Error to run Magiori FCN

[resnow92]

Hi,
I tried to run the create_savedmodel_maggiori17_fullyconv.py to train a model with my 15 bands images but I obtained this error:

I specified the number of channels (15). I tried with the create_savedmodel_simple_fcn.py and all is running well.

Thanks for your help.

[remicres]

There was an error in the documentation: wrong patches sizes during training step (I just corrected it in 3daef29).
You must change the patches sizes for input and output, because the Maggiori model inputs patches of size 80x80 and produce output labels of size 16x16. To generate the two sets of patches, you can use the PatchExtraction application with OTB_TF_NSOURCES=2

[resnow92]

Thanks for your answer.
Why i have to use OTB_TF_NSOURCES=2 ?
Sorry if i don't understand, why we have to generate 2 set of patchs ? you mean the label data have to be 16*16 ?
Last question, can we use this script for 3 classes classifications ? or is it for binary classification only ?
Thanks for your time, I have a too small background in deep learning and I don't understand well how to change parameters and data format to be read by tensorflow.

[remicres]

Setting OTB_TF_NSOURCES=2 enables to extract the two set of patches in one time. If one patch lies outside the image, it is rejected from the sampling, and that guarantee that patches are sampled simultaneously for all sources. If you don't set OTB_TF_NSOURCES, it's harder to guarantee that you have the same patches. Suppose you have 2 images to sample simultaneously, like one label image and one remote sensing image: if your images have not exactly the same extend, some patches may fall outside one image but not the other, and this would lead to a unpaired dataset. Since 5704527, PatchesExtraction also allows to set a no-data value for each source: with that you can reject every set of patches that contains some undesired data in at least one source.

[resnow92]

Ok thanks I think I understood, I have nice results with create_savedmodel_simple_fcn.py but i wanted to try this one.

[resnow92]

Sorry, I tried what you said, so I generate patch extraction with NSOURCES=2. As a result I have as you mentioned, 2 set of patchs:

• 80x80 satellite patchs
• 16x16 satellite patchs
  And I have one label file. Labels are 1 pixel in width.

So my question is, which file to use in the ModelTrain function for the training.source2.il parameter ? As my label are not 16x16 but 1 pixel width.
Or I made something wrong with the patch extraction ?
thanks

[remicres]

I admit that this outlabel parameter is confusing. Actually, the outlabel creates a 1x1 patch image that contains the field value of the input vector data (field name is set with the field parameter). It is convenient for patch-based classification (when your model outputs a single value for a given input patch, and when the class is stored in a vector data field). But as soon as you want to do something else using rasters (e.g. semantic segmentation, image-to-image translation, ...) and that you have output patches as reference (e.g. 16x16), the outlabel is not used anymore.

In your case, you just have to use the 80x80 (for input) and the 16x16 (for labels) patches. Just throw away the 1x1 patches, not used at all by the Maggiori model. By the way, I think that outlabel is not mandatory parameter, but you still have to provide the field parameter: not very logic if outlabel is not used, I conceed!

There is a little room for improvement in PatchesExtraction. In the future major release of otbtf, In next release, it could be great to improve this app parameters.

[resnow92]

Hum OK, many thanks for this very helpful precision ! I think I have all elements I need. Indeed, it was not clear for me, but as I said I am not an expert :)

[resnow92]

I have an error, don't know why, here is the log.
I will explore this issue tomorrow.

[remicres]

There is a typo in your command line: replace is.training=true with is_training=true

[resnow92]

new error concerning shapes of input images ?

My patchs 16x16 have 94 672 pixels height and patchs 80x80 have 473 360 pixels height

[remicres]

I don't know. How many channels have your images? Do you have modified the model? If yes, can you copy/paste it on pastebin or something like that

[resnow92]

I have 15 channels and I put the option -- n_channels 15 when I ran the python script. I didn't change the model.

[remicres]

The maggiori model is used for binary semantic segmentation. You must provide to y some binary labels (i.e. 0 or 1 pixels). The paper: https://hal.inria.fr/hal-01350706/document

[resnow92]

Ah, i think is the thing i missed ! I don't have binary raster label, only label from shapefile.

[remicres]

You can use the OTB application Rasterization to rasterize your vector over your image (otbcli_Rasterization -im yourimage.tif -vec yourvector.shp ... -out outputlabelimage.tif)

[resnow92]

Nice, thanks ! I will try this :)

[remicres]

to be continued in #34