min-dalle-test/min_dalle/min_dalle_torch.py

114 lines
4.2 KiB
Python
Raw Normal View History

2022-06-29 02:22:54 +00:00
import os
from PIL import Image
2022-06-28 16:16:44 +00:00
from typing import Dict
from torch import LongTensor
2022-06-27 15:57:56 +00:00
import torch
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 .load_params import (
convert_and_save_mega_torch_params,
load_dalle_bart_flax_params
)
from .min_dalle_base import MinDalleBase
2022-06-27 19:46:04 +00:00
from .models.dalle_bart_encoder_torch import DalleBartEncoderTorch
from .models.dalle_bart_decoder_torch import DalleBartDecoderTorch
2022-06-27 15:57:56 +00:00
class MinDalleTorch(MinDalleBase):
def __init__(
self,
is_mega: bool,
2022-06-30 15:25:24 +00:00
is_reusable: bool = True,
token_count: int = 256
):
print("initializing MinDalleTorch")
super().__init__(is_mega)
2022-06-30 15:25:24 +00:00
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_mega_torch_params(is_mega, self.model_path)
2022-06-30 15:25:24 +00:00
if is_reusable:
self.init_encoder()
self.init_decoder()
self.init_detokenizer()
2022-06-27 15:57:56 +00:00
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
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):
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
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
def init_detokenizer(self):
super().init_detokenizer()
if torch.cuda.is_available():
self.detokenizer = self.detokenizer.cuda()
2022-06-27 15:57:56 +00:00
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()
2022-06-27 15:57:56 +00:00
2022-06-30 15:25:24 +00:00
if not self.is_reusable: self.init_encoder()
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
2022-06-27 15:57:56 +00:00
2022-06-30 15:25:24 +00:00
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)
2022-06-30 15:25:24 +00:00
if not self.is_reusable: del self.decoder
return image_tokens
2022-06-27 15:57:56 +00:00
def generate_image(self, text: str, seed: int) -> Image.Image:
image_tokens = self.generate_image_tokens(text, seed)
2022-06-30 15:25:24 +00:00
if not self.is_reusable: self.init_detokenizer()
print("detokenizing image")
image = self.detokenizer.forward(image_tokens).to(torch.uint8)
2022-06-30 15:25:24 +00:00
if not self.is_reusable: del self.detokenizer
image = Image.fromarray(image.to('cpu').detach().numpy())
return image