Custom Models
While DeepStack provides many functionalities out of the box, it allows you to also deploy image recognition models trained on your own dataset.
For example, you can train a model on a dataset of different classes of plants. With DeepStack, you can deploy this model to actually classify plants in a production environment.
DeepStack supports custom image classification models in ONNX, Tensorflow and Keras. With the ONNX support, you can train a model in any deep learning framework including Pytorch, MxNet, Chainer, CNTK and more. and deploy them to production with DeepStack.
In this guide, we shall walk through deploying a custom model using the three supported formats. The model we are deploying here is trained to recognize different classes of professions by their mode of dressing.
Starting DeepStack on Docker
Below we start DeepStack with with no prebuilt API enabled, since we are using only custom models here
sudo docker run -v localstorage:/datastore -p 80:5000 deepquestai/deepstack
Basic Parameters
-v localstorage:/datastore This specifies the local volume where deepstack will store all data.
-p 80:5000 This makes deepstack accessible via port 80 of the machine.
Start the DeepStack App, Click Start Server, deselect any selected API and click Start Now
ONNX is a universal model format supported by the most popular deep learning frameworks. A model trained in a framework like Pytorch can be easily exported to onnx.
Download the trained IdenProf onnx model
The configuration file contains all the information about the preprocessing and labels for your model.
{"sys-version": "1.0","framework":"ONNX","mean":0.5,"std":255,"width":224,"height":224,"map": {"0": "chef", "1": "doctor", "2": "engineer", "3": "farmer","4": "firefighter", "5": "judge", "6": "mechanic","7": "pilot", "8": "police", "9": "waiter"}}
Config Parameters
sys-version This is a constant used internally by DeepStack.
framework This specifies the model format, supported values are ONNX, TF AND KERAS
std The input image is divided by this value, (Standard Deviation)
mean This is subtracted from the image after dividing by the std
map This is a mapping of the class indexes to the actual labels.
import requestsfrom io import openmodel = open("idenprof.onnx","rb").read()config = open("config.json","rb").read()response = requests.post("http://localhost:80/v1/vision/addmodel",files={"model":model,"config":config},data={"name":"profession"}).json()print(response)
The code above, uploads the model and the config file to your local DeepStack server, also, the {“name”:”profession”} specifies the unique name for the model. This model will be served on the endpoint http://localhost:80/v1/vision/custom/profession
Below, we shall attempt to use our custom model to predict the class of the image below
import requestsimage_data = open("test-custom-image.jpg","rb").read()response = requests.post("http://localhost:80/v1/vision/custom/profession",files={"image":image_data}).json()print("Label:",response["label"])print(response)
{'label': 'farmer', 'success': True, 'confidence': 0.584346}
Keras is a popular deep learning framework focussed on ease of use.
Deploying keras models follows the same process as onnx models.
You can download the keras model here Idenprof Keras Model
Note that when using your custom keras models, the model file must contain both the weights and the architecture
In your keras code, you can save the weights and architecture by using
model.save("model.h5")
The config file is essentially the same, except that the framework should be changed to KERAS
{"sys-version": "1.0","framework":"KERAS","mean":0.5,"std":255,"width":224,"height":224,"map": {"0": "chef", "1": "doctor", "2": "engineer", "3": "farmer","4": "firefighter", "5": "judge", "6": "mechanic","7": "pilot", "8": "police", "9": "waiter"}}
Config Parameters
sys-version This is a constant used internally by DeepStack.
framework This specifies the model format, supported values are ONNX, TF AND KERAS
std The input image is divided by this value, (Standard Deviation)
mean This is subtracted from the image after dividing by the std
map This is a mapping of the class indexes to the actual labels.
Now, we can register the model in the same way.
import requestsfrom io import openmodel = open("idenprof.h5","rb").read()config = open("config.json","rb").read()response = requests.post("http://localhost:80/v1/vision/addmodel",files={"model":model,"config":config},data={"name":"profession"}).json()print(response)
You can test your Keras model using the Example Test Code
Tensorflow is a very popular DL framework from Google.
Deploying tensorflow models follows the same process as onnx and keras.
You can download the tensorflow model here Idenprof Tensorflow Model
The config file for tensorflow models must contain the input_name and the output_name
{"sys-version": "1.0","framework":"TF","mean":0.5,"std":255,"width":224,"height":224,"input_name":"input_1:0","output_name":"output_1:0","map": {"0": "chef", "1": "doctor", "2": "engineer", "3": "farmer","4": "firefighter", "5": "judge", "6": "mechanic","7": "pilot", "8": "police", "9": "waiter"}}
Config Parameters
sys-version This is a constant used internally by DeepStack.
framework This specifies the model format, supported values are ONNX, TF AND KERAS
input_name The name of the input node in the Tensorflow model
output_name The name of the output node in the Tensorflow model
std The input image is divided by this value, (Standard Deviation)
mean This is subtracted from the image after dividing by the std
map This is a mapping of the class indexes to the actual labels.
Now, we can register the model in the same way.
import requestsfrom io import openmodel = open("idenprof.pb","rb").read()config = open("config.json","rb").read()response = requests.post("http://localhost:80/v1/vision/addmodel",files={"model":model,"config":config},data={"name":"profession"}).json()print(response)
You can test your Tensorflow model using the Example Test Code
Unlike the prebuilt APIs, custom models are auto started the moment you run DeepStack. When you add, delete or update a model, you need to restart the server for these changes to take effect.