Skip to content

🔥 Use pre-trained models in PyTorch to extract vector embeddings for any image

License

Notifications You must be signed in to change notification settings

christiansafka/img2vec

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

62 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Image 2 Vec with PyTorch

Medium post on building the first version from scratch: https://becominghuman.ai/extract-a-feature-vector-for-any-image-with-pytorch-9717561d1d4c

Applications of image embeddings:

  • Ranking for recommender systems
  • Clustering images to different categories
  • Classification tasks
  • Image compression

Available models

Model name Return vector length
Resnet-18 512
Alexnet 4096
Vgg-11 4096
Densenet 1024
efficientnet_b0 1280
efficientnet_b1 1280
efficientnet_b2 1408
efficientnet_b3 1536
efficientnet_b4 1792
efficientnet_b5 2048
efficientnet_b6 2304
efficientnet_b7 2560

Installation

Tested on Python 3.6 and torchvision 0.11.0 (nightly, 2021-09-25)

Requires Pytorch: http://pytorch.org/

conda install -c pytorch-nightly torchvision

pip install img2vec_pytorch

Run test

python -m img2vec_pytorch.test_img_to_vec

Using img2vec as a library

from img2vec_pytorch import Img2Vec
from PIL import Image

# Initialize Img2Vec with GPU
img2vec = Img2Vec(cuda=True)

# Read in an image (rgb format)
img = Image.open('test.jpg')
# Get a vector from img2vec, returned as a torch FloatTensor
vec = img2vec.get_vec(img, tensor=True)
# Or submit a list
vectors = img2vec.get_vec(list_of_PIL_images)
For running the example, you will additionally need:
  • Pillow: pip install Pillow
  • Sklearn pip install scikit-learn

Running the example

git clone https://github.com/christiansafka/img2vec.git

cd img2vec/example

python test_img_similarity.py

Expected output

Which filename would you like similarities for?
cat.jpg
0.72832 cat2.jpg
0.641478 catdog.jpg
0.575845 face.jpg
0.516689 face2.jpg

Which filename would you like similarities for?
face2.jpg
0.668525 face.jpg
0.516689 cat.jpg
0.50084 cat2.jpg
0.484863 catdog.jpg

Try adding your own photos!

Img2Vec Params

cuda = (True, False)   # Run on GPU?     default: False
model = ('resnet-18', 'efficientnet_b0', etc.)   # Which model to use?     default: 'resnet-18'

Advanced users


Read only file systems

If you use this library from the app running in read only environment (for example, docker container), specify writable directory where app can store pre-trained models.

export TORCH_HOME=/tmp/torch

Additional Parameters

layer = 'layer_name' or int   # For advanced users, which layer of the model to extract the output from.   default: 'avgpool'
layer_output_size = int   # Size of the output of your selected layer
gpu = (0, 1, etc.)   # Which GPU to use?     default: 0

Defaults: (layer = 'avgpool', layer_output_size = 512)
Layer parameter must be an string representing the name of a layer below

conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False)
bn1 = nn.BatchNorm2d(64)
relu = nn.ReLU(inplace=True)
maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
layer1 = self._make_layer(block, 64, layers[0])
layer2 = self._make_layer(block, 128, layers[1], stride=2)
layer3 = self._make_layer(block, 256, layers[2], stride=2)
layer4 = self._make_layer(block, 512, layers[3], stride=2)
avgpool = nn.AvgPool2d(7)
fc = nn.Linear(512 * block.expansion, num_classes)

Defaults: (layer = 2, layer_output_size = 4096)
Layer parameter must be an integer representing one of the layers below

alexnet.classifier = nn.Sequential(
            7. nn.Dropout(),                  < - output_size = 9216
            6. nn.Linear(256 * 6 * 6, 4096),  < - output_size = 4096
            5. nn.ReLU(inplace=True),         < - output_size = 4096
            4. nn.Dropout(),		      < - output_size = 4096
            3. nn.Linear(4096, 4096),	      < - output_size = 4096
            2. nn.ReLU(inplace=True),         < - output_size = 4096
            1. nn.Linear(4096, num_classes),  < - output_size = 4096
        )

Defaults: (layer = 2, layer_output_size = 4096)

vgg.classifier = nn.Sequential(
            nn.Linear(512 * 7 * 7, 4096),
            nn.ReLU(True),
            nn.Dropout(),
            nn.Linear(4096, 4096),
            nn.ReLU(True),
            nn.Dropout(),
            nn.Linear(4096, num_classes),
        )

Defaults: (layer = 1 from features, layer_output_size = 1024)

densenet.features = nn.Sequential(OrderedDict([
	('conv0', nn.Conv2d(3, num_init_features, kernel_size=7, stride=2,
						padding=3, bias=False)),
	('norm0', nn.BatchNorm2d(num_init_features)),
	('relu0', nn.ReLU(inplace=True)),
	('pool0', nn.MaxPool2d(kernel_size=3, stride=2, padding=1)),
]))

Defaults: (layer = 1 from features, layer_output_size = 1280 for efficientnet_b0 model)

To-do

  • Benchmark speed and accuracy
  • Add ability to fine-tune on input data
  • Export documentation to a normal place

About

🔥 Use pre-trained models in PyTorch to extract vector embeddings for any image

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages