• 演绎提示，生成我们的“高度详细，锐利的焦点，插图，宏伟，史诗”图像的3d渲染

1.比赛目标

这个竞赛的目标不是从文本提示生成图像，而是创建一个模型，可以在给定生成图像的情况下预测文本提示（你有一堆提示词，你预测是否该提示词参与了图像的生成）?您将在包含由Stable Diffusion 2.0生成的各种(提示，图像)对的数据集上进行预测，以了解潜在关系的可逆程度。

2.内容

文本到图像模型的流行已经摒弃了提示工程的一个全新领域。一部分是艺术，一部分是悬而未决的科学，ML从业者和研究人员正在迅速努力理解提示和它们生成的图像之间的关系。在提示符上添加“4k”是使其更逼真的最佳方法吗?提示中的小扰动会导致高度发散的图像吗?提示关键字的顺序如何影响生成的场景?这个竞赛的任务是创建一个模型，可以可靠地反转生成给定图像的扩散过程。

为了以一种稳健的方式计算提示的相似性——这意味着“史诗猫”的得分与“威严的小猫”相似（meaning that “epic cat” is scored as similar to “majestic kitten” in spite of character-level differences），尽管它们在字符级别上存在差异——你将提交你预测的提示的嵌入。是直接建模嵌入，还是先预测提示，然后转换为嵌入，都取决于您!祝你好运，并愿你在此创建“高质量、锐利焦点、复杂、详细、不真实的健壮交叉验证风格”的模型。

3评价指标

使用预测和实际提示嵌入向量之间的平均余弦相似度评分来评估提交。如何为groundtruth提示计算嵌入的精确细节见

数据

• images/ - 是一些从提示词中产生的图像;你的任务是预测是哪些提示词用来产生这个图像.隐藏的测试数据集包含大约16000张图片。
• prompts.csv - 用来产生图像的提示词。These are provided as illustrative examples only. It is up to each competitor to develop their own strategy of creating a training set of images, using pre-trained models, etc. Note that this file is not contained in the re-run test set, and thus referencing it in a Notebook submission will result in a failure.
• sample_submission.csv - 一个正确的示范 The values found in this file are embeddings of the prompts in the prompts.csv file and thus can be used validate your embedding pipeline. This notebook demonstrates how to calculate embeddings.

Baseline Stable Diffusion ViT Baseline Train

• 数据

Library

import os
import random
import numpy as np
import pandas as pd
from PIL import Image
from tqdm.notebook import tqdm
from scipy import spatial
from sklearn.model_selection import train_test_split
import torch
from torch import nn
from torch.utils.data import Dataset, DataLoader
from torch.optim.lr_scheduler import CosineAnnealingLR
from torchvision import transforms
import timm
from timm.utils import AverageMeter
import sys
sys.path.append('../input/sentence-transformers-222/sentence-transformers')
from sentence_transformers import SentenceTransformer
import warnings
warnings.filterwarnings('ignore')

Config

class CFG:model_name = 'vit_base_patch16_224'input_size = 224batch_size = 64num_epochs = 3lr = 1e-4seed = 42

seed

def seed_everything(seed):os.environ['PYTHONHASHSEED'] = str(seed)random.seed(seed)np.random.seed(seed)torch.manual_seed(seed)if torch.cuda.is_available(): torch.cuda.manual_seed(seed)torch.backends.cudnn.deterministic = Trueseed_everything(CFG.seed)

Dataset

class DiffusionDataset(Dataset):def __init__(self, df, transform):self.df = df#图像增强self.transform = transformdef __len__(self):return len(self.df)def __getitem__(self, idx):row = self.df.iloc[idx]#打开图片image = Image.open(row['filepath'])#图像增强image = self.transform(image)#标签prompt = row['prompt']return image, promptclass DiffusionCollator:def __init__(self):self.st_model = SentenceTransformer('/kaggle/input/sentence-transformers-222/all-MiniLM-L6-v2',device='cpu')def __call__(self, batch):images, prompts = zip(*batch)images = torch.stack(images)prompt_embeddings = self.st_model.encode(prompts, show_progress_bar=False, convert_to_tensor=True)return images, prompt_embeddingsdef get_dataloaders(trn_df,val_df,input_size,batch_size):#图像增强设置transform = transforms.Compose([transforms.Resize(input_size),transforms.ToTensor(),transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),])#datasettrn_dataset = DiffusionDataset(trn_df, transform)val_dataset = DiffusionDataset(val_df, transform)collator = DiffusionCollator()#dataloaderdataloaders = {}dataloaders['train'] = DataLoader(dataset=trn_dataset,shuffle=True,batch_size=batch_size,pin_memory=True,num_workers=2,drop_last=True,collate_fn=collator)dataloaders['val'] = DataLoader(dataset=val_dataset,shuffle=False,batch_size=batch_size,pin_memory=True,num_workers=2,drop_last=False,collate_fn=collator)return dataloaders

Train

#评价指标
def cosine_similarity(y_trues, y_preds):return np.mean([1 - spatial.distance.cosine(y_true, y_pred) for y_true, y_pred in zip(y_trues, y_preds)])

Train

def train(trn_df,val_df,model_name,input_size,batch_size,num_epochs,lr
):#设置运行环境device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')#dataloaderdataloaders = get_dataloaders(trn_df,val_df,input_size,batch_size)#创建模型model = timm.create_model(model_name,pretrained=True,num_classes=384#给定的词表长度为384)#梯度检查点，可以降低显存使用model.set_grad_checkpointing()#模型迁移到GPU上model.to(device)#优化器optimizer = torch.optim.AdamW(model.parameters(), lr=lr)#周期计算，用于CosineAnnealingLRttl_iters = num_epochs * len(dataloaders['train'])#学习率调整策略scheduler = CosineAnnealingLR(optimizer, T_max=ttl_iters, eta_min=1e-6)#eta_min.为最小的学习率#评价指标criterion = nn.CosineEmbeddingLoss()best_score = -1.0for epoch in range(num_epochs):train_meters = {'loss': AverageMeter(),'cos': AverageMeter(),}#模型设置为训练模式model.train()for X, y in tqdm(dataloaders['train'], leave=False):X, y = X.to(device), y.to(device)#梯度清零optimizer.zero_grad()X_out = model(X)target = torch.ones(X.size(0)).to(device)#计算损失loss = criterion(X_out, y, target)#反向传播loss.backward()#更新optimizer.step()scheduler.step()trn_loss = loss.item()trn_cos = cosine_similarity(X_out.detach().cpu().numpy(), y.detach().cpu().numpy())train_meters['loss'].update(trn_loss, n=X.size(0))train_meters['cos'].update(trn_cos, n=X.size(0))print('Epoch {:d} / trn/loss={:.4f}, trn/cos={:.4f}'.format(epoch + 1,train_meters['loss'].avg,train_meters['cos'].avg))val_meters = {'loss': AverageMeter(),'cos': AverageMeter(),}model.eval()for X, y in tqdm(dataloaders['val'], leave=False):X, y = X.to(device), y.to(device)with torch.no_grad():X_out = model(X)target = torch.ones(X.size(0)).to(device)loss = criterion(X_out, y, target)val_loss = loss.item()val_cos = cosine_similarity(X_out.detach().cpu().numpy(), y.detach().cpu().numpy())val_meters['loss'].update(val_loss, n=X.size(0))val_meters['cos'].update(val_cos, n=X.size(0))print('Epoch {:d} / val/loss={:.4f}, val/cos={:.4f}'.format(epoch + 1,val_meters['loss'].avg,val_meters['cos'].avg))if val_meters['cos'].avg > best_score:best_score = val_meters['cos'].avgtorch.save(model.state_dict(), f'{model_name}.pth')

准备训练数据

df = pd.read_csv('/kaggle/input/diffusiondb-data-cleansing/diffusiondb.csv')
trn_df, val_df = train_test_split(df, test_size=0.1, random_state=CFG.seed)

训练

train(trn_df, val_df, CFG.model_name, CFG.input_size, CFG.batch_size, CFG.num_epochs, CFG.lr)

模型推理

import numpy as np
import pandas as pd
from pathlib import Path
from PIL import Image
from tqdm.notebook import tqdm
import torch
from torch.utils.data import Dataset, DataLoader
from torchvision import transforms
import timm

class CFG:model_path = '/kaggle/input/stable-diffusion-vit-baseline-train/vit_base_patch16_224.pth'model_name = 'vit_base_patch16_224'input_size = 224batch_size = 64

dataset

class DiffusionTestDataset(Dataset):def __init__(self, images, transform):self.images = imagesself.transform = transformdef __len__(self):return len(self.images)def __getitem__(self, idx):image = Image.open(self.images[idx])image = self.transform(image)return image

inference

def predict(images,model_path,model_name,input_size,batch_size
):device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')transform = transforms.Compose([transforms.Resize(input_size),transforms.ToTensor(),transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225]),])dataset = DiffusionTestDataset(images, transform)dataloader = DataLoader(dataset=dataset,shuffle=False,batch_size=batch_size,pin_memory=True,num_workers=2,drop_last=False)model = timm.create_model(model_name,pretrained=False,num_classes=384)state_dict = torch.load(model_path)model.load_state_dict(state_dict)model.to(device)model.eval()preds = []for X in tqdm(dataloader, leave=False):X = X.to(device)with torch.no_grad():X_out = model(X)preds.append(X_out.cpu().numpy())return np.vstack(preds).flatten()

images = list(Path('/kaggle/input/stable-diffusion-image-to-prompts/images').glob('*.png'))
imgIds = [i.stem for i in images]
EMBEDDING_LENGTH = 384
imgId_eId = ['_'.join(map(str, i)) for i in zip(np.repeat(imgIds, EMBEDDING_LENGTH),np.tile(range(EMBEDDING_LENGTH), len(imgIds)))]prompt_embeddings = predict(images, CFG.model_path, CFG.model_name, CFG.input_size, CFG.batch_size)
submission = pd.DataFrame(index=imgId_eId,data=prompt_embeddings,columns=['val']
).rename_axis('imgId_eId')
submission.to_csv('submission.csv')