Select to view content in your preferred language

How to train deep learning model with image chips from multiple rasters, How to train deep learning model with image chips from multiple rasters

1799
4
08-01-2019 08:07 AM
PhilippZacharias
Occasional Contributor

I'm using a similar approach to deep learning with ArcGIS API for python like this tutorial describes:  https://learn.arcgis.com/en/projects/use-deep-learning-to-assess-palm-tree-health/lessons/detect-palm-trees-with-a-deep-learning-model.htm

how can merge multiple image chip exports from ExportTrainingDataForDeepLearning-Tool in one training? Specifically the FastAI SingeShotDetector?

And another thing I was wondering: How can i test the results of the SSD, apart from using DetectObjectsUsingDeepLearning Tool and looking at the resulting SHP file? Is there no test method?

0 Kudos
4 Replies
PhilippZacharias
Occasional Contributor

Nobody has an answer to this?

0 Kudos
SandeepKumar11
Emerging Contributor

Hey, Philipp we can always add more training data to a folder where we have already exported some training data, it will append data not override. This will fail if you change any parameter like chip_size in a later batch. The ssd results can be visualized in the notebook itself using the sho_results() method.

0 Kudos
by Anonymous User
Not applicable

Hey Philip did you try it ?

0 Kudos
RichardFairhurst
MVP Honored Contributor

With ArcMap 10.6 the Export Training Data tool has a Start Index parameter that will overwrite the existing tiles if you specify a starting index that overlaps any of the indexes used by previous tiles.  If you are using that version of the tool you have to manually set the Start Index to a number that comes after the last existing tile index to append tiles and not overwrite tiles.

I believe the tool automatically never overwrites and always appends only if you are using the Export Training Data tool at ArcMap 10.7 or ArcGIS Pro 2.4.  The ArcMap 10.7 and ArcGIS Pro 2.4 versions of the tool dropped the Start Index parameter and apparently starts at 0 if there are no existing tiles in a directory or else it starts indexing tiles after the last used tile index to avoid overwriting the outputs.  I think the notebook requires or assumes that you are using the ArcMap 10.7 and ArcGIS Pro 2.4 version of the tool.  This version of the tool also creates different output files that are needed by the learn.ai module that the ArcMap 10.6 version did not create, so I am pretty sure you have to use the ArcMap 10.7 and ArcGIS Pro 2.4 version of the tool to follow the notebook.

0 Kudos