simplified
This commit is contained in:
Brett Kuprel
@@ -1,70 +0,0 @@
|
||||
import os
|
||||
import json
|
||||
import numpy
|
||||
import torch
|
||||
from PIL import Image
|
||||
from typing import Tuple, List
|
||||
|
||||
from .text_tokenizer import TextTokenizer
|
||||
from .models.vqgan_detokenizer import VQGanDetokenizer
|
||||
from .load_params import load_vqgan_torch_params
|
||||
|
||||
|
||||
def load_dalle_bart_metadata(path: str) -> Tuple[dict, dict, List[str]]:
|
||||
print("loading model")
|
||||
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 ascii_from_image(image: Image.Image, size: int) -> str:
|
||||
rgb_pixels = image.resize((size, int(0.55 * size))).convert('L').getdata()
|
||||
chars = list('.,;/IOX')
|
||||
chars = [chars[i * len(chars) // 256] for i in rgb_pixels]
|
||||
chars = [chars[i * size: (i + 1) * size] for i in range(size // 2)]
|
||||
return '\n'.join(''.join(row) for row in chars)
|
||||
|
||||
|
||||
def save_image(image: numpy.ndarray, path: str) -> Image.Image:
|
||||
if os.path.isdir(path):
|
||||
path = os.path.join(path, 'generated.png')
|
||||
elif not path.endswith('.png'):
|
||||
path += '.png'
|
||||
print("saving image to", path)
|
||||
image: Image.Image = Image.fromarray(numpy.asarray(image))
|
||||
image.save(path)
|
||||
return image
|
||||
|
||||
|
||||
def tokenize(
|
||||
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 detokenize_torch(
|
||||
image_tokens: numpy.ndarray,
|
||||
model_path: str
|
||||
) -> numpy.ndarray:
|
||||
print("detokenizing image")
|
||||
params = load_vqgan_torch_params(model_path)
|
||||
detokenizer = VQGanDetokenizer()
|
||||
detokenizer.load_state_dict(params)
|
||||
image_tokens = torch.tensor(image_tokens).to(torch.long)
|
||||
image = detokenizer.forward(image_tokens).to(torch.uint8)
|
||||
return image.detach().numpy()
|
||||
@@ -1,50 +1,9 @@
|
||||
import jax
|
||||
from jax import numpy as jnp
|
||||
import numpy
|
||||
import argparse
|
||||
|
||||
from load_params import load_dalle_bart_flax_params
|
||||
from image_from_text import (
|
||||
load_dalle_bart_metadata,
|
||||
tokenize,
|
||||
detokenize_torch,
|
||||
save_image,
|
||||
ascii_from_image
|
||||
)
|
||||
from models.dalle_bart_encoder_flax import DalleBartEncoderFlax
|
||||
from models.dalle_bart_decoder_flax import DalleBartDecoderFlax
|
||||
|
||||
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument(
|
||||
'--text',
|
||||
help='text to generate image from',
|
||||
type=str
|
||||
)
|
||||
parser.add_argument(
|
||||
'--seed',
|
||||
help='random seed',
|
||||
type=int,
|
||||
default=0
|
||||
)
|
||||
parser.add_argument(
|
||||
'--image_path',
|
||||
help='generated image path',
|
||||
type=str,
|
||||
default='generated.png'
|
||||
)
|
||||
parser.add_argument(
|
||||
'--dalle_bart_path',
|
||||
help='pretraied dalle bart path',
|
||||
type=str,
|
||||
default='./pretrained/dalle_bart_mini'
|
||||
)
|
||||
parser.add_argument(
|
||||
'--vqgan_path',
|
||||
help='pretraied vqgan path',
|
||||
type=str,
|
||||
default='./pretrained/vqgan'
|
||||
)
|
||||
from .models.dalle_bart_encoder_flax import DalleBartEncoderFlax
|
||||
from .models.dalle_bart_decoder_flax import DalleBartDecoderFlax
|
||||
|
||||
|
||||
def encode_flax(
|
||||
@@ -67,6 +26,7 @@ def encode_flax(
|
||||
del encoder
|
||||
return encoder_state
|
||||
|
||||
|
||||
def decode_flax(
|
||||
text_tokens: jnp.ndarray,
|
||||
encoder_state: jnp.ndarray,
|
||||
@@ -95,32 +55,25 @@ def decode_flax(
|
||||
del decoder
|
||||
return image_tokens
|
||||
|
||||
|
||||
def generate_image_tokens_flax(
|
||||
text: str,
|
||||
seed: int,
|
||||
dalle_bart_path: str
|
||||
text_tokens: numpy.ndarray,
|
||||
seed: int,
|
||||
config: dict,
|
||||
params: dict
|
||||
) -> numpy.ndarray:
|
||||
config, vocab, merges = load_dalle_bart_metadata(dalle_bart_path)
|
||||
text_tokens = tokenize(text, config, vocab, merges)
|
||||
params_dalle_bart = load_dalle_bart_flax_params(dalle_bart_path)
|
||||
encoder_state = encode_flax(text_tokens, config, params_dalle_bart)
|
||||
encoder_state = encode_flax(
|
||||
text_tokens,
|
||||
config,
|
||||
params
|
||||
)
|
||||
image_tokens = decode_flax(
|
||||
text_tokens,
|
||||
encoder_state,
|
||||
config, seed,
|
||||
params_dalle_bart
|
||||
)
|
||||
return numpy.array(image_tokens)
|
||||
|
||||
if __name__ == '__main__':
|
||||
args = parser.parse_args()
|
||||
|
||||
image_tokens = generate_image_tokens_flax(
|
||||
args.text,
|
||||
args.seed,
|
||||
args.dalle_bart_path
|
||||
config,
|
||||
seed,
|
||||
params
|
||||
)
|
||||
image_tokens = numpy.array(image_tokens)
|
||||
print("image tokens", list(image_tokens))
|
||||
image = detokenize_torch(image_tokens, args.vqgan_path)
|
||||
image = save_image(image, args.image_path)
|
||||
print(ascii_from_image(image, size=128))
|
||||
return image_tokens
|
||||
113
min_dalle/min_dalle_torch.py
Normal file
113
min_dalle/min_dalle_torch.py
Normal file
@@ -0,0 +1,113 @@
|
||||
import numpy
|
||||
import torch
|
||||
from torch import Tensor
|
||||
from typing import Dict
|
||||
|
||||
from .models.vqgan_detokenizer import VQGanDetokenizer
|
||||
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
|
||||
)
|
||||
|
||||
|
||||
def encode_torch(
|
||||
text_tokens: numpy.ndarray,
|
||||
config: dict,
|
||||
params: dict
|
||||
) -> Tensor:
|
||||
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")
|
||||
text_tokens = torch.tensor(text_tokens).to(torch.long)
|
||||
encoder_state = encoder(text_tokens)
|
||||
del encoder
|
||||
return encoder_state
|
||||
|
||||
|
||||
def decode_torch(
|
||||
text_tokens: Tensor,
|
||||
encoder_state: Tensor,
|
||||
config: dict,
|
||||
seed: int,
|
||||
params: dict,
|
||||
image_token_count: int
|
||||
) -> Tensor:
|
||||
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
|
||||
|
||||
print("sampling image tokens")
|
||||
torch.manual_seed(seed)
|
||||
text_tokens = torch.tensor(text_tokens).to(torch.long)
|
||||
image_tokens = decoder.forward(text_tokens, encoder_state)
|
||||
return image_tokens
|
||||
|
||||
|
||||
def generate_image_tokens_torch(
|
||||
text_tokens: numpy.ndarray,
|
||||
seed: int,
|
||||
config: dict,
|
||||
params: dict,
|
||||
image_token_count: int
|
||||
) -> numpy.ndarray:
|
||||
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.detach().numpy()
|
||||
|
||||
|
||||
def detokenize_torch(image_tokens: numpy.ndarray) -> numpy.ndarray:
|
||||
print("detokenizing image")
|
||||
model_path = './pretrained/vqgan'
|
||||
params = load_vqgan_torch_params(model_path)
|
||||
detokenizer = VQGanDetokenizer()
|
||||
detokenizer.load_state_dict(params)
|
||||
image_tokens = torch.tensor(image_tokens).to(torch.long)
|
||||
image = detokenizer.forward(image_tokens).to(torch.uint8)
|
||||
return image.detach().numpy()
|
||||
|
||||
Reference in New Issue
Block a user