81 lines
2.8 KiB
Python
81 lines
2.8 KiB
Python
import torch
|
||
|
||
import fcbh.model_management as model_management
|
||
|
||
from transformers import AutoTokenizer, AutoModelForCausalLM, set_seed
|
||
from modules.path import fooocus_expansion_path
|
||
from fcbh.model_patcher import ModelPatcher
|
||
|
||
# limitation of np.random.seed(), called from transformers.set_seed()
|
||
SEED_LIMIT_NUMPY = 2**32
|
||
|
||
|
||
fooocus_magic_split = [
|
||
', extremely',
|
||
', intricate,',
|
||
]
|
||
dangrous_patterns = '[]【】()（）|:：'
|
||
|
||
|
||
def safe_str(x):
|
||
x = str(x)
|
||
for _ in range(16):
|
||
x = x.replace(' ', ' ')
|
||
return x.strip(",. \r\n")
|
||
|
||
|
||
def remove_pattern(x, pattern):
|
||
for p in pattern:
|
||
x = x.replace(p, '')
|
||
return x
|
||
|
||
|
||
class FooocusExpansion:
|
||
def __init__(self):
|
||
self.tokenizer = AutoTokenizer.from_pretrained(fooocus_expansion_path)
|
||
self.model = AutoModelForCausalLM.from_pretrained(fooocus_expansion_path)
|
||
self.model.eval()
|
||
|
||
load_device = model_management.text_encoder_device()
|
||
offload_device = model_management.text_encoder_offload_device()
|
||
|
||
# MPS hack
|
||
if model_management.is_device_mps(load_device):
|
||
load_device = torch.device('cpu')
|
||
offload_device = torch.device('cpu')
|
||
|
||
use_fp16 = model_management.should_use_fp16(device=load_device)
|
||
|
||
if use_fp16:
|
||
self.model.half()
|
||
|
||
self.patcher = ModelPatcher(self.model, load_device=load_device, offload_device=offload_device)
|
||
print(f'Fooocus Expansion engine loaded for {load_device}, use_fp16 = {use_fp16}.')
|
||
|
||
def __call__(self, prompt, seed):
|
||
if self.patcher.current_device != self.patcher.load_device:
|
||
print('Fooocus Expansion loaded by itself.')
|
||
model_management.load_model_gpu(self.patcher)
|
||
|
||
seed = int(seed) % SEED_LIMIT_NUMPY
|
||
set_seed(seed)
|
||
origin = safe_str(prompt)
|
||
prompt = origin + fooocus_magic_split[seed % len(fooocus_magic_split)]
|
||
|
||
tokenized_kwargs = self.tokenizer(prompt, return_tensors="pt")
|
||
tokenized_kwargs.data['input_ids'] = tokenized_kwargs.data['input_ids'].to(self.patcher.load_device)
|
||
tokenized_kwargs.data['attention_mask'] = tokenized_kwargs.data['attention_mask'].to(self.patcher.load_device)
|
||
|
||
# https://huggingface.co/blog/introducing-csearch
|
||
# https://huggingface.co/docs/transformers/generation_strategies
|
||
features = self.model.generate(**tokenized_kwargs,
|
||
num_beams=1,
|
||
max_new_tokens=256,
|
||
do_sample=True)
|
||
|
||
response = self.tokenizer.batch_decode(features, skip_special_tokens=True)
|
||
result = response[0][len(origin):]
|
||
result = safe_str(result)
|
||
result = remove_pattern(result, dangrous_patterns)
|
||
return result
|