DL in production
Text+Image recognition Trends spotting and analysis for Fashion and Beauty
- Tag Images(hat, jacket, handbag)
- Detect objects(marks)
Q: Model trained end2end? A: Seperately to assure the quality of each model
Transfer Learning
TF+Keras Docker GPU for training CPU/GPU for prod
ResNet, Inception, DenseNet pre-trained models + fine-tuning
Q: Retrain the whole model or only the last layer? A: whole
Q: On-line Learning? A: Yes
Q: Pricing in Training Data ? A: Yes
Manage classes
Thesaurus: one entry point for each class when adding new classes, influence all model
Pipeline Test Data totally external from train/valid set
Re-label => consistency strong label for training
Inference
Data => Kafka => DL(models) => Kafka(metadata) => indexing Kafka based => Decoupling + Robustness + Queue monitoring + Scalability
Two machines with GPU => Kafka
Send ack(commit) only after the processing to avoid crushing during processing
for one partiion, we can only have one consumer, but not reversely
timer => how long it takes to process each batch
Kafka lag
Nvidia-Docker => use GPU
Eg. Efficient LSTM TF XLA spotting instances discount
Learning plateform
generate exercises for personalized homework
aim: understand what the students know, and how well do they know
Word2Vec to measure the difficulty level of exercises
LSTM fot all diff categories, students, exercises, subjects, etc.
Time to respond as a feature to decide difficulty of exercise.
RAMP
DL & GPU
20, 21, 22 Nov lpma-paris.fr/dlhpc