Fine-tune Models#

Although CLIP-as-service has provided you a list of pre-trained models, you can also fine-tune your models. This guide will show you how to use Finetuner to fine-tune models and use them in CLIP-as-service.

For installation and basic usage of Finetuner, please refer to Finetuner documentation. You can also learn more details about fine-tuning CLIP.

This tutorial requires finetuner >=v0.6.4, clip_server >=v0.6.0.

Prepare Training Data#

Finetuner accepts training data and evaluation data in the form of DocumentArray. The training data for CLIP is a list of (text, image) pairs. Each pair is stored in a Document which wraps two chunks with image and text modality respectively. You can push the resulting DocumentArray to the cloud using the push() method.

We use fashion captioning dataset as a sample dataset in this tutorial. The following are examples of descriptions and image urls from the dataset. We also include a preview of each image.

Description

Image URL

Preview

subtly futuristic and edgy this liquid metal cuff bracelet is shaped from sculptural rectangular link

https://n.nordstrommedia.com/id/sr3/
58d1a13f-b6b6-4e68-b2ff-3a3af47c422e.jpeg

high quality leather construction defines a hearty boot one-piece on a tough lug sole

https://n.nordstrommedia.com/id/sr3/
21e7a67c-0a54-4d09-a4a4-6a0e0840540b.jpeg

this shimmering tricot knit tote is traced with decorative whipstitching and diamond cut chain the two hallmark of the falabella line

https://n.nordstrommedia.com/id/sr3/
1d8dd635-6342-444d-a1d3-4f91a9cf222b.jpeg

You can use the following script to transform the first three entries of the dataset to a DocumentArray and push it to the cloud using the name fashion-sample.

from docarray import Document, DocumentArray

train_da = DocumentArray(
    [
        Document(
            chunks=[
                Document(
                    content='subtly futuristic and edgy this liquid metal cuff bracelet is shaped from sculptural rectangular link',
                    modality='text',
                ),
                Document(
                    uri='https://n.nordstrommedia.com/id/sr3/58d1a13f-b6b6-4e68-b2ff-3a3af47c422e.jpeg',
                    modality='image',
                ),
            ],
        ),
        Document(
            chunks=[
                Document(
                    content='high quality leather construction defines a hearty boot one-piece on a tough lug sole',
                    modality='text',
                ),
                Document(
                    uri='https://n.nordstrommedia.com/id/sr3/21e7a67c-0a54-4d09-a4a4-6a0e0840540b.jpeg',
                    modality='image',
                ),
            ],
        ),
        Document(
            chunks=[
                Document(
                    content='this shimmering tricot knit tote is traced with decorative whipstitching and diamond cut chain the two hallmark of the falabella line',
                    modality='text',
                ),
                Document(
                    uri='https://n.nordstrommedia.com/id/sr3/1d8dd635-6342-444d-a1d3-4f91a9cf222b.jpeg',
                    modality='image',
                ),
            ],
        ),
    ]
)
train_da.push('fashion-sample')

The full dataset has been converted to clip-fashion-train-data and clip-fashion-eval-data and pushed to the cloud which can be directly used in Finetuner.

Start Finetuner#

You may now create and run a fine-tuning job after login to Jina ecosystem.

import finetuner

finetuner.login()
run = finetuner.fit(
    model='ViT-B-32::openai',
    run_name='clip-fashion',
    train_data='clip-fashion-train-data',
    eval_data='clip-fashion-eval-data',  # optional
    epochs=5,
    learning_rate=1e-5,
    loss='CLIPLoss',
    to_onnx=True,
)

After the job started, you may use status() to check the status of the job.

import finetuner

finetuner.login()
run = finetuner.get_run('clip-fashion')
print(run.status())

When the status is FINISHED, you can download the tuned model to your local machine.

import finetuner

finetuner.login()
run = finetuner.get_run('clip-fashion')
run.save_artifact('clip-model')

You should now get a zip file containing the tuned model named clip-fashion.zip under the folder clip-model.

Use the Model#

After unzipping the model you get from the previous step, a folder with the following structure will be generated:

.
└── clip-fashion/
    ├── config.yml
    ├── metadata.yml
    ├── metrics.yml
    └── models/
        ├── clip-text/
        │   ├── metadata.yml
        │   └── model.onnx
        ├── clip-vision/
        │   ├── metadata.yml
        │   └── model.onnx
        └── input-map.yml

Since the tuned model generated from Finetuner contains richer information such as metadata and config, we now transform it to simpler structure used by CLIP-as-service.

  • Firstly, create a new folder named clip-fashion-cas or name of your choice. This will be the storage of the models to use in CLIP-as-service.

  • Secondly, copy the textual model clip-fashion/models/clip-text/model.onnx into the folder clip-fashion-cas and rename the model to textual.onnx.

  • Similarly, copy the visual model clip-fashion/models/clip-vision/model.onnx into the folder clip-fashion-cas and rename the model to visual.onnx.

This is the expected structure of clip-fashion-cas:

.
└── clip-fashion-cas/
    ├── textual.onnx
    └── visual.onnx

In order to use the fine-tuned model, create a custom YAML file finetuned_clip.yml like below. Learn more about Flow YAML configuration and clip_server YAML configuration.

jtype: Flow
version: '1'
with:
  port: 51000
executors:
  - name: clip_o
    uses:
      jtype: CLIPEncoder
      metas:
        py_modules:
          - clip_server.executors.clip_onnx
      with:
        name: ViT-B-32::openai
        model_path: 'clip-fashion-cas' # path to clip-fashion-cas
    replicas: 1

You can use finetuner.describe_models() to check the supported models in finetuner, you should see:

                                                                Finetuner backbones                                                                      
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓
┃                                             name            task  output_dim  architecture                                                 description ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┩
│                                  bert-base-cased    text-to-text         768   transformer  BERT model pre-trained on BookCorpus and English Wikipedia │
│                     openai/clip-vit-base-patch16   text-to-image         512   transformer                          CLIP base model with patch size 16 │
│                     openai/clip-vit-base-patch32   text-to-image         512   transformer                                             CLIP base model │
│                openai/clip-vit-large-patch14-336   text-to-image         768   transformer                         CLIP large model for 336x336 images │
│                    openai/clip-vit-large-patch14   text-to-image        1024   transformer                         CLIP large model with patch size 14 │
│                                  efficientnet_b0  image-to-image        1280           cnn                     EfficientNet B0 pre-trained on ImageNet │
│                                  efficientnet_b4  image-to-image        1792           cnn                     EfficientNet B4 pre-trained on ImageNet │
│                                    RN101::openai   text-to-image         512   transformer                             Open CLIP "RN101::openai" model │
│                          RN101-quickgelu::openai   text-to-image         512   transformer                   Open CLIP "RN101-quickgelu::openai" model │
│                         RN101-quickgelu::yfcc15m   text-to-image         512   transformer                  Open CLIP "RN101-quickgelu::yfcc15m" model │
│                                   RN101::yfcc15m   text-to-image         512   transformer                            Open CLIP "RN101::yfcc15m" model │
│                                      RN50::cc12m   text-to-image        1024   transformer                               Open CLIP "RN50::cc12m" model │
│                                     RN50::openai   text-to-image        1024   transformer                              Open CLIP "RN50::openai" model │
│                            RN50-quickgelu::cc12m   text-to-image        1024   transformer                     Open CLIP "RN50-quickgelu::cc12m" model │
│                           RN50-quickgelu::openai   text-to-image        1024   transformer                    Open CLIP "RN50-quickgelu::openai" model │
│                          RN50-quickgelu::yfcc15m   text-to-image        1024   transformer                   Open CLIP "RN50-quickgelu::yfcc15m" model │
│                                  RN50x16::openai   text-to-image         768   transformer                           Open CLIP "RN50x16::openai" model │
│                                   RN50x4::openai   text-to-image         640   transformer                            Open CLIP "RN50x4::openai" model │
│                                  RN50x64::openai   text-to-image        1024   transformer                           Open CLIP "RN50x64::openai" model │
│                                    RN50::yfcc15m   text-to-image        1024   transformer                             Open CLIP "RN50::yfcc15m" model │
│                          ViT-B-16::laion400m_e31   text-to-image         512   transformer                   Open CLIP "ViT-B-16::laion400m_e31" model │
│                          ViT-B-16::laion400m_e32   text-to-image         512   transformer                   Open CLIP "ViT-B-16::laion400m_e32" model │
│                                 ViT-B-16::openai   text-to-image         512   transformer                          Open CLIP "ViT-B-16::openai" model │
│                 ViT-B-16-plus-240::laion400m_e31   text-to-image         640   transformer          Open CLIP "ViT-B-16-plus-240::laion400m_e31" model │
│                 ViT-B-16-plus-240::laion400m_e32   text-to-image         640   transformer          Open CLIP "ViT-B-16-plus-240::laion400m_e32" model │
│                            ViT-B-32::laion2b_e16   text-to-image         512   transformer                     Open CLIP "ViT-B-32::laion2b_e16" model │
│                          ViT-B-32::laion400m_e31   text-to-image         512   transformer                   Open CLIP "ViT-B-32::laion400m_e31" model │
│                          ViT-B-32::laion400m_e32   text-to-image         512   transformer                   Open CLIP "ViT-B-32::laion400m_e32" model │
│                                 ViT-B-32::openai   text-to-image         512   transformer                          Open CLIP "ViT-B-32::openai" model │
│                ViT-B-32-quickgelu::laion400m_e31   text-to-image         512   transformer         Open CLIP "ViT-B-32-quickgelu::laion400m_e31" model │
│                ViT-B-32-quickgelu::laion400m_e32   text-to-image         512   transformer         Open CLIP "ViT-B-32-quickgelu::laion400m_e32" model │
│                       ViT-B-32-quickgelu::openai   text-to-image         512   transformer                Open CLIP "ViT-B-32-quickgelu::openai" model │
│                             ViT-L-14-336::openai   text-to-image         768   transformer                      Open CLIP "ViT-L-14-336::openai" model │
│                                 ViT-L-14::openai   text-to-image         768   transformer                          Open CLIP "ViT-L-14::openai" model │
│                                        resnet152  image-to-image        2048           cnn                           ResNet152 pre-trained on ImageNet │
│                                         resnet50  image-to-image        2048           cnn                            ResNet50 pre-trained on ImageNet │
│ sentence-transformers/msmarco-distilbert-base-v3    text-to-text         768   transformer                     Pretrained BERT, fine-tuned on MS Marco │
└──────────────────────────────────────────────────┴────────────────┴────────────┴──────────────┴───────────────────────────────────────────────────────────

You can now start the clip_server using fine-tuned model to get a performance boost:

python -m clip_server finetuned_clip.yml

That’s it, enjoy 🚀