Has anyone else upgraded to ArcGIS Pro 3.5 and is testing the deep learning tools package for 3.5? I had no issues using the versions of the tools for ArcGIS Pro 3.3 and 3.4, but I just installed 3.5 and ran into the following issue: "Error:PytorchStreamReader failed reading zip archive: failed finding central directory.Training was not successful."
I did not get the error with version 3.4, and I am using a set of training data that did not give such an error previously. If there's no suggestions to resolve this, I will have to downgrade to 3.4.
Solved! Go to Solution.
FYI to all who might be following this - apparently the issue I had was idiosyncratic to my workstation, as a similarly configured workstation did not experience the issue. With the assistance of Esri colleagues, I was able to get model training and inferencing working on my computer, following reinstallation of ArcGIS Pro 3.5 and the associated deep learning framework package. I am still unable to train a DL model using the RESNET18 backbone model, but with other backbone models, I'm not getting the PyTorch error I was getting previously. Thanks to the Esri colleagues for their suggestions!
FYI - The solution to this issue, as suggested by Esri colleagues, was to delete the C:\Users\username\.cache\torch\hub\checkpoints folder [with ArcGIS Pro closed] and then re-opening ArcGIS Pro, which generated a new checkpoints folder.
Apparently, my computer had an old version of a file called resnet18-f37072fd.pth, which triggered the issue. By deleting the checkpoints folder (and that .pth file), it caused ArcGIS Pro to generate a new version of that file, which resolved both the model training and inferencing (classification) issue I was dealing with.
Kudos to Pavan and Sangeet from Esri for finding this solution!
By the way, I was able to execute the deep learning model training in an ArcGIS Pro v. 3.4 instance I have on a different machine. However, when I move said model to the machine with ArcGIS Pro v. 3.5, I get an error. I also even get an error with an older deep learning model that worked under v.3.4 (screenshot below). Something tells me that the issue is with the deep learning package for version 3.5, but I am open to suggestions.
And by the way, yes, the deep learning package is properly installed (screenshot below), per the instructions from https://support.esri.com/en-us/knowledge-base/how-to-verify-the-installation-of-deep-learning-framew....
FYI to all who might be following this - apparently the issue I had was idiosyncratic to my workstation, as a similarly configured workstation did not experience the issue. With the assistance of Esri colleagues, I was able to get model training and inferencing working on my computer, following reinstallation of ArcGIS Pro 3.5 and the associated deep learning framework package. I am still unable to train a DL model using the RESNET18 backbone model, but with other backbone models, I'm not getting the PyTorch error I was getting previously. Thanks to the Esri colleagues for their suggestions!
FYI - The solution to this issue, as suggested by Esri colleagues, was to delete the C:\Users\username\.cache\torch\hub\checkpoints folder [with ArcGIS Pro closed] and then re-opening ArcGIS Pro, which generated a new checkpoints folder.
Apparently, my computer had an old version of a file called resnet18-f37072fd.pth, which triggered the issue. By deleting the checkpoints folder (and that .pth file), it caused ArcGIS Pro to generate a new version of that file, which resolved both the model training and inferencing (classification) issue I was dealing with.
Kudos to Pavan and Sangeet from Esri for finding this solution!
Hi Dan, I think you posted the link to the current post. Is there somewhere else I should post this issue to? Thank you!
edited, and repeated here
@servir_bz1 Can we connect and troubleshoot it? my email address is pyadav AT esri DOT com?
Thank you, Pavan. I just emailed you.