Image-Text Retrieval Embdding with CLIP

author: David Wang

Description

This operator extracts features for image or text with CLIP which can generate embeddings for text and image by jointly training an image encoder and text encoder to maximize the cosine similarity.

Code Example

Load an image from path './teddy.jpg' to generate an image embedding.

Read the text 'A teddybear on a skateboard in Times Square.' to generate an text embedding.

Write a pipeline with explicit inputs/outputs name specifications:

from towhee import pipe, ops, DataCollection

img_pipe = (
    pipe.input('url')
    .map('url', 'img', ops.image_decode.cv2_rgb())
    .map('img', 'vec', ops.image_text_embedding.clip(model_name='clip_vit_base_patch16', modality='image'))
    .output('img', 'vec')
)

text_pipe = (
    pipe.input('text')
    .map('text', 'vec', ops.image_text_embedding.clip(model_name='clip_vit_base_patch16', modality='text'))
    .output('text', 'vec')
)

DataCollection(img_pipe('./teddy.jpg')).show()
DataCollection(text_pipe('A teddybear on a skateboard in Times Square.')).show()

Factory Constructor

Create the operator via the following factory method

clip(model_name, modality)

Parameters:

​ model_name: str

​ The model name of CLIP. Supported model names:

• clip_vit_base_patch16
• clip_vit_base_patch32
• clip_vit_large_patch14
• clip_vit_large_patch14_336

​ modality: str

​ Which modality(image or text) is used to generate the embedding.

checkpoint_path: str

The path to local checkpoint, defaults to None. If None, the operator will download and load pretrained model by model_name from Huggingface transformers.

Interface

An image-text embedding operator takes a towhee image or string as input and generate an embedding in ndarray.

save_model(format='pytorch', path='default')

Save model to local with specified format.

Parameters:

format: str

​ The format of saved model, defaults to 'pytorch'.

path: str

​ The path where model is saved to. By default, it will save model to the operator directory.

from towhee import ops

op = ops.image_text_embedding.clip(model_name='clip_vit_base_patch16', modality='image').get_op()
op.save_model('onnx', 'test.onnx')

Parameters:

​ data: towhee.types.Image (a sub-class of numpy.ndarray) or str

​ The data (image or text based on specified modality) to generate embedding.

Returns: numpy.ndarray

​ The data embedding extracted by model.

supported_model_names(format=None)

Get a list of all supported model names or supported model names for specified model format.

Parameters:

format: str

​ The model format such as 'pytorch', 'torchscript'.

from towhee import ops


op = ops.image_text_embedding.clip(model_name='clip_vit_base_patch16', modality='image').get_op()
full_list = op.supported_model_names()
onnx_list = op.supported_model_names(format='onnx')
print(f'Onnx-support/Total Models: {len(onnx_list)}/{len(full_list)}')

Fine-tune

Requirement

If you want to train this operator, besides dependency in requirements.txt, you need install these dependencies. There is also an example to show how to finetune it on a custom dataset.

! python -m pip install datasets evaluate

Get start

import towhee

clip_op = towhee.ops.image_text_embedding.clip(model_name='clip_vit_base_patch16', modality='image').get_op()

data_args = {
    'dataset_name': 'ydshieh/coco_dataset_script',
    'dataset_config_name': '2017',
    'max_seq_length': 77,
    'data_dir': path_to_your_coco_dataset,
    'image_mean': [0.48145466, 0.4578275, 0.40821073],
    'image_std': [0.26862954, 0.26130258, 0.27577711]
}
training_args = {
    'num_train_epochs': 3, # you can add epoch number to get a better metric.
    'per_device_train_batch_size': 8,
    'per_device_eval_batch_size': 8,
    'do_train': True,
    'do_eval': True,
    'remove_unused_columns': False,
    'output_dir': './tmp/test-clip',
    'overwrite_output_dir': True,
}
model_args = {
    'freeze_vision_model': False,
    'freeze_text_model': False,
    'cache_dir': './cache'
}

clip_op.train(data_args=data_args, training_args=training_args, model_args=model_args)

Dive deep and customize your training

You can change the training script in your customer way. Or your can refer to the original hugging face transformers training examples.

More Resources

• CLIP Object Detection: Merging AI Vision with Language Understanding - Zilliz blog: CLIP Object Detection combines CLIP's text-image understanding with object detection tasks, allowing CLIP to locate and identify objects in images using texts.
• Supercharged Semantic Similarity Search in Production - Zilliz blog: Building a Blazing Fast, Highly Scalable Text-to-Image Search with CLIP embeddings and Milvus, the most advanced open-source vector database.
• The guide to clip-vit-base-patch32 | OpenAI: clip-vit-base-patch32: a CLIP multimodal model variant by OpenAI for image and text embedding.
• Image Embeddings for Enhanced Image Search - Zilliz blog: Image Embeddings are the core of modern computer vision algorithms. Understand their implementation and use cases and explore different image embedding models.
• From Text to Image: Fundamentals of CLIP - Zilliz blog: Search algorithms rely on semantic similarity to retrieve the most relevant results. With the CLIP model, the semantics of texts and images can be connected in a high-dimensional vector space. Read this simple introduction to see how CLIP can help you build a powerful text-to-image service.