We recently released two blogs discussing the new GeoAI toolbox in ArcGIS Pro 3.0. * AutoML in ArcGIS Pro 3.0 * Text Analysis using Natural Language Processing (NLP) in ArcGIS Pro 3.0 This discussion is for questions related to the blogs or the tools themselves, or to share how you are using these tools or any feedback you have on them.

Thank you for the quick responses Karthik - answers below: * Training took 50 seconds. screen shots of the input and output attached. * Basic model training on a single ecozone fails in 3 seconds for all model types (warning 260151). * Display name is the same as actual raster name. * Zip attached with the 4 ecozone types (corresponds to all screen shots I've shared) and the successfully trained model (.dlpk) attempting to predict 'ecozone' based on the other six variables in the file. I could clip rasters to a smaller area, but I still wouldn't be able to attach those without exceeding upload limit. 4class_trainModel_input.png 4class_trainModel.png mdl_4types.zip

nav[aria-label="Primary Navigation"] { padding: 0; & ul { list-style: none; width: 100%; display: flex; flex-direction: row; justify-content: start; align-items: start; gap: 30px; padding: 0; & li { margin: 0; } & ul li { list-style: none; } } }

Blog discussion - GeoAI toolbox (AutoML and Text Analysis tools in ArcGIS Pro 3.0)

NicholasGiner1

We recently released two blogs discussing the new GeoAI toolbox in ArcGIS Pro 3.0.

This discussion is for questions related to the blogs or the tools themselves, or to share how you are using these tools or any feedback you have on them.

Encontre mais publicações marcadas com

Analysis

Accepted answers

All comments

KManl

Hi,

Quick question:

I am training using autoML and have successfully done this before with the tool, but now the training tool will not work. Every time I try to train the tool shows me it is processing features, and then gets stuck at the same point every time and eventually fails. I have identified that it is getting stuck processing the one training distance feature. I have tried repairing the geometry, redownloading the data, etc. but nothing helps it always gets stuck at the same point of processing. This is also weird because I have used this exact same data for training a distance feature previously and have changed nothing. it just doesn't work now. Any help would be amazing, thank you.

NicholasGiner1

Hi @KManl, thanks for the question. So you've used the Train Using AutoML tool successfully, and it just started not working? If you remove the one problematic training distance feature, does the tool run successfully?

KManl

Hi @NicholasGiner1,

I have fixed the issue now. But I have run into a new one. I am running the autoML tool to predict species richness (predict to raster) based on all 19 bioclimatic variables (rasters). When predicting for much different future scenarios the prediction result is always the same even though the model inputs are different for each scenario (i.e. under different future climate scenarios). What am I doing wrong here?

Thank you for your help.

jiewilliam

Hello, I want to ask why I put grid data in the training grid to explain that the dataset does not exist or is not supported.

PriyankaTuteja

@jiewilliam Could you provide data and steps to reproduce the issue?

PriyankaTuteja

@KManl Could you provide data and steps to reproduce the issue?

jiewilliam

First of all, thank you very much for your reply. I have solved the above problems. But I have a new question to ask you. When I use AutoML to predict, why will memory continue to grow, leading to the final software flash back? At the same time, I also use pychar's arcpy to try this data, but still encounter the problem of memory continue to grow, leading to software flash back. My running memory is 16G, with 3060 gpu and i5-12450 cpu.I don't know if it's my software settings. When running programs, the utilization rate of CPU and gpu is very low, but the memory has been growing.

PriyankaTuteja

What are the total number of records In prediction set ? Do you see this behavior while predicting even when the number of records is small ? Or does it happen when number of records exceed a specific number ?

jiewilliam

That happens when the number of samples is 31488499. When I reduce the number of samples to 85555, it can run normally.

PriyankaTuteja

@jiewilliam Wow!! that's big number!!

I think 31M records is too much to be processed on a 16GB machine. I do not think RAM is sufficient in that case. I would suggest to get the prediction results in smaller chunks.

ChristopherBerry1

Hi Nicholas,

Thanks for the Blog. I am reasonably new to ML and have previously had some limited experience with RF. I have begun trialling the AutoML tool in Arc and have a number of questions I was hoping I might be able to get you to help me with.

When running the application with the 'Algorithms' section left blank the only alogrithm being used initally was a decision tree. When I tried manually selecting other models it would not process and give me an error (as seen in the below Grab 1). I think I have resolved this by increasing the time limit from the default 60 to 240, would this have been the issue? Now when I select all the algorithms, or leave the mode field blank, a Decision Tree and a Light GBM algorithm are used (As seen below in Grab 2). Is this still ignoring the other model types? I also note that in my example multiple decision trees are being used (See Grab 2). Are these just Decision Trees with different hyperparameters?
In Grab 2, and in your example, there is an ensemble model. Is this model a combination of all the models or a selection? Is there any way to investigate the statistics of this model with regards to the other models it has used, the weighting etc?
Following on from model training is it the 'best' model from the ML Leaderboard used in the *.dlpk file? Is there a way to select the algorithm you want to use, for example if I wanted to use the ensemble as oppose to the Decision Tree?

In Predict using AutoML the fields are somewhat different to the Train tool. For the 'Input Prediction Features' section, if I chose my points shapefile which has been used in training, how does it know which field to predict? Is this somehow defined in the *.dlpk file?

Thanks again for your blog and any help you may be able to offer in helping me grasp the fundamentals of the tool.

GRAB 1

GRAB 2

SupratimBanik

Hi Christopher,

I will try to answer some of your questions:

1. If you leave the algorithm feild blank, then the tool will try to fit all available models on your data. However, if some of the model training fails during the process then that model is ignored and the tool proceeds to fit the next model on your data. In your case it looks like the a decision tree and a LightGBM models were trained and the other models couldnt be fitted.

Second part of the first question - Yes. These are different variations of decision tree with different hyperparameters and different variable combinations.

2. Yes. Ensemble is a combination of multiple models. The composition, performance and other details of the ensemble model can be found in the report that the tool generates. You can also use this report to compare other aspects of the evaluated models. Please note that the report is an additional optional parameter which you can choose to generate (Please lookout for the report parameter under additional outputs section of the training tool).

3. Yes. The best performing model is what is saved in the *dlpk. At the moment, there is no option to choose any other model apart from the best model for inference.

4. The variables that were used during the training are saved in the dlpk and when you provide a feature class/shape file to the inference tool, the tool tries to find the variables with the same names in the provided shape file. In case the names of your fields in inference shape file are different from what you had used during the training, then you will need to provide a mapping of the fields in the Map Explanatory variables section of the Inference tool.

Hope this helps.

~ Karthik

Edit - You can also try to run the tool in Advanced mode. The tool might pickup more models when you choose this mode.

ChristopherBerry1

Hi Karthik,

Thanks very much for your detailed response.

I haven't been able to find any details of the compisition of the ensemble. Will this always be what is used? I would imagine an ensemble is always going to give a bettter result than an individual model?

TranDuy

Hi Nicholas,

I am running the Train using AutoML tool in ArcGIS Pro, it works in my PC (version 3.0) but in my laptop with ArcGIS Pro 3.1, it's not working, although I used the same dataset. I have attached the problem here. Can you please let me know how to fix this? Many thanks

AutoML trained model issue.jpg

Many thanks,

Duy

SupratimBanik

Hi Christopher,

An ensemble will always be created but it does not necessarily mean that it will be the best model. There will be a few cases where the individual models work better compared to ensemble. But in majority of the cases ensemble will be better than individual.

In order to get the composition, you will need to generate the report. In the report, if you click on the ensemble model, the composition will be displayed. You can generate the report by populating the parameter Report (In additional outputs) before you run the tool. The tool output window will then show you a link to the report that was generated.

Karthik

SupratimBanik

Hello Duy,

Would you be able to share the dataset? We will have to try recreate this issue.

Karthik

xingchenc

Hi @NicholasGiner1

Very interested in this rideshare demands prediction demo(EsriDevEvents/ds-usa-plenary-automl-2023: This demo highlights new AutoML capabilities within the GeoAI toolbox during the Day 2 DevSummit 2023 Plenary. (github.com)), but unable to downlaod the ArcGIS Pro project package file listed on the webpage, need to login with an Esri credential, can you please make it public?

xingchenc

@NicholasGiner1

Hi Nicholas,

I would like to report one possible bug, when I am using the AutoML tool in a disconnected environment or when current portal is set to a portal with self-signed certificate, the tool will return error messages.

Upon checking the source code, I noticed it trying to connect to the portal, can you explain why this is necessary? Becuase I didn't see any parameters in the tool that related to the portal.

PriyankaTuteja

@xingchenc Thanks for reporting the issue. We have logged this as a bug internally. Please expect the fix with Pro 3.3 release.

xingchenc

@PriyankaTuteja

Hi Priyanka,

Thank you ,that's good to know!

NicholasGiner1

Hi @xingchenc - thanks for reaching out about this. I've reshared the link publicly, please let me know if you have any issues accessing it.

DevSummit2023_AutoML_PlenaryDemo_April2023.ppkx

thanks,

Nick Giner

Ol_VUW

@PriyankaTuteja @NicholasGiner1

Hi,

I am using the Train AutoML and Predict AutoML tools in 3.1. I can successfully train my model using raster data as input, however when using the predict to raster tool it keeps asking for a point file (i.e. the red asterisk never disappears) and so the tool won't run. "Parameter missing error".

If I train the tool on a point dataset (i.e. extract values from the rasters first), and then also run predict on a point dataset (again with values extracted from the rasters) everything works fine.

Do you have any suggestions why predict to raster won't run?

Thanks for your help.

Cheers,

Ollie

mattCSimon1

My goal is to predict 28 different ecozone types (acidic cove, northern hardwood, etc) using plot data (n=5,000ish) and about 44 different environmental variables as explanatory variables. Our study is area is large (~7 million acres) and the spatial resolution is fairly fine (5m). I'm running Pro 3.4.0 and can successfully train model on a smaller clipped out area (~700k acres) using only 6 explanatory variables for a smaller subset of ecozone types (n=18, where the total number is greater than 1% of total plots). However, the Predict Using AutoML errors out consistently after 2.5 hours and provides a generic Error 260204.

We are testing this modeling approach to see if it works better than our current approach and I have 10 more days to figure it out as I'm using a trial license to gain tool access.

Any help would be appreciated. These seems like a good application of the tools but maybe I'm missing something simple.

Attached images show result of trying to predict just 4 of the ecozones based on a model trained for those 4 ecozones only.

2024-11-18_15-47-20.png

2024-11-18_15-47-26.png

SupratimBanik

Thank you for reporting this.

Can you please share the following details:

1. How long did the Training tool run? If you have the tool output, please share that.
2. Can you try running the prediction on a single ecozone and see if that runs fine? For the purpose of testing, you can also try running it on the same data on which the training was done.
3. Can you please check if your rasters display name is different than actual raster name? If yes, try changing the display name to the actual name. (This is a workaround until you have the fix which has been made in the code recently)
4. If the above step is not applicable or does not address your problem, would it be possible to share some sample data and trained model with us to replicate the issue?

Regards

Karthik

mattCSimon1

Thank you for the quick responses Karthik - answers below:

Training took 50 seconds. screen shots of the input and output attached.
Basic model training on a single ecozone fails in 3 seconds for all model types (warning 260151).
Display name is the same as actual raster name.
Zip attached with the 4 ecozone types (corresponds to all screen shots I've shared) and the successfully trained model (.dlpk)

attempting to predict 'ecozone' based on the other six variables in the file.

I could clip rasters to a smaller area, but I still wouldn't be able to attach those without exceeding upload limit.

4class_trainModel_input.png

4class_trainModel.png

mdl_4types.zip

mattCSimon1

@Ol_VUW Did you ever figure this out? I might be having the same issue.

SupratimBanik

Thank you for sharing the details. We would like to know more. Can you please reach out to me at sbanik@esri.com?

Thanks

Supratim

Participe da Conversa

Membros conectados podem postar, seguir atualizações e mais. Novo aqui? Registre uma conta gratuita.

Links Rápidos

Community Help
Find useful guides, FAQs, and documents to help you navigate and make the most of Esri Community.