Deep Learning as Cloud APIs.
This project provides an image classification cloud service for research on Black-box Adversarial Attacks.
docker run -p 8080:8080 wuhanstudio/deepapi
$ pip install deepapi
$ python -m deepapi
Serving on port 8080...
By default, we enable all models on the server. Use deepapi -h to see more options.
The website and API service are available at https://localhost:8080.
To initiate black-box adversarial attacks, we can get predictions from a cloud API using model.predict().
Behind the scene, this model makes predictions by sending a POST request to http://localhost:8080/vgg16_cifar10.
import numpy as np
from PIL import Image
from deepapi.api import DeepAPI_VGG16_Cifar10
# Load the image
x = Image.open("dog.jpg")
x = np.array(x)
# Initialize the model
model = DeepAPI_VGG16_Cifar10('http://localhost:8080', concurrency=8)
# Predict
y = model.predict(np.array([x]))[0]
# Print the result
model.print(y)
export IMAGE_FILE=test/cat.jpg
(echo -n '{"file": "'; base64 $IMAGE_FILE; echo '"}') | \
curl -H "Content-Type: application/json" \
-d @- http://127.0.0.1:8080/vgg16_cifar10
You can also implement the API client from scratch using the request module.
def classification(url, file):
# Load the input image and construct the payload for the request
image = Image.open(file)
buff = BytesIO()
image.save(buff, format="JPEG")
data = {'file': base64.b64encode(buff.getvalue()).decode("utf-8")}
return requests.post(url, json=data).json()
res = classification('http://127.0.0.1:8080/vgg16_cifar10', 'cat.jpg')
This python script is available in the test folder. You should see prediction results by running python3 minimal.py:
cat 0.99804
deer 0.00156
truck 0.00012
airplane 0.00010
dog 0.00009
bird 0.00005
ship 0.00003
frog 0.00001
horse 0.00001
automobile 0.00001
Sending 5 concurrent requests to the API server:
$ python3 multi-client.py --num_workers 5 cat.jpg
You should see the result:
----- start -----
Sending requests
Sending requests
Sending requests
Sending requests
Sending requests
------ end ------
Concurrent Requests: 5
Total Runtime: 2.441638708114624