search

Esri / raster-deep-learning

ArcGIS built-in python raster functions for deep learning to get you started fast.
Apache License 2.0
190 stars 88 forks source link

MaskRCNN Doesnt work in Arcgis 2.6.2 #35

Open Boukary opened 4 years ago

Boukary commented 4 years ago

I tried this maskrcnn model for building segmentation before and it was working, but now it is not. I installed all the deep learning frameworks following this guide : https://github.com/Esri/deep-learning-frameworks/blob/master/README.md?rmedium=links_esri_com_b_d&rsource=https%3A%2F%2Flinks.esri.com%2Fdeep-learning-framework-install

I receive this error : 2020-10-05 (2) can you help me with this, please?

lingtangraster commented 4 years ago

In ArcGIS Pro 2.6, if you use the new deep learning frameworks installer provided by Esri here, Tensorflow 2.1.0, Keras 2.3.1, and scikit-image 0.17.2 will be installed. These versions are needed in other DL functionalities/workflows in Pro 2.6, but they are incompatible with the current Pro built-in Keras MaskRCNN support, mainly because Matterport implementation of MaskRCNN their repo has not been upgraded to support Tensorflow 2. Therefore, my suggestion is either install the compatible libraries (TensorFlow 1.14.0, Keras 2.2.4, and scikit-image 0.15.0) for your current model, or you could train a MaskRCNN model inside Pro 2.5/2.6 using our Train Deep Learning Model Geoprocessing tool (using Pytorch/fastai backend). The whole instance segmentation workflow can be done inside Pro, from exporting image chips with Meta Data Format “RCNN Masks”, to inference using Detect Objects Using Deep Learning tool.

lingtangraster commented 4 years ago

@Boukary Please refer to the above comment, thanks

Boukary commented 4 years ago

Thanks a lot @lingtangraster

Actually, that was my next step after I posted this question yesterday, I have created a conda env with compatible libraries following this guide https://pro.arcgis.com/en/pro-app/tool-reference/image-analyst/pdf/deep_learning_install.pdf, but i still get the same error.

Is this the MaskRCNN implementation used to build that model? https://github.com/matterport/Mask_RCNN Can I use your model directly with inference code from the matterport repo ?

Thanks a lot