Happy Tuesday, one and all.
The Shopin team has been hard at work relentlessly pushing forward different aspects of our solution for the world of retail on the blockchain.
Our current core focus has been advancements in our Visual AI engine. This involves intensive mapping and tagging of new categories in fashion retail, and training our artificial intelligence models.
These models power our ShopperIQ onboarding games for retailers as well as recognition of a retailer’s inventory, the shopper’s uploaded images and purchase history.
Below is a string of updates from our team:
Things to know:
Similar recommendations — recommendations within the same category:
Eg: Recommending more similar bags to the bags you like.
Cross-category recommendations — recommendations across categories:
We recommend items from other categories to the one you choose. Eg: You like a scarf and we recommend a shirt, jacket or other “non-scarf” items that you may like.
Complete the look-suggest items that look good with your choices
This is one of our most exciting products. We recommend other items that would look great to complete the look you’re purchasing. Eg: That scarf would look great with those shoes, briefcase, gloves, and suit. It’s a programmatic solution that delivers the experience of having a personal stylist with you wherever you go.
Product Update Commits Update 10/23/2018:
Complete the look (CTL)
- Development for experiments on multi-layer BiLSTM
- Research on hyper-parameters tuning for BiLSTM
- Evaluation of co-occurrence model on test data
- Submitting multiple training jobs to test hyper-parameters
Product Normalization — Text:
- Finalizing dataset for NER — for all categories
- Deploying of NER for Intimate wear in AI pipeline
Product Normalization — Image:
- Training of Jumpsuits length model
- Bench-marking of the multi-task model with separate Image models
Similar Recommendations:
- Training for Skirts & Shorts model in progress
- Integration of multi-layer in Bags model
Object Detection:
- Deployment of “human” classifier
- Testing of RestAPI
- Deployment of Internal OD
Product Update Commits Update 10/10/2018:
Object Detection:
- Training for Internal OD — Iteration 2 has been completed. Currently performing error analysis
- Finished error analysis for men's onboarding — Iteration 2
- Working on lambda to sagemaker conversion
Complete the look (CTL)
- Completed inference code for multidimensional co-occurrence model. While running the final model, the team observed that it’s taking a considerable amount of time to give outputs due to the very large size of the matrix. Currently, this task has been put on hold as the team will need to do research for a possible solution
- Previously the model was showing output for “complete the look” in random orders, i.e.- first tops, then shoes and then jeans. The team has modified the code so that the output is shown in a fixed logical order, i.e. — first tops, then jeans & then shoes
- Calculated Top K 96% precision for multi-query multi recommendation system
- Modifying the AI pipeline lambda based on the results of the testing. This is expected to complete today
Product Normalization — Text:
- Started working on the Classifier model for category & gender. Currently, we are utilizing output from Tree search logic. The data has been sent for tagging which will be used as input later on
- Creating a combined PN model which will be deployed on lambda
Product Normalization — Image:
- Multitask learning experiment on the pattern
- Experimenting for Jumpsuits neck model. Currently, the Top 1 accuracy that we have been able to reach is 75%. The major portion of inaccuracy is due to one jumpsuit having multiple styles for the neckline, resulting in confusion during the tagging the “Neck” style. The base data has been sent for tagging which will be used to improve the model
- Training the model: Shorts
Similar Recommendations:
Trying to implement multitask learning so that a single model can be used for multiple attributes
DevOps/Engineering:
- Performing a comparison between AWS & GCP for CICD
Product Update Commits Update 10/8/2018:
Object Detection:
- Training for shoes finished on complete model images. The team is currently performing the error analysis for the same
- Training completed for suit model for men's Onboarding — Iteration 2
- Started lambda to sagemaker conversion of OD model
Complete the look (CTL)
- Working on Inference coding for multi-dimension co-occurrence model, expected to complete today
- Development of autoencoder for training and inference is in its final stage and expected to complete today itself
- Completed first round of testing on AI pipeline. The team found an issue in which if we feed multiple files to the AI pipeline at the same time, there is no output for most of the files. For example — when we tried inputting 60 files to the pipeline, output was only available for 8 of them. The team is currently looking into it
Product Normalization — Text:
- Finalizing the dataset which will be used as input for classifier model. Once finished, the team will send the data to taggers for tagging and train the model on that
- Evaluation of NER Iteration-2 for intimate wear
Product Normalization — Image:
- Performing error analysis for Neck model of jumpsuits. The team noticed there are multiple styles for the same collar, i.e.- A neck can come in the style of V- Neck as well as collared while model will give only one output (V-Neck or Collared) which results in the lower accuracy of the model
- The data for shorts length model has been sent for tagging. Once finished, we will train the model for the same
- For the Coat & Jacket neck model, we noticed that the model is confusing between the collar of coat and tops. It picks up the collar of tops and gives a result based on that. Also, there can be multiple neck style for the same coat. We are currently looking into possible ways of dealing with these challenges
- Multitask learning experiment on patterns
Similar Recommendations:
- Finalized the shoes dataset for tagging and sent it to taggers
- Trying to implement multitask learning so that a single model can be used for multiple attributes
DevOps/Engineering
- Planned a sprint to do a check between AWS & GCP for CICD and started working on the same
Product Update Commits Update 10/5/2018:
Object Detection (Onboarding):
- Training for shoes images from complete model images done. Currently working on evaluation for the same
- Retraining Suiting model for men’s On-boarding — Iteration 2
- Complete the look team was facing issues in the OD model where they were seeing a drastic drop in the images count after the OD run. Helped the team in rectifying it
Complete the look (CTL)
- DevOps for multi-dimension co-occurrence model completed.
- Working on Inference coding for multi-dimension co-occurrence model
- Resolved issues with the storage of High dimension matrix in multi-dimension co-occurrence model
- Testing the lambda functions for both inputting training images into the AI pipeline & data shaping (JSON creation).
- Explored alternatives to ANNOY files. FALCONN & NMSLib seems to be good alternatives and will require testing to measure the performance between the three
- DevOps for Auto-encoder loss completed
Product Normalization — Text:
- Changing tree search code for changes in logic & taxonomy
- Trained logistic & neural-net classifier model. The neural-net classifier is giving a high accuracy of 92%
- Completed Sprint planning for the next 2 weeks
- Evaluation of NER Iteration-2 for intimate wear
- working on custom code for Protege so that the annotation task can be distributed to multiple taggers without any duplicates
Product Normalization — Image
- Testing models for all the jumpsuit attributes
- Working on Jumpsuit neck model to improve accuracy
- Performing error analysis on OD output for Sweater, Coats & Jackets
Similar Recommendations:
- Finalized classes to be trained for every style
DevOps/Engineering:
- Built a POC for DynamoDB
- Working on benchmarking of FAISS with Annoy
Product Update Commits Update 10/4/2018:
Object Detection (Onboarding):
- Sent additional Shoes data for tagging. Will be used to improve model’s performance for detecting shoes from complete model images
- Men’s model for onboarding- Iteration 1 is complete.
- Retraining Suiting model for men's Onboarding — Iteration 2
- Improving pre-processing scripts and also, documenting the code for the same
Complete the look (CTL)
- Completed the 2nd Sprint, training the model on ASOS data (matching Tops to Bottoms). Tested the model against Tommy Hilfiger etc
- DevOps for multi-dimension co-occurrence model is expected to be completed today. Will start working on coding tomorrow.
- Created lambda function for inputting training images into the AI pipeline. It also takes care of creating batches of 100 input files.
- Working on lambda for data shaping (JSON creation) for training the CTL model.
- Kickstarted next Sprint for CTL.
Product Normalization — Text:
- 2nd iteration of Named Entity Recognition (NER) model trained to extract 8 important attributes like Support, Fabric, Neckline, Pattern,
- NER model evaluation on the precision of score of 91% across all vital categories
- Performed error analysis of Style & Sub-Style extraction code output from tree search logic
- Pre-processing of the complete dataset
- Training count-vectorizer model. Will train logistic & neural net classifier model after this and perform error analysis
Product Normalization — Image:
- Committed Color & Pattern demos to Github
- Testing models for all the jumpsuit attributes
- 2nd iteration of single multi-task learning model is complete. One single model is able to detect Neckline, Color, Length, Sleeve and Category with an average 92% Top-1 and 95% Top-2 accuracy without any test-time augmentation.
- In production, this allows a reduction in the number of models being maintained (from 4 to 1) for attribute prediction.
- Validated sleeves output from OD for Jumpsuits (using the dress sleeve model). Accuracy — 97%
- Tried Jumpsuit neck model on two architectures, densenet(77%) & squeezenet (69.5%). The low accuracy maybe contributed to duplicates in the input dataset. Will be sending the data for tagging
- Performing error analysis on OD output for Sweater, Coats & Jackets
Similar Recommendations:
- Prioritized the future pipeline based on the Inventory stats gotten at the Category-Style Level.
- Discussion for model standardization by building one single classification model has been kick-started.
DevOps/Engineering
- Deployed combined PN model for intimate wear on Fargate
- Resolved pending issues in CICD of OD model
- Completed the dockerization of tagging tool
- Created lambda function for batching of retailer CSV
I hope you’ve enjoyed the update — We will continue to bring you more of these with regularity to give you a window into our technical development.
Thank you for staying tuned in.
