89 points by fashiontech 6 months ago flag hide 21 comments
fashiontechguru 6 months ago next
Fascinating project! I've often wondered how fashion recommendations work. I'd love to learn more about the technical details behind this AI-powered system.
aiexpert123 6 months ago next
Definitely! We trained a deep learning model with a large dataset of fashion images and descriptions, using style transfer and NLP techniques to learn the connections between different garments and create personalized recommendations.
machinelearningfan 6 months ago next
Would love to learn more about the NLP techniques! Can you give an overview of the methods you used to analyze and create connections with the text descriptions?
fabulouscode 6 months ago next
@MachineLearningFan: Absolutely! We relied heavily on word embeddings and RNNs. We wanted to understand how different words could be related in a description, as well as identify the most relevant words for recommendations.
dataanalystberlin 6 months ago next
Thanks for explaining the NLP techniques. Do you know of any resources that could help someone expand from this to image-based descriptions, like from Instagram posts or blog articles?
dataninja 6 months ago next
I have an idea of the resources to use for community based metadata collection. Do you think exploring this would help enhance the recommendation system?
deeplearninglover 6 months ago prev next
What types of connections do your style transfer techniques make? Do they focus on color schemes, patterns, or certain types of garments? Is it possible to learn more about this in the code?
aiexpert123 6 months ago next
@DeepLearningLover: Our style transfer methods specifically extract information about color schemes and patterns in garments for the connections, but could be expanded by using more advanced methods like CycleGAN or StarGAN.
deeplearningfan 6 months ago next
@AIExpert123 Would love to learn more about how you're currently extracting information for the color schemes and patterns! I'm guessing you're using a neural network model?
aiexpert123 6 months ago next
@DeepLearningFan: That's correct! We used a combination of pre-training12-layer VGG16 on the Fashion Style Dataset for color information and pre-trained CycleGAN for pattern extraction. This completely answered your question?
fashiontechguru 6 months ago prev next
@AIExpert123 - thanks for the detailed response! Are there any publicly available datasets or tutorials that could help me get started with building similar models?
aiexpert123 6 months ago next
@FashionTechGuru: Unfortunately, we mostly used in-house data for the model. Some public datasets you could check are the DeepFashion and Fashion Gen datasets. For tutorials, I would recommend the Tensorflow or PyTorch websites.
fashionistabot 6 months ago next
That's interesting about the datasets! Are there any datasets you wish you had access to or any data you'd like to gather more of for future iterations?
fashiontechies 6 months ago next
@FashionistaBot: For future iterations, I'd be looking to aggregate data from multiple sources such as blogs, influencers, and stores to create a detailed and continually updated dataset for training.
dev_cool 6 months ago prev next
Great work! Have you considered incorporating size and fit data into your model's recommendations? I think this would greatly enhance the user experience.
dev_cool 6 months ago next
@dev_cool: We originally planned to implement size and fit, but ran out of time. We really want to add it in future updates. Great point!
codeitbaba 6 months ago prev next
Could you open-source the backend as well? It would make a great educational resource to see the implementation details. :)
opensourcemate 6 months ago next
@CodeItBaBa: We'd love to open-source the backend, but it's currently tightly coupled with our proprietary systems and components. Perhaps in the future, we can break it out and share it with the community!
datasciencepro 6 months ago next
That's totally understandable, FashionTechGuru.czyj miałbyś ochotę na współpracę z open source community ?
opensourceloon 6 months ago next
@DataSciencePro: As a possible contributor to the open-source solution, I'm eager to help! What kind of features would you like to see in the final product, FashionTechGuru?
aiexpert123 6 months ago next
@OpenSourceLoon: It would be great to include better regularization techniques, DR techniques and user feedback through reward Models. This would improve the generalization and also personalization of recommendations. But first things first, would you like to review the source?