Few Shot Classification GUI-based Training

Demo: Youtube

Try at: Streamlit

• This is a Streamlit based tool to train and develop a Few Shot Classification ML model very rapidly.
• It uses Prototypical Networks with Resnet18 as the backbone network.
• During training, model parameters are also saved on disk.

Requirements (installable via pip)

• torch
• torchvision
• easyfsl
• matplotlib
• streamlit
• pillow

What I used?

1. PyTorch - An open source machine learning framework that accelerates the path from research prototyping to production deployment.
   • Resnet18 CNN
   • Loading datasets
   • Image transformations
2. Omniglot Data set - for Demo - It is designed for developing more human-like learning algorithms. It contains 1623 different handwritten characters from 50 different alphabets. Each of the 1623 characters was drawn online via Amazon's Mechanical Turk by 20 different people.
3. Streamlit - for GUI - Streamlit is an open-source app framework for Machine Learning and Data Science teams.
4. Matplotlib - for plotting loss function - Matplotlib is a comprehensive library for creating static, animated, and interactive visualizations in Python. Matplotlib makes easy things easy and hard things possible.

How to run the program?

1. Download this GitHub repository

   • Either Clone the repository

      git clone https://github.com/Kunal-Attri/Few-Shot-Classification-GUI-based-Training.git
     

   • Or download and extract the zip archive of the repository.

2. Download & Install requirements

   • Ensure that you have Python 3 installed.
   • Open terminal in the Repository folder on your local machine.
   • Run the following command to install requirements.

      pip3 install -r requirements.txt
     

3. Run the Program

   streamlit run train.py
   

   Expected Interface:
   

4. Preparing the Data set

   • To try: Download omniglot_dataset: It has the full Omniglot Data set which has characters of 30 different languages.
   • Create a folder, then sub-folders(which will essentially represent different classes), and then put images for those classes in those sub-folders.
     • It's a good idea to keep no. of images for N classes to be almost equal if not exactly equal. They must be quality images as per ML standards.
     • Then, zip the main folder!

5. Training model on the data set

   • Set parameters. Unless, you have your own data set, you can try demo values.
     • Train-test split ratio: That is the ratio in which you want to divide your data set for training and testing phases. Generally 60 to 90 percent is a good ratio. For demo: 75 %
     • Unique Classes in Data set: No of different classes in the data. For demo: 2 classes
     • Episodes in Train set: It's like the Epoch in ML. More complex the data set, higher episodes will benefit. Generally, 200-600 is a good number for training. For demo: 400
     • Episodes in Test set: Generally, 20-200 is a good number for testing. For demo: 50
     • Images in each class of Support Set: These are the no. of images that will be supplied in each support set during training phase. Generally, 3-10 is a good choice. For demo: 7
     • Images in each class of Query Set: No. of images that will be supplied in each set during testing phase. Generally 5-20 is good value. For demo: 10
     • Data set Path: It refers to the location at which data set is located. It can be either relative or absolute path. For demo: train
     • Image size for Data augmentation: That refers to the transformed size of images for training and testing. Low data -> Low res, else higher res. For demo: 100
   • Demo Parameters:
     
     
   • Then check the Train Model checkbox. It will now train the model. Should take 3-7 minutes. Model is saved every 25 % of model training progress, and we also get the Training Loss plot vs Episodes.
   • Trained Model:
     
     
     
   • NOTE: While setting parameters Don't set too high values or too low values, as it can cause Overfitting or Underfitting respectively.

6. Testing Model on Test Set

   • Uncheck the Train Model checkbox.
   • And, check the Evaluate Model on Test Set checkbox. This will now test the model on Test set from the data set.
   • After evaluation, it will give the Accuracy on test set as the output. Here, our model has 97.6% accuracy!!!
   • Evaluation on Test set:
     
     
     

7. Testing Model on Custom Image Input

   • Uncheck all checkboxes.
   • Check the Evaluate Model on Single Image Input checkbox. This will enable a image upload bar.
   • You can now drag-n-drop a image or upload one, to get predicted output from the model.
   • Evaluation on Image:
     
     
     

References

• Few Shot Classification
• Prototypical Networks
• Resnet18
• PyTorch
• Omniglot Data set
• Streamlit
• Matplotlib