Week 4

Jan 01/23/2023

1. EPA
   1. Cloned 3 repos that were shared by Jake.
   2. All of them are Django apps.
   3. EPA does not have a coding standard. Going with the defacto most popular option, black and flake8
2. Over-head
   1. Updates from Vannary
   2. Giving her next steps
3. ORSANCO
   1. Got details about the size of the database
   2. Created an additional disk on the VM
   3. Followed steps to mount the drive on the VM

Jan 01/24/2023

1. Documentation of the runs from Vannary
   1. Model fully trained on SD1 --> validation and test scores
   2. Model trained on DNV --> test scores for SD1
   3. Model trained on SewerML --> test scores for SD1
   4. Model trained on DNV --> Tuned on SD1 --> validation and test scores?
   5. Model trained on SewerML --> Tuned on SD1 --> validation and test scores?
   6. What is the class distribution of the test images used for SD1?
      1. What kind of accuracy are we getting on them?
      2. While training using SD1 images, what is the class distribution?
   7. Other insights
      1. What % of images are classified correctly? (when trained on DNV, predicted on SD1)
   8. Are the final 2 videos labelled for SD1?

Jan 01/25/2023

1. There is some trouble using the Transform and then to write it to disk.
2. The issue was solved by manually creating the transforms using PIL.
3. The notebook was created then shared with Vannary

Jan 01/26/2023

New data from DNV

1. Sean is going to share data almost 440 GBs in size.
2. Here in this part of the document, we are trying to decide what are going to do with that data.

Size estimates

1. If a single video is 80 MBs in size, after capturing the original frames, the size of the folder becomes 300 MBs in size. This is the same pattern observed across multiple videos.
2. In our current pipeline, we are saving the original as well as the blurred images.
3. ⚠️ Note: We capture frames every 1 second in the video.
4. This means that, after we capture the frames, the size of the extracted images will be 3.75x of the original video size (for the original frames) and 3.75x for the blurred images.
5. Hence, the total size would become 7x of the original size of the video. (400 x 7 → 2.8 TB of storage)
6. If we wish to save augmented images as well, the size of the images will keep on multiplying.

Data analysis

1. Once the data is on our drive, we need to
   1. intelligently analyze the data
   2. extract insights, class distributions, defect density in the images, and much more
2. Once, we have some insights, we need to create a baseline model for comparision,
   1. training on a set number of epochs, fixed architecture, no transforms and progress iteratively

Observations from current training routine

1. EDA (exploratory data analysis) could be a lot more thorough
   1. We don't have any systematic insights about the SD1 images that Vannary has labelled
   2. Same is the same with the images
2. ResNet backbones aren't going to cut it, need a better architecture
   1. With this scale of data, we should probably train from scratch.
      1. No use of pretrained models
3. Better choice of backbone
   1. EfficientNet
   2. Xception
   3. or something else?
4. Some really nice notebooks from Jeremy comparing the best models available in CV right now. The notebooks are now 6-8 months old, meaning the best models are likely to change.
   1. Which image models are best? | Kaggle
   2. The best vision models for fine-tuning | Kaggle

Jan 01/27/2023

1. Discussion with Sudhir
2. Next steps with Vannary
   1. Next steps on CCTV
   2. Next steps on DNV

Meeting with Brian

Questions discussed (with answers below)

1. List of requested features
   1. Ability to read model defined parameter within the python script block
   2. Ability to update submodel parameter from the notebook using collimator python-api
   3. Provide a menu option or script to convert old submodel block to the current submodel block
   4. Ability to handle algebraic loop problem (the unit delay block uses global step and does not address the problem)
   5. Ability to search for a model/file in the project folder
   6. A way to organize different files present inside the project folder (subfolders?)
   7. Ability to save runs? If a simul takes 3 hrs to finish, and you forget to save the results, will collimator have the ability to save those results?
2. Marketplace mechanism, if GQC creates custom blocks inside Collimator, what would the licensing mechanism be? Where would those blocks get hosted?
   1. How do we encrypt custom blocks?
3. Share some updates about the publication
4. Share the results from GQC collimator's workspace
5. We want to have a shiny interface to the Collimator model and results that would be similar to the SCADA HMI. How would we go about doing that?

Answers from Brian

1. List of requested features
   1. to-be-filled
   2. to-be-filled
   3. to-be-filled
   4. to-be-filled
   5. Are upcoming features
   6. Are upcoming features
   7. You can download all the results from the top-right menu bar. Results are stored for 3 days and then new results are overwritten in a cyclic fashion.
2. He would need to discuss this with Kevin
   1. This is an upcoming feature. No ETA on the topic though.
3. This was done. Updates were shared
4. Results were shared. Results from the Grafana Dashboard were shown to Brian.
5. He would need to discuss those things with Kevin.