在自定义数据集上实现OpenAI CLIP

在2021年1月，OpenAI宣布了两个新模型:DALL-E和CLIP，它们都是以某种方式连接文本和图像的多模态模型。CLIP全称是Contrastive Language–Image Pre-training，一种基于对比文本-图像对的预训练方法。为什么要介绍CLIP呢？因为现在大火得Stable Diffusion 并不是单一模型，而是多个模型组成。其中会用到一个 Text encoder 将用户的文本输入进行编码，这个 text encoder 就是 CLIP 模型中 text encoder

CLIP模型在训练时，可以给它一个输入句子，并提取最相关的图像来配合它。CLIP学习了一个完整的句子和它所描述的图像之间的关系。也就是说它是在完整的句子上训练的，而不是像“汽车”、“狗”等离散的分类，这一点对于应用至关重要。当训练完整的短语时，模型可以学习更多的东西，并识别照片和文本之间的模式。他们还证明，当在相当大的照片和与之相对应的句子数据集上进行训练时，该模型是可以作为分类器的。CLIP在发布的时候能在无任何微调的情况下（zero-shot ），在 ImageNet 数据集上的分类表现超 ResNets-50 微调后的效果，也就是说他是非常有用的。

所以在本文中，我们将使用PyTorch中从头开始实现CLIP模型，以便我们对CLIP有一个更好的理解

这里就需要用到2个库：timm和transformers，我们先导入代码

下一步就是预处理数据和通用配置config。config是一个普通的python文件，我们将所有的超参数放在里面，如果使用Jupyter Notebook的情况下，它是一个在Notebook开头定义的类。

class CFG:

  debug = False

  image_path = "../input/flickr-image-dataset/flickr30k_images/flickr30k_images"

  captions_path = "."

  batch_size = 32

  num_workers = 4

  head_lr = 1e-3

  image_encoder_lr = 1e-4

  text_encoder_lr = 1e-5

  weight_decay = 1e-3

  patience = 1

  factor = 0.8

  epochs = 2

  device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

  model_name = 'resnet50'

  image_embedding = 2048

  text_encoder_model = "distilbert-base-uncased"

  text_embedding = 768

  text_tokenizer = "distilbert-base-uncased"

  max_length = 200

  pretrained = True # for both image encoder and text encoder

  trainable = True # for both image encoder and text encoder

  temperature = 1.0

  # image size

  size = 224

  # for projection head; used for both image and text encoders

  num_projection_layers = 1

  projection_dim = 256

  dropout = 0.1

还有一些我们自定义指标的辅助类

class AvgMeter:

  def __init__(self, name="Metric"):

      self.name = name

      self.reset()

  def reset(self):

      self.avg, self.sum, self.count = [0] * 3

  def update(self, val, count=1):

      self.count += count

      self.sum += val * count

      self.avg = self.sum / self.count

  def __repr__(self):

      text = f": "

      return text

def get_lr(optimizer):

  for param_group in optimizer.param_groups:

      return param_group["lr"]

我们的目标是描述图像和句子。所以数据集必须同时返回句子和图像。所以需要使用DistilBERT标记器对句子(标题)进行标记，然后将标记id (input_ids)和注意掩码提供给DistilBERT。DistilBERT比BERT 模型要小，但是模型的结果都差不多，所以我们选择使用它。

下一步就是使用HuggingFace tokenizer进行标记化。在__init__中获得的tokenizer对象，将在模型运行时加载。标题被填充并截断到预定的最大长度。在加载相关图像之前，我们将在__getitem__中加载一个编码的标题，这是一个带有键input_ids和attention_mask的字典，并对其进行转换和扩充(如果有的话)。然后把它变成一个张量，并以“image”作为键存储在字典中。最后我们将标题的原始文本与关键字“标题”一起输入字典。

class CLIPDataset(torch.utils.data.Dataset):

  def __init__(self, image_filenames, captions, tokenizer, transforms):

      """

      image_filenames and cpations must have the same length; so, if there are

      multiple captions for each image, the image_filenames must have repetitive

      file names

      """

      self.image_filenames = image_filenames

      self.captions = list(captions)

      self.encoded_captions = tokenizer(

          list(captions), padding=True, truncation=True, max_length=CFG.max_length

      )

      self.transforms = transforms

  def __getitem__(self, idx):

      item = {

          key: torch.tensor(values[idx])

          for key, values in self.encoded_captions.items()

      }

      image = cv2.imread(f"/")

      image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)

      image = self.transforms(image=image)['image']

      item['image'] = torch.tensor(image).permute(2, 0, 1).float()

      item['caption'] = self.captions[idx]

      return item

  def __len__(self):

      return len(self.captions)

def get_transforms(mode="train"):

  if mode == "train":

      return A.Compose(

          [

              A.Resize(CFG.size, CFG.size, always_apply=True),

              A.Normalize(max_pixel_value=255.0, always_apply=True),

          ]

      )

  else:

      return A.Compose(

          [

              A.Resize(CFG.size, CFG.size, always_apply=True),

              A.Normalize(max_pixel_value=255.0, always_apply=True),

          ]

      )

图像和文本编码器:我们将使用ResNet50作为图像编码器。

class ImageEncoder(nn.Module):

  """

  Encode images to a fixed size vector

  """

  def __init__(

      self, model_name=CFG.model_name, pretrained=CFG.pretrained, trainable=CFG.trainable

  ):

      super().__init__()

      self.model = timm.create_model(

          model_name, pretrained, num_classes=0, global_pool="avg"

      )

      for p in self.model.parameters():

          p.requires_grad = trainable

  def forward(self, x):

      return self.model(x)

使用DistilBERT作为文本编码器。使用CLS令牌的最终表示来获得句子的整个表示。

class TextEncoder(nn.Module):

  def __init__(self, model_name=CFG.text_encoder_model, pretrained=CFG.pretrained, trainable=CFG.trainable):

      super().__init__()

      if pretrained:

          self.model = DistilBertModel.from_pretrained(model_name)

      else:

          self.model = DistilBertModel(config=DistilBertConfig())

      for p in self.model.parameters():

          p.requires_grad = trainable

      # we are using the CLS token hidden representation as the sentence's embedding

      self.target_token_idx = 0

  def forward(self, input_ids, attention_mask):

      output = self.model(input_ids=input_ids, attention_mask=attention_mask)

      last_hidden_state = output.last_hidden_state

      return last_hidden_state[:, self.target_token_idx, :]

上面的代码已经将图像和文本编码为固定大小的向量(图像2048，文本768)，我们需要图像和文本具有相似的尺寸，以便能够比较它们，所以我们把2048维和768维向量投影到256维(projection_dim)，只有维度相同我们才能比较它们。

class ProjectionHead(nn.Module):

  def __init__(

      self,

      embedding_dim,

      projection_dim=CFG.projection_dim,

      dropout=CFG.dropout

  ):

      super().__init__()

      self.projection = nn.Linear(embedding_dim, projection_dim)

      self.gelu = nn.GELU()

      self.fc = nn.Linear(projection_dim, projection_dim)

      self.dropout = nn.Dropout(dropout)

      self.layer_norm = nn.LayerNorm(projection_dim)

  def forward(self, x):

      projected = self.projection(x)

      x = self.gelu(projected)

      x = self.fc(x)

      x = self.dropout(x)

      x = x + projected

      x = self.layer_norm(x)

      return x

所以最后我们的CLIP模型就是这样：

class CLIPModel(nn.Module):

  def __init__(

      self,

      temperature=CFG.temperature,

      image_embedding=CFG.image_embedding,

      text_embedding=CFG.text_embedding,

  ):

      super().__init__()

      self.image_encoder = ImageEncoder()

      self.text_encoder = TextEncoder()

      self.image_projection = ProjectionHead(embedding_dim=image_embedding)

      self.text_projection = ProjectionHead(embedding_dim=text_embedding)

      self.temperature = temperature

  def forward(self, batch):

      # Getting Image and Text Features

      image_features = self.image_encoder(batch["image"])

      text_features = self.text_encoder(

          input_ids=batch["input_ids"], attention_mask=batch["attention_mask"]

      )

      # Getting Image and Text Embeddings (with same dimension)

      image_embeddings = self.image_projection(image_features)

      text_embeddings = self.text_projection(text_features)

      # Calculating the Loss

      logits = (text_embeddings @ image_embeddings.T) / self.temperature

      images_similarity = image_embeddings @ image_embeddings.T

      texts_similarity = text_embeddings @ text_embeddings.T

      targets = F.softmax(

          (images_similarity + texts_similarity) / 2 * self.temperature, dim=-1

      )

      texts_loss = cross_entropy(logits, targets, reduction='none')

      images_loss = cross_entropy(logits.T, targets.T, reduction='none')

      loss = (images_loss + texts_loss) / 2.0 # shape: (batch_size)

      return loss.mean()

#这里还加了一个交叉熵函数

def cross_entropy(preds, targets, reduction='none'):

  log_softmax = nn.LogSoftmax(dim=-1)

  loss = (-targets * log_softmax(preds)).sum(1)

  if reduction == "none":

      return loss

  elif reduction == "mean":

      return loss.mean()

这里需要说明下，CLIP使用 symmetric cross entropy 作为损失函数，可以降低噪音影响，提高模型鲁棒性，我们这里为了简单只是用cross entropy 。

我们可以进行测试：

# A simple Example

batch_size = 4

dim = 256

embeddings = torch.randn(batch_size, dim)

out = embeddings @ embeddings.T

print(F.softmax(out, dim=-1))

下一步就是训练了，有一些函数可以帮助我们加载训练和验证的dataloader

def make_train_valid_dfs():

  dataframe = pd.read_csv(f"/captions.csv")

  max_id = dataframe["id"].max() + 1 if not CFG.debug else 100

  image_ids = np.arange(0, max_id)

  np.random.seed(42)

  valid_ids = np.random.choice(

      image_ids, size=int(0.2 * len(image_ids)), replace=False

  )

  train_ids = [id_ for id_ in image_ids if id_ not in valid_ids]

  train_dataframe = dataframe[dataframe["id"].isin(train_ids)].reset_index(drop=True)

  valid_dataframe = dataframe[dataframe["id"].isin(valid_ids)].reset_index(drop=True)

  return train_dataframe, valid_dataframe

def build_loaders(dataframe, tokenizer, mode):

  transforms = get_transforms(mode=mode)

  dataset = CLIPDataset(

      dataframe["image"].values,

      dataframe["caption"].values,

      tokenizer=tokenizer,

      transforms=transforms,

  )

  dataloader = torch.utils.data.DataLoader(

      dataset,

      batch_size=CFG.batch_size,

      num_workers=CFG.num_workers,

      shuffle=True if mode == "train" else False,

  )

  return dataloader

然后就是训练和评估

def train_epoch(model, train_loader, optimizer, lr_scheduler, step):

  loss_meter = AvgMeter()

  tqdm_object = tqdm(train_loader, total=len(train_loader))

  for batch in tqdm_object:

      batch =

      loss = model(batch)

      optimizer.zero_grad()

      loss.backward()

      optimizer.step()

      if step == "batch":

          lr_scheduler.step()

      count = batch["image"].size(0)

      loss_meter.update(loss.item(), count)

      tqdm_object.set_postfix(train_loss=loss_meter.avg, lr=get_lr(optimizer))

  return loss_meter

def valid_epoch(model, valid_loader):

  loss_meter = AvgMeter()

  tqdm_object = tqdm(valid_loader, total=len(valid_loader))

  for batch in tqdm_object:

      batch =

      loss = model(batch)

      count = batch["image"].size(0)

      loss_meter.update(loss.item(), count)

      tqdm_object.set_postfix(valid_loss=loss_meter.avg)

  return loss_meter

最后整合起来就是全部流程

def main():

  train_df, valid_df = make_train_valid_dfs()

  tokenizer = DistilBertTokenizer.from_pretrained(CFG.text_tokenizer)

  train_loader = build_loaders(train_df, tokenizer, mode="train")

  valid_loader = build_loaders(valid_df, tokenizer, mode="valid")

  model = CLIPModel().to(CFG.device)

  params = [

      {"params": model.image_encoder.parameters(), "lr": CFG.image_encoder_lr},

      {"params": model.text_encoder.parameters(), "lr": CFG.text_encoder_lr},

      {"params": itertools.chain(

          model.image_projection.parameters(), model.text_projection.parameters()

      ), "lr": CFG.head_lr, "weight_decay": CFG.weight_decay}

  ]

  optimizer = torch.optim.AdamW(params, weight_decay=0.)

  lr_scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(

      optimizer, mode="min", patience=CFG.patience, factor=CFG.factor

  )

  step = "epoch"

  best_loss = float('inf')

  for epoch in range(CFG.epochs):

      print(f"Epoch: ")

      model.train()

      train_loss = train_epoch(model, train_loader, optimizer, lr_scheduler, step)

      model.eval()

      with torch.no_grad():

          valid_loss = valid_epoch(model, valid_loader)

      if valid_loss.avg < best_loss:

          best_loss = valid_loss.avg

          torch.save(model.state_dict(), "best.pt")

          print("Saved Best Model!")

      lr_scheduler.step(valid_loss.avg)

应用：获取图像嵌入并找到匹配。

我们训练完成后如何实际应用呢？我们需要编写一个函数加载训练后的模型，为其提供验证集中的图像，并返回形状(valid_set_size, 256)和模型本身的image_embeddings。

def get_image_embeddings(valid_df, model_path):

  tokenizer = DistilBertTokenizer.from_pretrained(CFG.text_tokenizer)

  valid_loader = build_loaders(valid_df, tokenizer, mode="valid")

  model = CLIPModel().to(CFG.device)

  model.load_state_dict(torch.load(model_path, map_location=CFG.device))

  model.eval()

  valid_image_embeddings = []

  with torch.no_grad():

      for batch in tqdm(valid_loader):

          image_features = model.image_encoder(batch["image"].to(CFG.device))

          image_embeddings = model.image_projection(image_features)

          valid_image_embeddings.append(image_embeddings)

  return model, torch.cat(valid_image_embeddings)

_, valid_df = make_train_valid_dfs()

model, image_embeddings = get_image_embeddings(valid_df, "best.pt")

def find_matches(model, image_embeddings, query, image_filenames, n=9):

  tokenizer = DistilBertTokenizer.from_pretrained(CFG.text_tokenizer)

  encoded_query = tokenizer([query])

  batch = {

      key: torch.tensor(values).to(CFG.device)

      for key, values in encoded_query.items()

  }

  with torch.no_grad():

      text_features = model.text_encoder(

          input_ids=batch["input_ids"], attention_mask=batch["attention_mask"]

      )

      text_embeddings = model.text_projection(text_features)

  image_embeddings_n = F.normalize(image_embeddings, p=2, dim=-1)

  text_embeddings_n = F.normalize(text_embeddings, p=2, dim=-1)

  dot_similarity = text_embeddings_n @ image_embeddings_n.T

  values, indices = torch.topk(dot_similarity.squeeze(0), n * 5)

  matches = [image_filenames[idx] for idx in indices[::5]]

  _, axes = plt.subplots(3, 3, figsize=(10, 10))

  for match, ax in zip(matches, axes.flatten()):

      image = cv2.imread(f"/")

      image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)

      ax.imshow(image)

      ax.axis("off")

  plt.show()

调用方法如下：

find_matches(model,

            image_embeddings,

            query="one dog sitting on the grass",

            image_filenames=valid_df['image'].values,

            n=9)

可以看到我们自定义效果还是不错的（但是图里面有个猫，哈）。也就是说CLIP这种方法在小数据集上自定义也是可行的。

以下是本文的代码和数据集：

https://www.kaggle.com/code/jyotidabas/simple-openai-clip-implementation

作者:Jyoti Dabass, Ph.D