Pre-trained model for inference

[rohit7044]

First of all thank you for this wonderful work. This is a beautiful repo and I am pretty sure many people gets a lot of benefits. However due to lack of proper documentation and me being new to this field, I am having a hard time understanding tutorial.
Basically I have an basic plain RGB image and I don't have any satellite data. I want to apply various segmentation methods/models to remove vegetation, roads, buildings,trees,terrain,lanes, Can somebody give me any headsup maybe links or a tutorial on how to do that?
Thank You

[roybenhayun]

Hi Rohit, as a new comer to TrochGeo myself I can share the learning path I went through.

1. do the basic getting started tutorials. currently there are Getting Started, Custom Raster Datasets, Transforms, Indices, Lightning Trainers, Pretrained Weights. These are all basic and must-know. don't take long and pretty easy to follow.
2. watch the TorchGeo intro by Caleb
3. watch IADF School 2022: Change Detection (TorchGeo) also by Caleb Robinson
4. read the original paper. it's a must read IMHO.

there is probably more but these were the first steps I took before embarking on a real project. to start your own project, find an example that does something similar to what you need and work it from there.

I work on semantic segmentation and started from https://github.com/microsoft/satellite-imagery-labeling-tool/blob/main/example/training-example.ipynb

my 2c. HTH.

[rohit7044]

Thank you for the reply. However I am not looking for training a model or some sort. I just want to perform inference on an RGB image and segment various classes. That's the only scope between me and torchgeo right now. Do you have any leads on that?

[roybenhayun]

so you want to run prediction, but without training, right?
hmm. haven't seen this yet and would be also interested to see an answer on this.

[adamjstewart]

I would add to this list that TorchGeo uses the same API as torchvision/etc., so if you have never used torchvision before, it helps to go through some of their tutorials. We're planning on adding a lot more tutorials in the near future, we want to start having official tutorials at popular ML/RS conferences.

[rohit7044]

Yes, We can preprocess the image according to the model input and run inference since Torchgeo already has pre-trained weights. I am looking for some headsup or any information if it's even possible on this library. Regards, RD

…

On Tue, Oct 3, 2023, 04:31 Climate-Eyes ***@***.***> wrote: so you want to run prediction, but without training, right? hmm. haven't seen this yet and would be also interested to see an answer on this. — Reply to this email directly, view it on GitHub <#1607 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ALRFYF7LIJH3OPFDWBWDNOTX5MQDVAVCNFSM6AAAAAA5OL6JCGVHI2DSMVQWIX3LMV43SRDJONRXK43TNFXW4Q3PNVWWK3TUHM3TCNRZGE4DS> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[adamjstewart]

It's easy to do. You could:

1. Instantiate a model
2. Load the pre-trained weights
3. Loop over all images
4. Make predictions for all images

However, none of our pre-trained weights are pre-trained for semantic segmentation, and certainly not for the exact set of classes you need to use. You'll still need a labeled dataset to fine-tune the pre-trained weights. Pre-trained weights are designed to make this quicker and easier, not to skip it altogether.

[rohit7044]

Aah! Interesting
Thanks for this information