Identifying rose species from a given image

• 

  result 1 from model serving interface - accurately identified rosa canina or dog rose

• 

  result 2 from model serving interface - accurately identified rosa glauca

• 

  result 2 from model serving interface - accurately identified - rosa centifolia or cabbage rose

What inspired you

Wanted to build something which could be presented to masses. The use case was something which I wanted to be less technical in nature and people can easily connect with.

What it does

Identifies three species of roses, could be very useful for horticulturist as with the help of an application, which can correctly identify roses, population of various species can be tracked and compared with other locations.

How I built it

I used Pytorch and FastAI framework and used Transfer Learning.Took a pretrained model and retrained with new set of images. I have also scraped internet for building my image repository. I trained with 399 images.

Challenges I ran into

Productionizing model was something which took a lot of time.Sometimes docker did not work. I also faced some challenges in recording the video. I also noticed that changing a learning rate has degraded the model performance. The final model is trained with only ten epochs, I have earlier trained to a much higher number with not so great results.

Accomplishments that I'm proud of

I have completed my end to end learning of deep learning where data acquisition, model training and model serving were accomplished.

What I learned

How much to train a model and model serving are two important and tricky things to learn. I am glad that I gained some perspective here. Also working with Pytorch is a breeze, it truly lives up to the tagline "From Research to Production".

What's next for Identifying rose species from a given image - Identify more species, currently my model identifies three species of roses.

Built With

• docker
• fastai
• python
• pytorch
• uvicorn