I am currently working on training a deep learning model in ArcGIS Pro that will detect geothermal imagery in Landsat 8 imagery. I have been following this tutorial: Train Deep Learning Model—Help | Documentation .
I am currently stuck on the last step in the process, training the model itself. I have created a proper folder for all the exported training data and am working with Python now. Here is my code below:
from arcpy.ia import *
# Check out the ArcGIS Image Analyst extension license
#Define input parameters
in_folder = "C:\\Users\\gphy426\\Documents\\Geothermal Pombo\\training_data_geoproc"
out_folder = "C:\\Users\\gphy426\\Documents\\Geothermal Pombo\\trained_model"
max_epochs = 20
model_type = "FEATURE_CLASSIFIER"
batch_size = 2
arg = "# #"
learning_rate = None
backbone_model = "RESNET34"
pretrained_model = None
validation_percent = 10
stop_training = "STOP_TRAINING"
TrainDeepLearningModel(in_folder, out_folder, max_epochs, model_type,
batch_size, arg, learning_rate, backbone_model, pretrained_model,
The error I seem to be running into now is "unknown dataset_type = KITTI_rectangles".
I've attached an image of the exact error that I am receiving to this forum post. I have done a bit of forum reading to see how I could bypass this issue, but the main solution I've seen regarding pascal VOC objects does not work for my case, considering that the reference raster is two bands (not the optimal 1 or 3).
I run into the same issue when using the geoprocessing tool "train deep learning model".
Any thoughts or suggestions are appreciated.
The training api and tools use "transfer learning" to train our models which means your data needs to be 3 Band 8 bit. What I would suggest to make your dataset training capable, is to use a composite band tool/function on your dataset to make it 3 band( Here is some help How To: Convert individual raster bands into a color or multiband image ). If it is not 8 bit you can use the stretch function to make it 8 bit. Then run the Export training tool (using PASCAL VOC Objects option ) and then train your data. Hope this helps.