Closed 2320sharon closed 1 year ago
I updated doodleverse-utils to hopefully avoid this bug. please try it out https://pypi.org/project/doodleverse-utils/0.0.34/
This remains an issue even with doodleverse_utils 0.0.34. For a temporary solution these models will be taken out in coastseg 0.0.72
Due to information provided by @2320sharon , the est_label_multiclass
version implemented (for some reason!) in zoo was working but not the one in 'verse-utils-34 ... so I put the zoo versions of that function plus sister functions est_label_binary
and do_seg
into 'verse-utils/prediction_imports.
then I updated the Zoo and Gym functions so they use the doodleverse-utils versions of the above functions. Specifically, I updated model_functions
and model_inference_funcs
in zoo and tested various models with segment_orthomosaic.py
and `select_model_and_batch_process_folder.py. I also tested gym/seg_images_in_folder.py
I tested both 2- and multiclass models, and segformer and resunet models, on satellite imagery and other imagery, including RGB and MNDWI and NDWI. It worked great every time :tada:
please use https://pypi.org/project/doodleverse-utils/0.0.35/ and fingers crossed!!
I was worried this might not work, but its works great! I tested it with all the models in both BEST
and ENSEMBLE
model and it works great! I also tested this change with all the models and it works!
Due to the latest update to doodleverse-utils v0.0.33 MNDWI and NDWI models no longer function and throw error similar to the one pasted below. These models need the option to pass only a single band to the model such with code such as
tf.squeeze(model(tf.expand_dims(image[:,:,1], 0)))