Blog Post is here
Dataset is available over Kaggle
Published in Official PyTorch Medium Channel. Thank you PyTorch team !
- Builds a simple Convolutional Auto-encoder based Image similarity engine.
- This solves the problem of finding similar images using unsupervised learning. There are no labels for images.
Tasks that it can do.
- Similar images using Auto-encoders.
- Similar image search from image features.
- Clustering similar images.
=== data Read data from here. It is in gitignore so it won't appear here.
=== docs Documentation using mkdocs.
=== image_similarity
====== cluster_images.py Clusters the embeddings learnt using PCA and T-SNE.
====== torch_data.py Contains Dataset class to create PyTorch dataset from folder.
====== torch_model.py Convolutional Auto-enocder Model.
====== torch_engine.py Train_step and validation_step for training.
====== torch_train.py Trainng script. Trains Auto-enocder and saves the embeddings.
====== torch_infer.py Contains inference code
====== config.py Configurations of models and paths.
====== torch_inference.ipynb Inference code in .ipynb to play with.
====== torch_train.ipynb Stand-alone code to train in Jupyter.
=== tests Contains tests for CI
Please visit here for documentation.
It contains detailed description of work.
Some sample outputs of recommendations.
More outputs can be found in outputs folder.
- Some models trained using a sample dataset are given in this link.
- It contains encoder, decoder and numpy embeddings generated from auto-enocoders.
- These are obtained by running
torch_train.pyscript.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent