Skip to content

Latest commit

 

History

History
48 lines (36 loc) · 2.8 KB

journal.md

File metadata and controls

48 lines (36 loc) · 2.8 KB
Raw

Journal

Log of attempts:

  • Try 1. We used the top left-most cell for training the model. This resulted in predictions with a cluster of bounding boxes along the corners.

    • Dataset: projects/147301782967/locations/us-central1/datasets/4591639722429775872
    • Model: projects/147301782967/locations/us-central1/models/6748089887754289152
    • Endpoint: projects/147301782967/locations/us-central1/endpoint/7556530001332404224

  • Try 2. We used the 1th cell (2nd from top, second from left) cell for training the model. This also resulted in predictions with a cluster of bounding boxes along the corners.

    • Dataset: projects/147301782967/locations/us-central1/datasets/5740620577362673664
    • Model: projects/147301782967/locations/us-central1/models/8978216128232816640
    • Endpoint: projects/147301782967/locations/us-central1/endpoint/3245177783055286272

  • Try 3. MVP. We placed bounding boxes over ALL the cells in the training data.

    • Dataset: projects/147301782967/locations/us-central1/datasets/6572379133542662144
    • Model: projects/147301782967/locations/us-central1/models/6227079705862864896
    • Endpoint: projects/147301782967/locations/us-central1/endpoint/6490584264529149952

  • Try 4. After creating an MVP online model, we tried to create an exportable, high-accuracy TF model. Unfortunately, our results, a seen in the model evaluations, were not very promising.

    • Dataset: Same as Try 3.
    • Model: projects/147301782967/locations/us-central1/models/1526408012376309760
    • Endpoint: projects/147301782967/locations/us-central1/endpoint/1207721164135202816

  • Try 5. After the previous unsuccessful attempt to create an edge model (Try #4), we decided to try changing the training parameters a little bit: increasing the milli node hours to 100K; changing the test train split to 80/10/10 (was 70/20/10).

    • Dataset: Same as Try 3.
    • Failure. The model evaluations showed an average precision and recall of 0 (!).

  • Try 6. Created a pipeline that:

    • Trains two models, one online and the other Edge
    • Creates a new dataset based upon increased data
      • New dataset created by batch prediction using existing online model (MVP).
      • After batch prediction, the pipeline adds the new map & map coords data into the GCS bucket for the main pipeline flow.

Alternatives considered:

Process / work items:

  1. Get an exportable TF model.
  2. Generate C library (minimal) that calls into TF model
  3. Test C library with real files
  4. OSSPO process for library
  5. Release!