@cordmaur that would be great. Your notebooks and samples were super helpful. There were lots of little details that would have taken me forever to figure out myself.
I could help to check whether these can be combined with your work.
1 Like
@mayrajeo. I took a look. nice work! seems like we were working along similar lines, although your library is addressing more issues.
As I said in another thread, I’ve avoided adding functionality that belongs in the data preparation stage, because I don’t understand the use cases well enough to know what can / should be generalized.
I also didn’t want to add additional dependencies or decide which specific libraries to depend on. For example, I found it much easier to install and work with rasterio, as compared to gdal, but many people seem to use the latter.
I left the specific functions for file I/O out of the main library, and added sample demo functions in the test module that show the code to load the tensor for each channel.
I’ve largely stuck to things that help with the actual training, or with things that need to be parameterized for training.
I also tried not to introduce new subclasses to the basic fastai model, with the major exception of TensorImageMS. Most of the higher-level classes are basically factory classes that produce normal fastai datablocks, learners, etc.
I view the lower-level API as something that could be merged into fastai with no fuss or blockers.
That’s the general philosophy of the library design.
Based on what I saw on the link that you sent me, here are my comments (probably a little cursory since I haven’t had time for a detailed look)
- Opening functions. Should probably not be in the main library. Could be added as a demo or test notebook.
- show functions. Am I right in my analysis that these are various ways to compress the multi-channel data to one image? I’m already covering the false colour and monochrome cases, the other cases (mean_spectra and normalized_spectral_index) seem interesting, and we could bring that functionality in.
- MultiChannelTensorImage - seems more or less the same idea as TensorImageMS, so any additional functionality could be merged into TensorImageMS.
- The various data loaders seem to include io as well, so it would be worth refactoring the io out and see what abstractions remain.
The two completely new areas of functionality are RegressionMask and the MultiChannelTensorImageTuple. Although I haven’t had a chance to work on a problem that requires either, they seem like they could be useful. My big-picture observations here are - are there non MS applications for these abstractions, and how have those been implemented (in Fastai or an extension). I view the MS functionality as being an extension of what would naturally be done in the regular image learning pipeline. In principle I’m fine with adding the functionality, although it may require some restructuring.
What do you think?
Hi Folks,
After some fits and starts with FastAI I’m pretty far down the rabbit hole again. I’m one of the developers for xView2 that a few people have mentioned but it’s been on the application development end as opposed to the model side. I’m looking to take a step across the aisle. That said, may be able to help with some more general Python/Geospatial questions.
Anyhow, I’m curios if anyone has accomplished loading geospatial data directly with FastAI? I ask this with the caveat that I have not read every linked page. I’m a proponent of proposing a fastai.geospatial feature. Additionally, this package is quite good I’ve been told and I have worked with the developers tangentially GitHub - microsoft/torchgeo: TorchGeo: datasets, samplers, transforms, and pre-trained models for geospatial data. Perhaps wrapping that into a DataBlock for FastAI?? I’m spiff balling currently, as you can probably tell. Open to talk geospatial, particularly in the HADR space.
1 Like
Thannk you for sharing this.
DGCustomerFirst
Thank you for sharing this…
Thank you for sharing this…
NCEDCloud login
If you’re still looking, this link was useful to me. Note that it might be outdated due to the lack of updates from the repository maintainers: pip - How to install solaris (on colab or locally)? - Stack Overflow