An adaptation of Google's Teachable Machine library to classify a food item and look up nutrition information.
This code is an adaptation of the code provided with Google's Teachable Machine library. I'm making this for a school project. While Image models are not provided in this code repository, you can make your own model easily & for free with Google's Teachable Machine, or, if you don't want to try and gather images of food, you can download my model from this link (Note that this model file is NOT pre-trained, it's just the training data. To convert it to the h5 file the code needs, see the guide below.)
The version on this branch (websv) has two files, an updated classifier script that only outputs JSON, and an HTML server you can make a POST request to and get a response. Simply keep websv.py running and both py files in the same directory, then make a POST request. Here's an example output, with a picture of steak passed:
% curl -X POST http://localhost:8080 -d '{"arg1": "IMG_2942.jpeg"}'
{"name": "Steak", "nutrition_info": {"Calories": "273.4", "Service size": "100.0g", "Saturated Fat": "7.3g", "Total Fat": "18.8g", "Protein": "26.0g", "Sodium": "52mg", "Potassium": "194mg", "Cholesterol": "95mg", "Total Carbohydrates": "0.0g", "Fiber": "0.0g", "Sugar": "0.0g"}}%
One thing I would like to see incorporated with this is a script to where a user can take a picture (of food) with their phone, and it will SFTP it (or upload the file to the remote folder one way or another), and then the name of the uploaded file is passed to the POST request, and the results are fed back. This has been accomplished, with the Java version. See branch Java-and-py.
To get an image model, go to Google's Teachable Machine.
If you want to build your own training database, go right ahead! Simply create a new image classifier, and train your model how you want it, then skip to downloading your model.
If you would rather use my example model, first download the Food training.tm library from either this link or the one above (they're the same). Then, select "Open an existing project from file".
When your model is tuned to your liking, you can hit "Export Model".
You will be presented with a screen like this, just go to the "Tensorflow" tab, make sure "Keras" is selected, and hit "Download my model". Then, take the files from the downloaded folder and paste them into the folder with your python scripts. That's it!
To prepare this program for use:
- Clone this branch of the repository with
git clone -b websv --single-branch https://github.com/THEWHITEBOY503/TM-ImageClassifier.git - Install the dependancies with
pip install tensorflow keras pillow requests(You may need to install tensorflow another way) - Paste your ML model files into the cloned directory
- Paste your image files into the directory
- Set up your API key
- This project utilizes API-Ninja's nutrition API. You can read more about the API and sign up to get a key here. Once you have your key, simply find and replace
--YOUR-API-KEY-HERE--with your API key.
- This project utilizes API-Ninja's nutrition API. You can read more about the API and sign up to get a key here. Once you have your key, simply find and replace
- Start the web server with
python3 websv.py
That's it!
Right now, there's no way to upload an image to the web server using a POST request. So, for the time being, you will need to have the image you want to process in the same directory as the scripts.
Let's say I have this picture stored on my server in the same directory as my scripts under the file name IMG_6216.jpeg:
(mmm, yummy...)
In order to process this, I have to send the server this data with my POST request:
{"arg1": "IMG_2819.jpeg"}
So, if I was going to make a cURL request to process this photo, I'd run curl -X POST http://localhost:8080 -d '{"arg1": "IMG_2819.jpeg"}'
(Note-- in this example the server (right side) is being run on a Ubuntu VM running inside my MacBook Pro on Parallels, while the left side is just the terminal from macOS)