argusvision 1.1.0

Downloads pretrained Argus Vision models

Argus Vision

argusvision package provides easy to use access to image embedding models pretrained on Bing data - Argus Vision models. Interface is based on popular torchvision.

In version 1.0, supported model is Argus Vision V6 - resnext101 32x8d

Please note that provided models are offering image embeddings

Installation

pip install argusvision

For Windows, torch needs to be install from wheel file. Please download the latest torch from here: https://download.pytorch.org/whl/torch_stable.html and then install it using: pip install <.whl file>

Usage

Input images should be in BGR format of shape (3 x H x W), where H and W are expected to be at least 224. The images have to be loaded in to a range of [0, 1] and then normalized using mean = [0.485, 0.456, 0.406] and std = [0.229, 0.224, 0.225].

Example script:

import argusvision
import torch

# This will load pretrained model
model = argusvision.models.resnext101_32x8d()

# This will initialize weights with default values
model = argusvision.models.resnext101_32x8d(pretrained=False) 

# Load model to CPU memory, interface is the same as torchvision
model = argusvision.resnext101_32x8d(map_location=torch.device('cpu')) 

Example of creating image embeddings:

import argusvision
from torchvision import transforms
import torch
from PIL import Image

def get_image():
    img = cv2.imread('example.jpg', cv2.IMREAD_COLOR)
    img = cv2.resize(img, (256, 256))
    img = img[16:256-16, 16:256-16]
    preprocess = transforms.Compose([
        transforms.ToTensor(),
        transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    ])
    return preprocess(image).unsqueeze(0) # Unsqueeze only required when there's 1 image in images batch

model = argusvision.models.resnext101_32x8d(map_location=torch.device('cpu'))
features = model(get_image())
print(features.shape)

Should output

...
torch.Size([1, 2048])

Benchmarks

Here are the evaluations of the popular datasets

Model	CIFAR-10	STL-10
Torchvision, ResNext101 32x8d	90%	81.1%
Argusvision, ResNext101 32x8d	92.6%	84.2%