Skip to main content

Week 04

Sept 19 2022​

  • Document on water rescue training as a GQC research idea that builds on GQC's prior work (NIST IoT, EPANET, HEC-RAS). Email RE: Hurricane Fiona hits Dominican Republic after slamming Puerto Rico and more| September 19, 2022/GQC Research Idea
  • send to pavan about the shell file. how to implement --help and see what input I need
  • learning rate tuning
  • Implement f2 score weight (at the end of the run)?
  • Implement F1 normal score (at the end of the run)?
  • Implement the optimal batchsize function?

Try to find the best hyper parameters​

  • learning rate with sewer-ml on colab and implement in compute canada using fastai. Right now validation loss is used as metric. (if you left as default, what is the loss function used for validation loss), Can we change it? Mention this in Thesis.
  • sweep
  • sweep to find the ideal hyperparameter
  • learn.summary() to get information
  • Sensitivity analysis on optimizer and loss function
  • Change optimizer will change the way to calculate loss function
  • Grab the image from Sean video and apply to model to his image and compare to ACCESS database

Talk to Brian​

  • Collimator has a controller that is better than PDI. He has a idea that we can run AI model in notebook to predict the next state and feed the result to the control and asking the system to change the system (such as increase the oxygen level) base on the predicted state.

Sept 20 2022​

  • implementing learning rate tracking in fastai training the model on train00. Submit the job to compute canada. There is error with wandbcallback as the wandb trying to log every parameter when wandb.log() is called before fine_tune()
  • add fbeta as part of metric for fine_tune()
  • find an example of sweep and write to Deven on how I am going to use it
  • use ffmpeg to extract images from DNV CCTV footage

Sept 21 2022​

  • EPANET to simulate the break/leak
  • Need to investigate why there is negative pressure when running PDA in EPANET
  • Fix the issue with wandb.log() by calling it after fine_tune() and submit the job to commpute canada
  • Extract more images from video footages and create its csv file
  • predict the test images got from DNV footage using the Fastai model trained on train00 images
  • Implement F2 score in the learn.fine_tune by calling f2score_multi = FBetaMulti(2,average=None). None is chosen because we want to return the score for each class. We then can multiple each score by its corresponding weight to get the same evaluation as Sewer-ML paper
  • Need to figure out how to use batch size function
  • PACP_Inspections table of the Access Data contains the information regarding the watermain. ASSET_ID of the SanMain shape file can be found in Pipe_Segment_Reference column. The location of manhole can be found in Upstream_MH and Downstream_MH columns. The manhole can be found in SanFitting shape file.
Last updated on by Viraj