As an introductory project to myself, I built an end-to-end CNN Image Classification Model which identifies the food in your image.
I worked out with a pre-trained Image Classification Model that comes with Keras and then retrained it on the infamous Food101 Dataset.
The Model actually beats the DeepFood Paper's model which also trained on the same dataset.
The Accuracy aquired by DeepFood was 77.4% and our model's 85% . Difference of 8% ain't much, but the interesting thing is, DeepFood's model took 2-3 days to train while our's barely took 90min.
Finally after training the model, I have exported it as .hdf5 files and then integrated it with Streamlit Web App.
Streamlit turns data scripts into shareable web apps in minutes. Once I got the App working on my local device I then deployed it using Streamlit’s invite-only sharing feature
Check the deployed app, or the demo video
The app may take a couple of seconds to load for the first time, but it works perfectly fine.
app.video.mp4
Once an app is loaded,
- Upload an image of food. If you dont have one, use the images from
food-images/ - Once the image is processed,
Predictbutton appears. Click it. - Once you click the
Predictbutton, the model prediction takes place and the output will be displayed along with the model's Top-5 Predictions - And voilà, there you go.
If you actually want to know the Nuts and Bolts how the model was trained check out
model-training.ipynbNotebook
-
Imported Food101 dataset from Tensorflow Datasets Module.
-
Setup Global dtype policy to
mixed_float16to implement Mixed Precision TrainingMixed precision is the use of both 16-bit and 32-bit floating-point types in a model during training to make it run faster and use less memory.
-
As we are dealing with a complex Neural Network (EfficientNetB0) its a good practice to have few callbacks set up. Few ones I will be using throughtout this Notebook are :
-
TensorBoard Callback : TensorBoard provides the visualization and tooling needed for machine learning experimentation
-
EarlyStoppingCallback : Used to stop training when a monitored metric has stopped improving.
-
ReduceLROnPlateau : Reduce learning rate when a metric has stopped improving.
-
-
Built a Fine Tuning Model
This part tool the longest. In Deep Learning, you have to know which nob does what. Once yoy get experienced you'll what nobs you should turn to get the results you want. Architecture :
EffficientNetB1
Again if you want to dive deeper on how the model was trained check out
model-training.ipynbNotebook
-
Once we have our model ready, its cruicial to evaluate it on our custom data : the data our model has never seen.
Training and evaluating a model on train and test data is cool, but making predictions on our own realtime images is another level.
Once we are satisfied with the results, we can export the model as a
.hdf5which can be used in future for model deployment.
Once the model is exported then there comes the Deployment part. Check out app.py to get more insight on How I integrated it with Streamlit.
At first glance the files in the repo may look intimidating and overwhelming. To avoid that, here is a quick guide :
.gitignore: tells what files/folders to ignore when committingapp.py: Our Food Vision app built using Streamlitutils.py: Some of used fuctions inapp.pymodel-training.ipynb: Google Colab Notebook used to train the modelmodel/: Contains all the models used as .hfd5 filesrequirements.txt: List of required dependencies required to runapp.pyextras/: Has some miscellaneous images and files used to write this README Document
Post your queries on the Discussions tab, else contact me : [email protected]