Inspiration
We chose to pursue this challenge since we had freedom of choosing whatever feels right for the specific objective of the challenge and also to put some skills to the test!
What it does
This finetuned neural network uses resnet18 (a pretained neural model for feature extraction) to reconstruct an altered (rotated) version of a fashion-related image and get feature embeddings along the way. This is done with the purpose of providing an accurate and fast model that's able to showcase some of the most similar options (images) for a particular query fashion image.
How we built it
We've mainly used the deep learning framework Pytorch to develop the model's architecture but we've also experimented with Keras using pretrained VGG16 to find where should our results lie in. We've used a pretrained neural network resnet18 and then added some finetuning to get some image embeddings and then with cosine similarity we get the similarity between the images.
Challenges we ran into
We've run through lots of errors and debugging them was a complete struggle! Although we faces some difficulties we've still managed to deliver a nice project.
Accomplishments that we're proud of
We're proud of having finished the project with our initial expectations.
What we learned
We've learned to work collaboratively in finetuning models in a speedy and accurate way.
What's next for Fashion_IndiNET
Developing a stronger UI and improving the model with more images.
Built With
- git
- gradio
- keras
- pillow
- python
- scikit-learn
- torch
Log in or sign up for Devpost to join the conversation.