sayanam / de_cifar10_assignment

Resnet on CIFAR10 Image dataset

DE_CIFAR10_Assignment

Resnet on CIFAR10 Image dataset

1. Built Resnet-18 Architecture.

2. Trained the model with CIFAR10 dataset

3. Various experiments were conducted by configuring values of parameters such as Image Augmentation, Normalization, Optimizers and Dropout. The results are available in Model Performance Log

4. The best model with image augmentation, normlization, SGD with dynamic learning rate and dropout(0.25).

5. The model was trained for 200 eppoch with the best accuracy of 90.98 on test data.

6. The following per class accuracy on test data were obtained.

    			aeroplane: 0.86
    			automobile: 0.96
    			bird: 0.865
    			cat: 0.762
    			deer: 0.916
    			dog: 0.85
    			frog: 0.947
    			horse: 0.935
    			ship: 0.951
    			truck: 0.935
   

   with confusion matrix of :

Inferences:
	Ships and automobiles have the highest accuracy amoung all the classes
	Cats are being misclassified as dogs and vice versa.

7. Application Information

Directory

Training

Use Training Notebook and Model Performance

Setting up the Environment

To setup the environment

python3 -m pip install -U virtualenv # install the virtualenv
virtualenv -p python3 dev_env # creating an environment
source dev_env/bin/activate # activate environment
git clone https://github.com/sayanam/DE_CIFAR10_Assignment #clone the repo
cd DE_CIFAR10_Assignment
pip install -r requirements.txt # install the requirements

For Inference

To generate validation data run the below code:

cd data
python generate_valid_data.py

Run application using Flask

Run the below command to commence the application

python flask_api/app.py

8. How to use the application?

9. Deployment in HEROKU

Configuring HEROKU

Note: requirement.txt and Procfile are required to be present in the project workspace for Hiroku to identify the initial file that has to run. a) Sign-up for Heroku at https://devcenter.heroku.com/ and agree to all the necessary terms and conditions. b) In the home page click on New -> Create a new app. c) Enter the app name and click on Create app d) Selected the selected app e) Select Deploy tab and connect to the github and configure the repository from which the code has to be deployed f) Select settings and click on Add buildpack and select python e) Goto Deploy tab again and click on to f) Click Open app for the app to open.

Note: How to check application logs ?? Select More --> View Logs