.PRNGKey(seed), |
||||
params.pop('decoder') |
||||
) |
||||
del decoder |
||||
return image_tokens |
||||
def init_encoder(self): |
||||
print("initializing DalleBartEncoderFlax") |
||||
self.encoder: DalleBartEncoderFlax = DalleBartEncoderFlax( |
||||
attention_head_count = self.config['encoder_attention_heads'], |
||||
embed_count = self.config['d_model'], |
||||
glu_embed_count = self.config['encoder_ffn_dim'], |
||||
text_token_count = self.config['max_text_length'], |
||||
text_vocab_count = self.config['encoder_vocab_size'], |
||||
layer_count = self.config['encoder_layers'] |
||||
).bind({'params': self.model_params.pop('encoder')}) |
||||
|
||||
|
||||
def generate_image_tokens_flax( |
||||
text_tokens: numpy.ndarray, |
||||
seed: int, |
||||
config: dict, |
||||
params: dict |
||||
) -> numpy.ndarray: |
||||
encoder_state = encode_flax( |
||||
text_tokens, |
||||
config, |
||||
params |
||||
) |
||||
image_tokens = decode_flax( |
||||
text_tokens, |
||||
encoder_state, |
||||
config, |
||||
seed, |
||||
params |
||||
) |
||||
image_tokens = numpy.array(image_tokens) |
||||
print("image tokens", list(image_tokens)) |
||||
return image_tokens |
||||
def init_decoder(self): |
||||
print("initializing DalleBartDecoderFlax") |
||||
self.decoder = DalleBartDecoderFlax( |
||||
image_token_count = self.config['image_length'], |
||||
text_token_count = self.config['max_text_length'], |
||||
image_vocab_count = self.config['image_vocab_size'], |
||||
attention_head_count = self.config['decoder_attention_heads'], |
||||
embed_count = self.config['d_model'], |
||||
glu_embed_count = self.config['decoder_ffn_dim'], |
||||
layer_count = self.config['decoder_layers'], |
||||
start_token = self.config['decoder_start_token_id'] |
||||
) |
||||
|
||||
|
||||
def generate_image(self, text: str, seed: int) -> Image.Image: |
||||
text_tokens = self.tokenize_text(text) |
||||
|
||||
if not self.is_reusable: self.init_encoder() |
||||
print("encoding text tokens") |
||||
encoder_state = self.encoder(text_tokens) |
||||
if not self.is_reusable: del self.encoder |
||||
|
||||
if not self.is_reusable: |
||||
self.init_decoder() |
||||
params = self.model_params.pop('decoder') |
||||
else: |
||||
params = self.model_params['decoder'] |
||||
print("sampling image tokens") |
||||
image_tokens = self.decoder.sample_image_tokens( |
||||
text_tokens, |
||||
encoder_state, |
||||
jax.random.PRNGKey(seed), |
||||
params |
||||
) |
||||
if not self.is_reusable: del self.decoder |
||||
|
||||
image_tokens = torch.tensor(numpy.array(image_tokens)) |
||||
|
||||
if not self.is_reusable: self.init_detokenizer() |
||||
print("detokenizing image") |
||||
image = self.detokenizer.forward(image_tokens).to(torch.uint8) |
||||
if not self.is_reusable: del self.detokenizer |
||||
image = Image.fromarray(image.to('cpu').detach().numpy()) |
||||
return image |
||||
@ -1,113 +1,114 @@ |
||||
import numpy |
||||
import os |
||||
from PIL import Image |
||||
from typing import Dict |
||||
from torch import LongTensor, FloatTensor |
||||
from torch import LongTensor |
||||
import torch |
||||
torch.no_grad() |
||||
torch.set_grad_enabled(False) |
||||
torch.set_num_threads(os.cpu_count()) |
||||
|
||||
from .models.vqgan_detokenizer import VQGanDetokenizer |
||||
from .load_params import ( |
||||
convert_and_save_torch_params, |
||||
load_dalle_bart_flax_params |
||||
) |
||||
from .min_dalle_base import MinDalleBase |
||||
from .models.dalle_bart_encoder_torch import DalleBartEncoderTorch |
||||
from .models.dalle_bart_decoder_torch import DalleBartDecoderTorch |
||||
|
||||
from .load_params import ( |
||||
load_vqgan_torch_params, |
||||
convert_dalle_bart_torch_from_flax_params |
||||
) |
||||
|
||||
class MinDalleTorch(MinDalleBase): |
||||
def __init__( |
||||
self, |
||||
is_mega: bool, |
||||
is_reusable: bool = True, |
||||
token_count: int = 256 |
||||
): |
||||
print("initializing MinDalleTorch") |
||||
super().__init__(is_mega) |
||||
self.is_reusable = is_reusable |
||||
self.token_count = token_count |
||||
|
||||
if not is_mega: |
||||
self.model_params = load_dalle_bart_flax_params(self.model_path) |
||||
|
||||
self.encoder_params_path = os.path.join(self.model_path, 'encoder.pt') |
||||
self.decoder_params_path = os.path.join(self.model_path, 'decoder.pt') |
||||
|
||||
is_converted = os.path.exists(self.encoder_params_path) |
||||
is_converted &= os.path.exists(self.decoder_params_path) |
||||
if not is_converted: |
||||
convert_and_save_torch_params(is_mega, self.model_path) |
||||
|
||||
if is_reusable: |
||||
self.init_encoder() |
||||
self.init_decoder() |
||||
self.init_detokenizer() |
||||
|
||||
def encode_torch( |
||||
text_tokens: LongTensor, |
||||
config: dict, |
||||
params: dict |
||||
) -> FloatTensor: |
||||
print("loading torch encoder") |
||||
encoder = DalleBartEncoderTorch( |
||||
layer_count = config['encoder_layers'], |
||||
embed_count = config['d_model'], |
||||
attention_head_count = config['encoder_attention_heads'], |
||||
text_vocab_count = config['encoder_vocab_size'], |
||||
text_token_count = config['max_text_length'], |
||||
glu_embed_count = config['encoder_ffn_dim'] |
||||
) |
||||
encoder_params = convert_dalle_bart_torch_from_flax_params( |
||||
params.pop('encoder'), |
||||
layer_count=config['encoder_layers'], |
||||
is_encoder=True |
||||
) |
||||
encoder.load_state_dict(encoder_params, strict=False) |
||||
del encoder_params |
||||
|
||||
print("encoding text tokens") |
||||
encoder_state = encoder(text_tokens) |
||||
del encoder |
||||
return encoder_state |
||||
def init_encoder(self): |
||||
print("initializing DalleBartEncoderTorch") |
||||
self.encoder = DalleBartEncoderTorch( |
||||
layer_count = self.config['encoder_layers'], |
||||
embed_count = self.config['d_model'], |
||||
attention_head_count = self.config['encoder_attention_heads'], |
||||
text_vocab_count = self.config['encoder_vocab_size'], |
||||
text_token_count = self.config['max_text_length'], |
||||
glu_embed_count = self.config['encoder_ffn_dim'] |
||||
) |
||||
params = torch.load(self.encoder_params_path) |
||||
self.encoder.load_state_dict(params, strict=False) |
||||
del params |
||||
if torch.cuda.is_available(): self.encoder = self.encoder.cuda() |
||||
|
||||
|
||||
def decode_torch( |
||||
text_tokens: LongTensor, |
||||
encoder_state: FloatTensor, |
||||
config: dict, |
||||
seed: int, |
||||
params: dict, |
||||
image_token_count: int |
||||
) -> LongTensor: |
||||
print("loading torch decoder") |
||||
decoder = DalleBartDecoderTorch( |
||||
image_vocab_size = config['image_vocab_size'], |
||||
image_token_count = config['image_length'], |
||||
sample_token_count = image_token_count, |
||||
embed_count = config['d_model'], |
||||
attention_head_count = config['decoder_attention_heads'], |
||||
glu_embed_count = config['decoder_ffn_dim'], |
||||
layer_count = config['decoder_layers'], |
||||
batch_count = 2, |
||||
start_token = config['decoder_start_token_id'], |
||||
is_verbose = True |
||||
) |
||||
decoder_params = convert_dalle_bart_torch_from_flax_params( |
||||
params.pop('decoder'), |
||||
layer_count=config['decoder_layers'], |
||||
is_encoder=False |
||||
) |
||||
decoder.load_state_dict(decoder_params, strict=False) |
||||
del decoder_params |
||||
def init_decoder(self): |
||||
print("initializing DalleBartDecoderTorch") |
||||
self.decoder = DalleBartDecoderTorch( |
||||
image_vocab_size = self.config['image_vocab_size'], |
||||
image_token_count = self.config['image_length'], |
||||
sample_token_count = self.token_count, |
||||
embed_count = self.config['d_model'], |
||||
attention_head_count = self.config['decoder_attention_heads'], |
||||
glu_embed_count = self.config['decoder_ffn_dim'], |
||||
layer_count = self.config['decoder_layers'], |
||||
batch_count = 2, |
||||
start_token = self.config['decoder_start_token_id'], |
||||
is_verbose = True |
||||
) |
||||
params = torch.load(self.decoder_params_path) |
||||
self.decoder.load_state_dict(params, strict=False) |
||||
del params |
||||
if torch.cuda.is_available(): self.decoder = self.decoder.cuda() |
||||
|
||||
print("sampling image tokens") |
||||
torch.manual_seed(seed) |
||||
image_tokens = decoder.forward(text_tokens, encoder_state) |
||||
return image_tokens |
||||
|
||||
def init_detokenizer(self): |
||||
super().init_detokenizer() |
||||
if torch.cuda.is_available(): |
||||
self.detokenizer = self.detokenizer.cuda() |
||||
|
||||
|
||||
def generate_image_tokens(self, text: str, seed: int) -> LongTensor: |
||||
text_tokens = self.tokenize_text(text) |
||||
text_tokens = torch.tensor(text_tokens).to(torch.long) |
||||
if torch.cuda.is_available(): text_tokens = text_tokens.cuda() |
||||
|
||||
def generate_image_tokens_torch( |
||||
text_tokens: numpy.ndarray, |
||||
seed: int, |
||||
config: dict, |
||||
params: dict, |
||||
image_token_count: int |
||||
) -> LongTensor: |
||||
text_tokens = torch.tensor(text_tokens).to(torch.long) |
||||
if torch.cuda.is_available(): text_tokens = text_tokens.cuda() |
||||
encoder_state = encode_torch( |
||||
text_tokens, |
||||
config, |
||||
params |
||||
) |
||||
image_tokens = decode_torch( |
||||
text_tokens, |
||||
encoder_state, |
||||
config, |
||||
seed, |
||||
params, |
||||
image_token_count |
||||
) |
||||
return image_tokens |
||||
if not self.is_reusable: self.init_encoder() |
||||
print("encoding text tokens") |
||||
encoder_state = self.encoder.forward(text_tokens) |
||||
if not self.is_reusable: del self.encoder |
||||
|
||||
if not self.is_reusable: self.init_decoder() |
||||
print("sampling image tokens") |
||||
torch.manual_seed(seed) |
||||
image_tokens = self.decoder.forward(text_tokens, encoder_state) |
||||
if not self.is_reusable: del self.decoder |
||||
return image_tokens |
||||
|
||||
|
||||
def detokenize_torch(image_tokens: LongTensor) -> numpy.ndarray: |
||||
print("detokenizing image") |
||||
model_path = './pretrained/vqgan' |
||||
params = load_vqgan_torch_params(model_path) |
||||
detokenizer = VQGanDetokenizer() |
||||
detokenizer.load_state_dict(params) |
||||
image = detokenizer.forward(image_tokens).to(torch.uint8) |
||||
return image.detach().numpy() |
||||
|
||||
def generate_image(self, text: str, seed: int) -> Image.Image: |
||||
image_tokens = self.generate_image_tokens(text, seed) |
||||
if not self.is_reusable: self.init_detokenizer() |
||||
print("detokenizing image") |
||||
image = self.detokenizer.forward(image_tokens).to(torch.uint8) |
||||
if not self.is_reusable: del self.detokenizer |
||||
image = Image.fromarray(image.to('cpu').detach().numpy()) |
||||
return image |
||||
@ -0,0 +1,23 @@ |
||||
import tempfile |
||||
from cog import BasePredictor, Path, Input |
||||
|
||||
from min_dalle.min_dalle_torch import MinDalleTorch |
||||
|
||||
class Predictor(BasePredictor): |
||||
def setup(self): |
||||
self.model = MinDalleTorch(is_mega=True) |
||||
|
||||
def predict( |
||||
self, |
||||
text: str = Input( |
||||
description="Text for generating images.", |
||||
), |
||||
seed: int = Input( |
||||
description="Specify the seed.", |
||||
), |
||||
) -> Path: |
||||
image = self.model.generate_image(text, seed) |
||||
out_path = Path(tempfile.mkdtemp()) / "output.png" |
||||
image.save(str(out_path)) |
||||
|
||||
return out_path |
||||
@ -1,2 +1,3 @@ |
||||
torch |
||||
flax==0.4.2 |
||||
wandb |
||||
|
||||
@ -1,15 +1,15 @@ |
||||
#!/bin/bash |
||||
|
||||
set -e |
||||
|
||||
pip install -r requirements.txt |
||||
|
||||
mkdir -p pretrained |
||||
mkdir -p pretrained/vqgan |
||||
|
||||
# download vqgan |
||||
git lfs install |
||||
git clone https://huggingface.co/dalle-mini/vqgan_imagenet_f16_16384 ./pretrained/vqgan |
||||
curl https://huggingface.co/dalle-mini/vqgan_imagenet_f16_16384/resolve/main/flax_model.msgpack -L --output ./pretrained/vqgan/flax_model.msgpack |
||||
|
||||
# download dalle-mini and dalle mega |
||||
pip install wandb |
||||
python -m wandb login |
||||
python -m wandb login --anonymously |
||||
python -m wandb artifact get --root=./pretrained/dalle_bart_mini dalle-mini/dalle-mini/mini-1:v0 |
||||
python -m wandb artifact get --root=./pretrained/dalle_bart_mega dalle-mini/dalle-mini/mega-1-fp16:v14 |
||||
|
||||
Loading…
Reference in new issue