Image Embedding with ISC

author: Jael Gu

Description

An image embedding operator generates a vector given an image. This operator extracts features for image top ranked models from Image Similarity Challenge 2021 - Descriptor Track. The default pretrained model weights are from The 1st Place Solution of ISC21 (Descriptor Track).

Code Example

Load an image from path './towhee.jpg' and use the pretrained ISC model ('resnet50') to generate an image embedding.

Write a pipeline with explicit inputs/outputs name specifications:

from towhee.dc2 import pipe, ops, DataCollection

p = (
    pipe.input('path')
        .map('path', 'img', ops.image_decode())
        .map('img', 'vec', ops.image_embedding.isc())
        .output('img', 'vec')
)

DataCollection(p('towhee.jpg')).show()

Factory Constructor

Create the operator via the following factory method

image_embedding.isc(skip_preprocess=False, device=None)

Parameters:

skip_preprocess: bool

The flag to control whether to skip image preprocess. The default value is False. If set to True, it will skip image preprocessing steps (transforms). In this case, input image data must be prepared in advance in order to properly fit the model.

device: str

The device to run this operator, defaults to None. When it is None, 'cuda' will be used if it is available, otherwise 'cpu' is used.

Interface

An image embedding operator takes a towhee image as input. It uses the pre-trained model specified by model name to generate an image embedding in ndarray.

Parameters:

img: towhee.types.Image (a sub-class of numpy.ndarray)

The decoded image data in numpy.ndarray.

Returns: numpy.ndarray

The image embedding extracted by model.

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_embedding.isc().get_op()
op.save_model('onnx', 'test.onnx')

Fine-tune

Requirements

If you want to fine tune this operator, make sure your timm version is 0.4.12, and the higher version would cause model collapse during training.

! python -m pip install tqdm augly timm==0.4.12 pytorch-metric-learning==0.9.99

Download dataset

Your need to download dataset from Facebook AI Image Similarity Challenge: Descriptor Track, and use the ./training_images folder as your training images root. It requires about 165G space.

Get started to fine-tune

Just call method op.train() and pass in your args.

import towhee

op = towhee.ops.image_embedding.isc().get_op()
op.train(training_args={
        'train_data_dir': './training_images',
        'distributed': False,
        'output_dir': './output',
        'gpu': 0,
        'epochs': 2,
        'batch_size': 128,
        'init_lr': 0.1
    })

These are each arg infos in training_args:

• output_dir

  • default: './output'
  • metadata_dict: {'help': 'output checkpoint saving dir.'}

• distributed

  • default: False
  • metadata_dict: {'help': 'If true, use all gpu in your machine, else use only one gpu.'}

• gpu

  • default: 0
  • metadata_dict: {'help': 'When distributed is False, use this gpu No. in your machine.'}

• start_epoch

  • default: 0
  • metadata_dict: {'help': 'Start epoch number.'}

• epochs

  • default: 6
  • metadata_dict: {'help': 'End epoch number.'}

• batch_size

  • default: 128
  • metadata_dict: {'help': 'Total batch size in all gpu.'}

• init_lr

  • default: 0.1
  • metadata_dict: {'help': 'init learning rate in SGD.'}

• train_data_dir

  • default: None
  • metadata_dict: {'help': 'The dir containing all training data image files.'}

Your custom training

Your can change training script in your way. Or your can refer to the original repo and paper to learn more about contrastive learning and image instance retrieval.