wehub-resource-sync
279 lines
11 KiB
Python
279 lines
11 KiB
Python
# Copyright (c) 2023 PaddlePaddle Authors. All Rights Reserved.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
|
|
import os
|
|
import random
|
|
|
|
import numpy as np
|
|
import paddle
|
|
import paddle.nn as nn
|
|
import paddle.nn.functional as F
|
|
from datasets import DatasetDict, concatenate_datasets
|
|
from paddle.io import BatchSampler, DataLoader, DistributedBatchSampler
|
|
from paddle.vision import BaseTransform, transforms
|
|
|
|
from paddlenlp.transformers import CLIPTextModel, CLIPTokenizer
|
|
from paddlenlp.utils.downloader import get_path_from_url_with_filelock
|
|
from paddlenlp.utils.log import logger
|
|
|
|
from .model_base import BenchmarkBase
|
|
|
|
|
|
def freeze_params(params):
|
|
for param in params:
|
|
param.stop_gradient = True
|
|
|
|
|
|
def url_or_path_join(*path_list):
|
|
return os.path.join(*path_list) if os.path.isdir(os.path.join(*path_list)) else "/".join(path_list)
|
|
|
|
|
|
class Lambda(BaseTransform):
|
|
def __init__(self, fn, keys=None):
|
|
super().__init__(keys)
|
|
self.fn = fn
|
|
|
|
def _apply_image(self, img):
|
|
return self.fn(img)
|
|
|
|
|
|
class StableDiffusion(nn.Layer):
|
|
def __init__(self, args):
|
|
super().__init__()
|
|
from ppdiffusers import AutoencoderKL, DDPMScheduler, UNet2DConditionModel
|
|
|
|
self.args = args
|
|
self.unet = UNet2DConditionModel.from_pretrained(url_or_path_join(args.model_name_or_path, "unet"))
|
|
self.vae = AutoencoderKL.from_pretrained(url_or_path_join(args.model_name_or_path, "vae"))
|
|
self.text_encoder = CLIPTextModel.from_pretrained(url_or_path_join(args.model_name_or_path, "text_encoder"))
|
|
# we only use self.noise_scheduler.alphas_cumprod
|
|
self.noise_scheduler = DDPMScheduler.from_pretrained(url_or_path_join(args.model_name_or_path, "scheduler"))
|
|
self.register_buffer("alphas_cumprod", self.noise_scheduler.alphas_cumprod)
|
|
freeze_params(self.vae.parameters())
|
|
freeze_params(self.text_encoder.parameters())
|
|
self.unet.train()
|
|
self.vae.eval()
|
|
self.text_encoder.eval()
|
|
if args.use_amp and args.amp_level == "O2":
|
|
self.vae.to(dtype=paddle.float16)
|
|
self.text_encoder.to(dtype=paddle.float16)
|
|
|
|
def add_noise(
|
|
self,
|
|
original_samples: paddle.Tensor,
|
|
noise: paddle.Tensor,
|
|
timesteps: paddle.Tensor,
|
|
) -> paddle.Tensor:
|
|
sqrt_alpha_prod = self.alphas_cumprod[timesteps] ** 0.5
|
|
sqrt_alpha_prod = sqrt_alpha_prod.flatten()
|
|
while len(sqrt_alpha_prod.shape) < len(original_samples.shape):
|
|
sqrt_alpha_prod = sqrt_alpha_prod.unsqueeze(-1)
|
|
|
|
sqrt_one_minus_alpha_prod = (1 - self.alphas_cumprod[timesteps]) ** 0.5
|
|
sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.flatten()
|
|
while len(sqrt_one_minus_alpha_prod.shape) < len(original_samples.shape):
|
|
sqrt_one_minus_alpha_prod = sqrt_one_minus_alpha_prod.unsqueeze(-1)
|
|
|
|
noisy_samples = sqrt_alpha_prod * original_samples + sqrt_one_minus_alpha_prod * noise
|
|
return noisy_samples
|
|
|
|
def forward(self, input_ids=None, pixel_values=None):
|
|
with paddle.no_grad():
|
|
latents = self.vae.encode(pixel_values).latent_dist.sample()
|
|
latents = latents * 0.18215
|
|
noise = paddle.randn(latents.shape)
|
|
timesteps = paddle.randint(0, self.noise_scheduler.num_train_timesteps, (latents.shape[0],))
|
|
noisy_latents = self.add_noise(latents, noise, timesteps)
|
|
|
|
encoder_hidden_states = self.text_encoder(input_ids)[0]
|
|
model_pred = self.unet(noisy_latents, timesteps, encoder_hidden_states).sample
|
|
|
|
# Get the target for loss depending on the prediction type
|
|
if self.noise_scheduler.config.prediction_type == "epsilon":
|
|
target = noise
|
|
elif self.noise_scheduler.config.prediction_type == "v_prediction":
|
|
target = self.noise_scheduler.get_velocity(latents, noise, timesteps)
|
|
else:
|
|
raise ValueError(f"Unknown prediction type {self.noise_scheduler.config.prediction_type}")
|
|
|
|
loss = F.mse_loss(model_pred.cast(paddle.float32), target.cast(paddle.float32), reduction="mean")
|
|
return loss
|
|
|
|
|
|
class StableDiffusionBenchmark(BenchmarkBase):
|
|
def __init__(self):
|
|
super().__init__()
|
|
|
|
@staticmethod
|
|
def add_args(args, parser):
|
|
parser.add_argument(
|
|
"--model_name_or_path",
|
|
type=str,
|
|
default="CompVis/stable-diffusion-v1-4",
|
|
help="Model name. Defaults to CompVis/stable-diffusion-v1-4. ",
|
|
)
|
|
parser.add_argument(
|
|
"--resolution",
|
|
type=int,
|
|
default=512,
|
|
help=(
|
|
"The resolution for input images, all the images in the train/validation dataset will be resized to this"
|
|
" resolution"
|
|
),
|
|
)
|
|
parser.add_argument(
|
|
"--center_crop",
|
|
action="store_true",
|
|
help="Whether to center crop images before resizing to resolution (if not set, random crop will be used)",
|
|
)
|
|
parser.add_argument(
|
|
"--random_flip",
|
|
action="store_true",
|
|
help="whether to randomly flip images horizontally",
|
|
)
|
|
parser.add_argument(
|
|
"--dataloader_num_workers",
|
|
type=int,
|
|
default=4,
|
|
help=(
|
|
"Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
|
|
),
|
|
)
|
|
|
|
def create_input_specs(self):
|
|
input_ids = paddle.static.InputSpec(name="input_ids", shape=[-1, self.model_max_length], dtype="int64")
|
|
dtype = "float16" if self.args.use_amp and self.args.amp_level == "O2" else "float32"
|
|
pixel_values = paddle.static.InputSpec(
|
|
name="pixel_values", shape=[-1, 3, self.args.resolution, self.args.resolution], dtype=dtype
|
|
)
|
|
return [input_ids, pixel_values]
|
|
|
|
def create_data_loader(self, args, **kwargs):
|
|
caption_column = "text"
|
|
image_column = "image"
|
|
self.tokenizer = tokenizer = CLIPTokenizer.from_pretrained(
|
|
url_or_path_join(args.model_name_or_path, "tokenizer")
|
|
)
|
|
|
|
def tokenize_captions(examples, is_train=True):
|
|
captions = []
|
|
for caption in examples[caption_column]:
|
|
if isinstance(caption, str):
|
|
captions.append(caption)
|
|
elif isinstance(caption, (list, np.ndarray)):
|
|
# take a random caption if there are multiple
|
|
captions.append(random.choice(caption) if is_train else caption[0])
|
|
else:
|
|
raise ValueError(
|
|
f"Caption column `{caption_column}` should contain either strings or lists of strings."
|
|
)
|
|
inputs = tokenizer(
|
|
captions,
|
|
max_length=tokenizer.model_max_length,
|
|
padding="do_not_pad",
|
|
truncation=True,
|
|
return_attention_mask=False,
|
|
)
|
|
return inputs.input_ids
|
|
|
|
# Preprocessing the datasets.
|
|
train_transforms = transforms.Compose(
|
|
[
|
|
transforms.Resize((args.resolution, args.resolution), interpolation="bilinear"),
|
|
transforms.CenterCrop(args.resolution) if args.center_crop else transforms.RandomCrop(args.resolution),
|
|
transforms.RandomHorizontalFlip() if args.random_flip else Lambda(lambda x: x),
|
|
transforms.ToTensor(),
|
|
transforms.Normalize([0.5], [0.5]),
|
|
]
|
|
)
|
|
|
|
def preprocess_train(examples):
|
|
images = [image.convert("RGB") for image in examples[image_column]]
|
|
examples["pixel_values"] = [train_transforms(image) for image in images]
|
|
examples["input_ids"] = tokenize_captions(examples)
|
|
|
|
return examples
|
|
|
|
from ppdiffusers.training_utils import main_process_first
|
|
from ppdiffusers.utils import PPDIFFUSERS_CACHE
|
|
|
|
file_path = get_path_from_url_with_filelock(
|
|
"https://paddlenlp.bj.bcebos.com/models/community/junnyu/develop/pokemon-blip-captions.tar.gz",
|
|
PPDIFFUSERS_CACHE,
|
|
)
|
|
dataset = DatasetDict.load_from_disk(file_path)
|
|
|
|
with main_process_first():
|
|
repeat_dataset = concatenate_datasets([dataset["train"]] * 250)
|
|
dataset["train"] = repeat_dataset
|
|
# Set the training transforms
|
|
train_dataset = dataset["train"].with_transform(preprocess_train)
|
|
|
|
train_sampler = (
|
|
DistributedBatchSampler(train_dataset, batch_size=args.batch_size, shuffle=False, drop_last=True)
|
|
if paddle.distributed.get_world_size() > 1
|
|
else BatchSampler(train_dataset, batch_size=args.batch_size, shuffle=False, drop_last=True)
|
|
)
|
|
|
|
def collate_fn(examples):
|
|
pixel_values = paddle.stack([example["pixel_values"] for example in examples])
|
|
if args.use_amp and args.amp_level == "O2":
|
|
pixel_values = pixel_values.cast(paddle.float16)
|
|
input_ids = [example["input_ids"] for example in examples]
|
|
input_ids = tokenizer.pad(
|
|
{"input_ids": input_ids},
|
|
padding="max_length",
|
|
max_length=tokenizer.model_max_length,
|
|
return_tensors="pd",
|
|
return_attention_mask=False,
|
|
).input_ids
|
|
return {"pixel_values": pixel_values, "input_ids": input_ids}
|
|
|
|
train_dataloader = DataLoader(
|
|
train_dataset, batch_sampler=train_sampler, collate_fn=collate_fn, num_workers=args.dataloader_num_workers
|
|
)
|
|
self.num_batch = len(train_dataloader)
|
|
self.model_max_length = tokenizer.model_max_length
|
|
return train_dataloader, None
|
|
|
|
def build_model(self, args, **kwargs):
|
|
model = StableDiffusion(args)
|
|
self.args = args
|
|
return model
|
|
|
|
def forward(self, model, args, input_data=None, **kwargs):
|
|
loss = model(**input_data)
|
|
return (
|
|
loss,
|
|
input_data["input_ids"].shape[0],
|
|
)
|
|
|
|
def logger(
|
|
self,
|
|
args,
|
|
step_id=None,
|
|
pass_id=None,
|
|
batch_id=None,
|
|
loss=None,
|
|
batch_cost=None,
|
|
reader_cost=None,
|
|
num_samples=None,
|
|
ips=None,
|
|
**kwargs
|
|
):
|
|
logger.info(
|
|
"global step %d / %d, loss: %f, avg_reader_cost: %.5f sec, avg_batch_cost: %.5f sec, avg_samples: %.5f, ips: %.5f sample/sec"
|
|
% (step_id, args.epoch * self.num_batch, loss, reader_cost, batch_cost, num_samples, ips)
|
|
)
|