finetuner 0.3.1.dev15

Finetuner allows one to tune the weights of any deep neural network for better embedding on search tasks.

Finetuning any deep neural network for better embedding on neural search tasks

Finetuner allows one to tune the weights of any deep neural network for better embeddings on search tasks. It accompanies Jina to deliver the last mile of performance for domain-specific neural search applications.

🎛 Designed for finetuning: a human-in-the-loop deep learning tool for leveling up your pretrained models in domain-specific neural search applications.

🔱 Powerful yet intuitive: all you need is finetuner.fit() - a one-liner that unlocks rich features such as siamese/triplet network, interactive labeling, layer pruning, weights freezing, dimensionality reduction.

⚛️ Framework-agnostic: promise an identical API & user experience on PyTorch, Tensorflow/Keras and PaddlePaddle deep learning backends.

🧈 Jina integration: buttery smooth integration with Jina, reducing the cost of context-switch between experiment and production.

How does it work

Install

Requires Python 3.7+ and one of PyTorch(>=1.9) or Tensorflow(>=2.5) or PaddlePaddle installed on Linux/MacOS.

pip install finetuner

Documentation

Usage

Usage	Do you have an embedding model?
Yes	No
Do you have labeled data?	Yes	🟠	🟡
	No	🟢	🔵

🟠 Have embedding model and labeled data

Perfect! Now embed_model and labeled_data are given by you already, simply do:

import finetuner

tuned_model = finetuner.fit(
    embed_model,
    train_data=labeled_data
)

🟢 Have embedding model and unlabeled data

You have an embed_model to use, but no labeled data for finetuning this model. No worry, that's good enough already! You can use Finetuner to interactive label data and train embed_model as below:

import finetuner

tuned_model = finetuner.fit(
    embed_model,
    train_data=unlabeled_data,
    interactive=True
)

🟡 Have general model and labeled data

You have a general_model which does not output embeddings. Luckily you provide some labeled_data for training. No worries, Finetuner can convert your model into an embedding model and train it via:

import finetuner

tuned_model = finetuner.fit(
    general_model,
    train_data=labeled_data,
    to_embedding_model=True,
    layer_name='my_embedding_layer',
    freeze=['layer_1', 'layer_2'],
)

🔵 Have general model and unlabeled data

You have a general_model which is not for embeddings. Meanwhile, you don't have labeled data for training. But no worries, Finetuner can help you train an embedding model with interactive labeling on-the-fly:

import finetuner

tuned_model = finetuner.fit(
    general_model,
    train_data=unlabeled_data,
    interactive=True,
    to_embedding_model=True,
    layer_name='my_embedding_layer',
    freeze=['layer_1', 'layer_2'],
)

Finetuning ResNet50 on CelebA

⚡ To get the best experience, you will need a GPU-machine for this example. For CPU users, we provide finetuning a MLP on FashionMNIST and finetuning a Bi-LSTM on CovidQA that run out the box on low-profile machines. Check out more examples in our docs!

1. Download CelebA-small dataset (7.7MB) and decompress it to './img_align_celeba'. Full dataset can be found here.
2. Finetuner accepts Jina DocumentArray/DocumentArrayMemmap, so we load CelebA image into this format using a generator:

   from jina.types.document.generators import from_files
   
   # please change the file path to your data path
   data = list(from_files('img_align_celeba/*.jpg', size=100, to_dataturi=True))
   
   for doc in data:
       doc.load_uri_to_image_blob(
           height=224, width=224
       ).set_image_blob_normalization().set_image_blob_channel_axis(
           -1, 0
       )  # No need for changing channel axes line if you are using tf/keras
   

3. Load pretrained ResNet50 using PyTorch/Keras/Paddle:
   • PyTorch

     import torchvision
     model = torchvision.models.resnet50(pretrained=True)
     

   • Keras

     import tensorflow as tf
     model = tf.keras.applications.resnet50.ResNet50(weights='imagenet')
     

   • Paddle

     import paddle
     model = paddle.vision.models.resnet50(pretrained=True)
     

4. Start the Finetuner:

   import finetuner
   
   finetuner.fit(
       model=model,
       interactive=True,
       train_data=data,
       freeze=True,
       to_embedding_model=True,
       input_size=(3, 224, 224),
       layer_name='my_embedding_layer',
       freeze=['layer_1', 'layer_2'],
   )
   

5. After downloading the model and loading the data (takes ~20s depending on your network/CPU/GPU), your browser will open the Labeler UI as below. You can now label the relevance of celebrity faces via mouse/keyboard. The ResNet50 model will get finetuned and improved as you are labeling. If you are running this example on a CPU machine, it may take up to 20 seconds for each labeling round.

Support

• Use Discussions to talk about your use cases, questions, and support queries.
• Join our Slack community and chat with other Jina community members about ideas.
• Join our Engineering All Hands meet-up to discuss your use case and learn Jina's new features.
  • When? The second Tuesday of every month
  • Where? Zoom (see our public events calendar/.ical) and live stream on YouTube
• Subscribe to the latest video tutorials on our YouTube channel

Join Us

Finetuner is backed by Jina AI and licensed under Apache-2.0. We are actively hiring AI engineers, solution engineers to build the next neural search ecosystem in opensource.