Skip to main content

Week 03

Feb 13​

  • SD1 fine tune
  • Image Argumentation
  • Weight loss function (recalculate the weight). The rarest label gets the biggest weight.

Feb 15​

This video was not used in training:

  • SANMN08983_19.mp4 (MSA at the beginning)

I found additional 7 videos that were not used to train the model. Somehow these videos were not uploaded to the google drive.

  • SANMN00478_30.mp4
  • SANMN02750_37.mp4
  • SANMN03277_31.mp4
  • SANMN03861_851.mp4
  • SANMN05054_33.mp4
  • SANMN08060_35.mp4
  • SANMN08061_36.mp4

from 73 videos used to train the model (excluding SANMN08983_19.mp4), there are 71 unique sewer ID.

  • SANMN03959 have two videos (SANMN03959_67.MP4 and SANM03959_70.MP4)
  • SANMN04050 have two videos (SANMN04050A_83.mp4 and SANMN04050_82.mp4)
  • SANMN86600 was not found inside the shape file.

Feb 16​

  • try fine-tuning DNV model with SD1 images using google Colab

  • Create a data flow diagram processing new uploaded DNV videos

    • WRC vs PACP
    • Get the runs going with PACP videos
    • Videos that did not have distance can be used for prediction to get the frame where the defect is found

  • Move forward with water level model (Multi classes)

  • Write up on Pipe Sleuth with new info

  • Continuous learning model

    • Example
    • Run with 80 videos and fine tune with another 80 videos
    • Need to make sure there have the same number of class to preserve the original class labels

  • Compare the results running with Single node Single GPU vs Single node multi GPUs

  • Check the conditional data (different from access database)

  • run the model with 7 additional videos (73+7 = 80 videos)

Presentation​

  • Start creating dataset slide for each time the video is added

  • how many length of main is covered. (both from access database and from the last frame of each videos)

  • Classes and Count are uppercase. Make the title smaller

  • Use Sewer Main instead of Sanitary main pipe. Location of 70 sewer mains as the title and try to find a different way to address SANMN86600 was not found.

  • Prepare Datasets slide. _> change to Prepare Data

    • Most of the videos have 30 frames per second (as the first bullet)

  • Baseline model slide

    • BackBone model
    • change weight from previous model to Pretain?
    • Number of Batch -> Batch sizes

  • Different Model Parameters -> Different processing pipeline (creating proccessing pipeline with visual

    • Changing data (image augmentation replace the original images)
    • Baseline plus image augementation ...etc
    • visual: incoming data -> model

  • Performance metric -> Performance Metric

  • Pre-trained Models -> Prediction accuracy from different processing pipeline using pre-trained models.

    • Split the graph to half and write the conclusions to the left.
    • Have a graph showing Pre-trained is better than no pre-trained model

  • Prediction Example -> Accurately predicted ISJ and SRI

  • rename prediction Example:

  • determine the location of defect using the combination of AI and GIS

    • if we know the access point manhole, we might be able to use this to find the location of defect from video that did not distance.

  • list of images misclassified by inspector.

  • Q/A

  • issues found (last_slide).

  • Next Steps

    • process the new videos
    • add continuous learning (fine tunes)
    • Model Detailed and Summary PACP

  • Make set of images of different defects

    • get the spectrum of images of different defects
    • illustrate how varies the images of different defects
    • lovely Tensors
    • histogram of the rgb -> pick the image that has different rgb

  • Model summary slide:

    • ResNet50: how many trainable parameters

Azure​

  • create resource group
  • create resource in the resource group - use cognitive services multi-service account

Condition DATA​

  • revived the dnv defect classification model pipeline. Include which column of access database that we use
  • create another notebook to deal with new condition data.
    • transform the csv function: changing the column name of the incoming condition data
  • create the github issue and describe the analysis of condition data
  • When query wizard, joining the jobnumber between cctv header and cctv detail, the order of code is messed up.
  • the data did not come in right order (CCTV_detail), need to sort by jobnumber and then sort by distance, then sort the code so that ST and AMH will always come before WL and MWL to get the right order.
  • Filter out only video with PACP code by
    • filter Date > #2016-12-31# (does not work when connect with detail csv)
    • filter using MWL at distance 0 then query with CCTV_header to get name of videos with PACP

Feb 17​

  • try using multi-gpus in the interactive nodes. While running fastai_distributed.py, the job only ran using 1 GPU not on 4 GPUs.

Meeting with Dr.Lence​

  • we went over the presentation
    • create a table for fbeta, f1, f2
    • make sure to create a visualization for all the pipeline
  • she suggested me adding more stuff to my outline. We will on sometime next week (probably on Friday) to review it.
Last updated on by Vannary