合并
This commit is contained in:
5
.gitignore
vendored
5
.gitignore
vendored
@@ -7,6 +7,11 @@ node_modules
|
|||||||
.output
|
.output
|
||||||
.env
|
.env
|
||||||
dist
|
dist
|
||||||
|
data
|
||||||
|
|
||||||
# python
|
# python
|
||||||
venv
|
venv
|
||||||
|
outputs
|
||||||
|
*.egg-info
|
||||||
|
*.egg
|
||||||
|
*.pyc
|
||||||
|
|||||||
363
server.py
363
server.py
@@ -213,209 +213,186 @@ def put_watermark(img, wm_encoder=None):
|
|||||||
import time
|
import time
|
||||||
import requests
|
import requests
|
||||||
|
|
||||||
# 获取model, 如果和之前的model不一样，重新加载
|
|
||||||
def get_model(model_name):
|
|
||||||
global model
|
|
||||||
global config
|
|
||||||
global device
|
|
||||||
if model_name != model_name:
|
|
||||||
config = OmegaConf.load(f"{opt.config}")
|
|
||||||
device = torch.device("cuda") if opt.device == "cuda" else torch.device("cpu")
|
|
||||||
model = load_model_from_config(config, f"{opt.ckpt}", device)
|
|
||||||
return model
|
|
||||||
|
|
||||||
# 使用指定的模型和配置文件进行推理一组参数
|
|
||||||
def drawing(model_name):
|
|
||||||
model = get_model(model_name)
|
|
||||||
if opt.plms:
|
|
||||||
sampler = PLMSSampler(model, device=device)
|
|
||||||
elif opt.dpm:
|
|
||||||
sampler = DPMSolverSampler(model, device=device)
|
|
||||||
else:
|
|
||||||
sampler = DDIMSampler(model, device=device)
|
|
||||||
|
|
||||||
|
|
||||||
def main_dev(opt):
|
def main_dev(opt):
|
||||||
|
model_name = '' # 默认模型
|
||||||
|
model = None # 默认模型
|
||||||
|
config = None # 默认配置
|
||||||
|
device = None # 默认设备
|
||||||
while True:
|
while True:
|
||||||
time.sleep(1) # 延时1s执行, 避免cpu占用过高
|
time.sleep(2) # 延时1s执行, 避免cpu占用过高
|
||||||
# 从局域网中获取一组参数
|
data = requests.get("http://localhost:3000/api/drawing").json() # 从局域网中获取一组参数
|
||||||
request = requests.get("http://localhost:3000/api/drawing")
|
print(data) # [{'model': '768-v-ema', 'prompt': '一只猫', 'watermark': '0'}, {'model': '768-v-ema', 'prompt': '一只狗', 'watermark': '0'}]
|
||||||
if request.status_code == 200:
|
# 遍历 data 返回dict
|
||||||
data = request.json()
|
for item in data:
|
||||||
print("data: ", data)
|
print(item) # {'model': '768-v-ema', 'prompt': '一只猫', 'watermark': '0'}
|
||||||
#drawing("model_name")
|
# 设置参数
|
||||||
|
if 'prompt' in item: opt.prompt = item['prompt'] # 描述
|
||||||
|
if 'n_samples' in item: opt.n_samples = item['n_samples'] # 列数
|
||||||
|
if 'n_rows' in item: opt.n_rows = item['n_rows'] # 行数
|
||||||
|
if 'scale' in item: opt.scale = item['scale'] # 比例
|
||||||
|
|
||||||
def main(opt):
|
# 如果模型不同，重新加载模型(注意释放内存)
|
||||||
seed_everything(opt.seed)
|
if item['model'] != model_name:
|
||||||
|
# 获取环境配置
|
||||||
config = OmegaConf.load(f"{opt.config}")
|
model_name = item['model']
|
||||||
device = torch.device("cuda") if opt.device == "cuda" else torch.device("cpu")
|
opt.config = f'/data/{model_name}.yaml'
|
||||||
model = load_model_from_config(config, f"{opt.ckpt}", device)
|
opt.ckpt = f'/data/{model_name}.ckpt'
|
||||||
|
opt.device = 'cuda'
|
||||||
if opt.plms:
|
print(f"config: {opt.config}", f"ckpt: {opt.ckpt}", f"device: {opt.device}")
|
||||||
sampler = PLMSSampler(model, device=device)
|
config = OmegaConf.load(f"{opt.config}")
|
||||||
elif opt.dpm:
|
device = torch.device("cuda") if opt.device == "cuda" else torch.device("cpu")
|
||||||
sampler = DPMSolverSampler(model, device=device)
|
# 加载模型(到显存)
|
||||||
else:
|
print(f"load model: {item['model']}..")
|
||||||
sampler = DDIMSampler(model, device=device)
|
model_name = item['model']
|
||||||
|
model = load_model_from_config(config, f"{opt.ckpt}", device)
|
||||||
os.makedirs(opt.outdir, exist_ok=True)
|
print(f"model_name: {model_name}")
|
||||||
outpath = opt.outdir
|
# 使用指定的模型和配置文件进行推理一组参数
|
||||||
|
if opt.plms:
|
||||||
print("Creating invisible watermark encoder (see https://github.com/ShieldMnt/invisible-watermark)...")
|
sampler = PLMSSampler(model, device=device)
|
||||||
wm = "SDV2"
|
elif opt.dpm:
|
||||||
wm_encoder = WatermarkEncoder()
|
sampler = DPMSolverSampler(model, device=device)
|
||||||
wm_encoder.set_watermark('bytes', wm.encode('utf-8'))
|
else:
|
||||||
|
sampler = DDIMSampler(model, device=device)
|
||||||
batch_size = opt.n_samples
|
# 检查输出目录是否存在
|
||||||
n_rows = opt.n_rows if opt.n_rows > 0 else batch_size
|
os.makedirs(opt.outdir, exist_ok=True)
|
||||||
if not opt.from_file:
|
outpath = opt.outdir
|
||||||
prompt = opt.prompt
|
# 创建水印编码器
|
||||||
assert prompt is not None
|
wm = "SDV2"
|
||||||
data = [batch_size * [prompt]]
|
wm_encoder = WatermarkEncoder()
|
||||||
|
wm_encoder.set_watermark('bytes', wm.encode('utf-8'))
|
||||||
else:
|
# x
|
||||||
print(f"reading prompts from {opt.from_file}")
|
batch_size = opt.n_samples
|
||||||
with open(opt.from_file, "r") as f:
|
n_rows = opt.n_rows if opt.n_rows > 0 else batch_size
|
||||||
data = f.read().splitlines()
|
if not opt.from_file:
|
||||||
data = [p for p in data for i in range(opt.repeat)]
|
prompt = opt.prompt
|
||||||
data = list(chunk(data, batch_size))
|
assert prompt is not None
|
||||||
|
data = [batch_size * [prompt]]
|
||||||
sample_path = os.path.join(outpath, "samples")
|
else:
|
||||||
os.makedirs(sample_path, exist_ok=True)
|
print(f"reading prompts from {opt.from_file}")
|
||||||
sample_count = 0
|
with open(opt.from_file, "r") as f:
|
||||||
base_count = len(os.listdir(sample_path))
|
data = f.read().splitlines()
|
||||||
grid_count = len(os.listdir(outpath)) - 1
|
data = [p for p in data for i in range(opt.repeat)]
|
||||||
|
data = list(chunk(data, batch_size))
|
||||||
start_code = None
|
# x
|
||||||
if opt.fixed_code:
|
sample_path = os.path.join(outpath, "samples")
|
||||||
start_code = torch.randn([opt.n_samples, opt.C, opt.H // opt.f, opt.W // opt.f], device=device)
|
os.makedirs(sample_path, exist_ok=True)
|
||||||
|
sample_count = 0
|
||||||
if opt.torchscript or opt.ipex:
|
base_count = len(os.listdir(sample_path))
|
||||||
transformer = model.cond_stage_model.model
|
grid_count = len(os.listdir(outpath)) - 1
|
||||||
unet = model.model.diffusion_model
|
# x
|
||||||
decoder = model.first_stage_model.decoder
|
start_code = None
|
||||||
additional_context = torch.cpu.amp.autocast() if opt.bf16 else nullcontext()
|
if opt.fixed_code:
|
||||||
shape = [opt.C, opt.H // opt.f, opt.W // opt.f]
|
start_code = torch.randn([opt.n_samples, opt.C, opt.H // opt.f, opt.W // opt.f], device=device)
|
||||||
|
# x
|
||||||
if opt.bf16 and not opt.torchscript and not opt.ipex:
|
if opt.torchscript or opt.ipex:
|
||||||
raise ValueError('Bfloat16 is supported only for torchscript+ipex')
|
transformer = model.cond_stage_model.model
|
||||||
if opt.bf16 and unet.dtype != torch.bfloat16:
|
unet = model.model.diffusion_model
|
||||||
raise ValueError("Use configs/stable-diffusion/intel/ configs with bf16 enabled if " +
|
decoder = model.first_stage_model.decoder
|
||||||
"you'd like to use bfloat16 with CPU.")
|
additional_context = torch.cpu.amp.autocast() if opt.bf16 else nullcontext()
|
||||||
if unet.dtype == torch.float16 and device == torch.device("cpu"):
|
shape = [opt.C, opt.H // opt.f, opt.W // opt.f]
|
||||||
raise ValueError("Use configs/stable-diffusion/intel/ configs for your model if you'd like to run it on CPU.")
|
if opt.bf16 and not opt.torchscript and not opt.ipex:
|
||||||
|
raise ValueError('Bfloat16 is supported only for torchscript+ipex')
|
||||||
if opt.ipex:
|
if opt.bf16 and unet.dtype != torch.bfloat16:
|
||||||
import intel_extension_for_pytorch as ipex
|
raise ValueError("Use configs/stable-diffusion/intel/ configs with bf16 enabled if " +
|
||||||
bf16_dtype = torch.bfloat16 if opt.bf16 else None
|
"you'd like to use bfloat16 with CPU.")
|
||||||
transformer = transformer.to(memory_format=torch.channels_last)
|
if unet.dtype == torch.float16 and device == torch.device("cpu"):
|
||||||
transformer = ipex.optimize(transformer, level="O1", inplace=True)
|
raise ValueError("Use configs/stable-diffusion/intel/ configs for your model if you'd like to run it on CPU.")
|
||||||
|
if opt.ipex:
|
||||||
unet = unet.to(memory_format=torch.channels_last)
|
import intel_extension_for_pytorch as ipex
|
||||||
unet = ipex.optimize(unet, level="O1", auto_kernel_selection=True, inplace=True, dtype=bf16_dtype)
|
bf16_dtype = torch.bfloat16 if opt.bf16 else None
|
||||||
|
transformer = transformer.to(memory_format=torch.channels_last)
|
||||||
decoder = decoder.to(memory_format=torch.channels_last)
|
transformer = ipex.optimize(transformer, level="O1", inplace=True)
|
||||||
decoder = ipex.optimize(decoder, level="O1", auto_kernel_selection=True, inplace=True, dtype=bf16_dtype)
|
unet = unet.to(memory_format=torch.channels_last)
|
||||||
|
unet = ipex.optimize(unet, level="O1", auto_kernel_selection=True, inplace=True, dtype=bf16_dtype)
|
||||||
if opt.torchscript:
|
decoder = decoder.to(memory_format=torch.channels_last)
|
||||||
with torch.no_grad(), additional_context:
|
decoder = ipex.optimize(decoder, level="O1", auto_kernel_selection=True, inplace=True, dtype=bf16_dtype)
|
||||||
# get UNET scripted
|
if opt.torchscript:
|
||||||
if unet.use_checkpoint:
|
with torch.no_grad(), additional_context:
|
||||||
raise ValueError("Gradient checkpoint won't work with tracing. " +
|
# get UNET scripted
|
||||||
"Use configs/stable-diffusion/intel/ configs for your model or disable checkpoint in your config.")
|
if unet.use_checkpoint:
|
||||||
|
raise ValueError("Gradient checkpoint won't work with tracing. " +
|
||||||
img_in = torch.ones(2, 4, 96, 96, dtype=torch.float32)
|
"Use configs/stable-diffusion/intel/ configs for your model or disable checkpoint in your config.")
|
||||||
t_in = torch.ones(2, dtype=torch.int64)
|
img_in = torch.ones(2, 4, 96, 96, dtype=torch.float32)
|
||||||
context = torch.ones(2, 77, 1024, dtype=torch.float32)
|
t_in = torch.ones(2, dtype=torch.int64)
|
||||||
scripted_unet = torch.jit.trace(unet, (img_in, t_in, context))
|
context = torch.ones(2, 77, 1024, dtype=torch.float32)
|
||||||
scripted_unet = torch.jit.optimize_for_inference(scripted_unet)
|
scripted_unet = torch.jit.trace(unet, (img_in, t_in, context))
|
||||||
print(type(scripted_unet))
|
scripted_unet = torch.jit.optimize_for_inference(scripted_unet)
|
||||||
model.model.scripted_diffusion_model = scripted_unet
|
print(type(scripted_unet))
|
||||||
|
model.model.scripted_diffusion_model = scripted_unet
|
||||||
# get Decoder for first stage model scripted
|
# get Decoder for first stage model scripted
|
||||||
samples_ddim = torch.ones(1, 4, 96, 96, dtype=torch.float32)
|
samples_ddim = torch.ones(1, 4, 96, 96, dtype=torch.float32)
|
||||||
scripted_decoder = torch.jit.trace(decoder, (samples_ddim))
|
scripted_decoder = torch.jit.trace(decoder, (samples_ddim))
|
||||||
scripted_decoder = torch.jit.optimize_for_inference(scripted_decoder)
|
scripted_decoder = torch.jit.optimize_for_inference(scripted_decoder)
|
||||||
print(type(scripted_decoder))
|
print(type(scripted_decoder))
|
||||||
model.first_stage_model.decoder = scripted_decoder
|
model.first_stage_model.decoder = scripted_decoder
|
||||||
|
prompts = data[0]
|
||||||
prompts = data[0]
|
print("Running a forward pass to initialize optimizations")
|
||||||
print("Running a forward pass to initialize optimizations")
|
uc = None
|
||||||
uc = None
|
if opt.scale != 1.0:
|
||||||
if opt.scale != 1.0:
|
uc = model.get_learned_conditioning(batch_size * [""])
|
||||||
uc = model.get_learned_conditioning(batch_size * [""])
|
if isinstance(prompts, tuple):
|
||||||
if isinstance(prompts, tuple):
|
prompts = list(prompts)
|
||||||
prompts = list(prompts)
|
with torch.no_grad(), additional_context:
|
||||||
|
for _ in range(3):
|
||||||
with torch.no_grad(), additional_context:
|
c = model.get_learned_conditioning(prompts)
|
||||||
for _ in range(3):
|
samples_ddim, _ = sampler.sample(S=5,
|
||||||
c = model.get_learned_conditioning(prompts)
|
|
||||||
samples_ddim, _ = sampler.sample(S=5,
|
|
||||||
conditioning=c,
|
|
||||||
batch_size=batch_size,
|
|
||||||
shape=shape,
|
|
||||||
verbose=False,
|
|
||||||
unconditional_guidance_scale=opt.scale,
|
|
||||||
unconditional_conditioning=uc,
|
|
||||||
eta=opt.ddim_eta,
|
|
||||||
x_T=start_code)
|
|
||||||
print("Running a forward pass for decoder")
|
|
||||||
for _ in range(3):
|
|
||||||
x_samples_ddim = model.decode_first_stage(samples_ddim)
|
|
||||||
|
|
||||||
precision_scope = autocast if opt.precision=="autocast" or opt.bf16 else nullcontext
|
|
||||||
with torch.no_grad(), \
|
|
||||||
precision_scope(opt.device), \
|
|
||||||
model.ema_scope():
|
|
||||||
all_samples = list()
|
|
||||||
for n in trange(opt.n_iter, desc="Sampling"):
|
|
||||||
for prompts in tqdm(data, desc="data"):
|
|
||||||
uc = None
|
|
||||||
if opt.scale != 1.0:
|
|
||||||
uc = model.get_learned_conditioning(batch_size * [""])
|
|
||||||
if isinstance(prompts, tuple):
|
|
||||||
prompts = list(prompts)
|
|
||||||
c = model.get_learned_conditioning(prompts)
|
|
||||||
shape = [opt.C, opt.H // opt.f, opt.W // opt.f]
|
|
||||||
samples, _ = sampler.sample(S=opt.steps,
|
|
||||||
conditioning=c,
|
conditioning=c,
|
||||||
batch_size=opt.n_samples,
|
batch_size=batch_size,
|
||||||
shape=shape,
|
shape=shape,
|
||||||
verbose=False,
|
verbose=False,
|
||||||
unconditional_guidance_scale=opt.scale,
|
unconditional_guidance_scale=opt.scale,
|
||||||
unconditional_conditioning=uc,
|
unconditional_conditioning=uc,
|
||||||
eta=opt.ddim_eta,
|
eta=opt.ddim_eta,
|
||||||
x_T=start_code)
|
x_T=start_code)
|
||||||
|
print("Running a forward pass for decoder")
|
||||||
x_samples = model.decode_first_stage(samples)
|
for _ in range(3):
|
||||||
x_samples = torch.clamp((x_samples + 1.0) / 2.0, min=0.0, max=1.0)
|
x_samples_ddim = model.decode_first_stage(samples_ddim)
|
||||||
|
precision_scope = autocast if opt.precision=="autocast" or opt.bf16 else nullcontext
|
||||||
for x_sample in x_samples:
|
with torch.no_grad(), \
|
||||||
x_sample = 255. * rearrange(x_sample.cpu().numpy(), 'c h w -> h w c')
|
precision_scope(opt.device), \
|
||||||
img = Image.fromarray(x_sample.astype(np.uint8))
|
model.ema_scope():
|
||||||
img = put_watermark(img, wm_encoder)
|
all_samples = list()
|
||||||
img.save(os.path.join(sample_path, f"{base_count:05}.png"))
|
for n in trange(opt.n_iter, desc="Sampling"):
|
||||||
base_count += 1
|
for prompts in tqdm(data, desc="data"):
|
||||||
sample_count += 1
|
uc = None
|
||||||
|
if opt.scale != 1.0:
|
||||||
all_samples.append(x_samples)
|
uc = model.get_learned_conditioning(batch_size * [""])
|
||||||
|
if isinstance(prompts, tuple):
|
||||||
# additionally, save as grid
|
prompts = list(prompts)
|
||||||
grid = torch.stack(all_samples, 0)
|
c = model.get_learned_conditioning(prompts)
|
||||||
grid = rearrange(grid, 'n b c h w -> (n b) c h w')
|
shape = [opt.C, opt.H // opt.f, opt.W // opt.f]
|
||||||
grid = make_grid(grid, nrow=n_rows)
|
samples, _ = sampler.sample(S=opt.steps,
|
||||||
|
conditioning=c,
|
||||||
# to image
|
batch_size=opt.n_samples,
|
||||||
grid = 255. * rearrange(grid, 'c h w -> h w c').cpu().numpy()
|
shape=shape,
|
||||||
grid = Image.fromarray(grid.astype(np.uint8))
|
verbose=False,
|
||||||
grid = put_watermark(grid, wm_encoder)
|
unconditional_guidance_scale=opt.scale,
|
||||||
grid.save(os.path.join(outpath, f'grid-{grid_count:04}.png'))
|
unconditional_conditioning=uc,
|
||||||
grid_count += 1
|
eta=opt.ddim_eta,
|
||||||
|
x_T=start_code)
|
||||||
print(f"Your samples are ready and waiting for you here: \n{outpath} \n"
|
x_samples = model.decode_first_stage(samples)
|
||||||
f" \nEnjoy.")
|
x_samples = torch.clamp((x_samples + 1.0) / 2.0, min=0.0, max=1.0)
|
||||||
|
for x_sample in x_samples:
|
||||||
|
x_sample = 255. * rearrange(x_sample.cpu().numpy(), 'c h w -> h w c')
|
||||||
|
img = Image.fromarray(x_sample.astype(np.uint8))
|
||||||
|
img = put_watermark(img, wm_encoder)
|
||||||
|
img.save(os.path.join(sample_path, f"{base_count:05}.png"))
|
||||||
|
base_count += 1
|
||||||
|
sample_count += 1
|
||||||
|
all_samples.append(x_samples)
|
||||||
|
# additionally, save as grid
|
||||||
|
grid = torch.stack(all_samples, 0)
|
||||||
|
grid = rearrange(grid, 'n b c h w -> (n b) c h w')
|
||||||
|
grid = make_grid(grid, nrow=n_rows)
|
||||||
|
# to image
|
||||||
|
grid = 255. * rearrange(grid, 'c h w -> h w c').cpu().numpy()
|
||||||
|
grid = Image.fromarray(grid.astype(np.uint8))
|
||||||
|
grid = put_watermark(grid, wm_encoder)
|
||||||
|
grid.save(os.path.join(outpath, f'grid-{grid_count:04}.png'))
|
||||||
|
grid_count += 1
|
||||||
|
print(f"Your samples are ready and waiting for you here: \n{outpath} \n"
|
||||||
|
f" \nEnjoy.")
|
||||||
|
break
|
||||||
|
|
||||||
if __name__ == "__main__":
|
if __name__ == "__main__":
|
||||||
opt = parse_args()
|
opt = parse_args()
|
||||||
|
|||||||
Reference in New Issue
Block a user