Working with Fashion Models
summary
A convolutional model trained as a multi-task multi-image CNN runs on multiple GPUs using Chainer, trained on about 4 million images (60M in-house) with ~60M parameters; training is hours to days. It yields ~80% overall accuracy and ~73% on color, improves color with 1×1 filters, and exposes learned features via deep‑dream visualizations. The learned visual embedding combines with word2vec plus a triplet loss to extract keywords from noisy product copy.
AI-written summary · about
chapters (12)
- 0:00Fashion data science intro
- 1:54Deep learning for fashion
- 7:25Visual model architecture
- 10:11Multi-image multi-label training
- 13:48Debugging model performance
- 16:03Visualizing learned representations
- 20:39Color representation challenges
- 23:49Improving color detection
- 26:42Deep dream and style transfer
- 27:46Image and text representations
- 32:02Composition and representation takeaways
- 33:56Q&A
related talks
34mComputer Vision and NLP for Multi-Task Fashion Attrb. Modeling
51mCognition and Computation
28mLarge Multimodal Models: Understanding the State of the Art in Compute
33mMachine Learning in Visual Search
42mDeep Learning for Computer Vision
01h 23mMultimodal Deep Learning in the Real World
40mHow to Implement a Production Ready Deep Learning Model in 12 Weeks
01h 01mA practical guide to deep learning