no training instances

[umkhanqta]

Trying to run the command python3 train_taskonomy.py -d=/taskonomy-sample-model-1-master/ -a=xception_taskonomy_new -j 4 -b 96 -lr=.1 --fp16 -sbn --tasks=dnerac -r

but found 0 training instances.

Kindly help me out

[tstandley]

How is your data arranged? The data must be aranged in inputs:
root/rgb/building/point_x_view_x.png labels: root/task/building/point_x_view_x.png

[umkhanqta]

The data is arranged as following image:

I did not find the "task" folder

Can you send the latest/correct dataset link?

[tstandley]

There is no download link. You must organize the data yourself. The code was written when the data was laid out differently. If you make a script to do it, would you mind sharing?

There is no folder with the name "task". That referred to the name of the task (for example, normals)

[umkhanqta]

so I need to add a "building" folder and put "rgb" in that. Also rename all the files and remove "_domain_rgb"?

[tstandley]

Sorry, I don't know how the data you have is organized. I downloaded the dataset years ago and they've since changed it.

[umkhanqta]

OK thanks for the help

[umkhanqta]

All the paths are correct but this section is not running. because "building" is not in the "model_whitelist" any suggestion?

[tstandley]

My guess is that I wasn't clear before that "building" isn't the string literal 'bulding', but the actual building name of the model. Each image comes from a 3d model of a building. There are about 400 buildings IIRC. Each building has a name. So the data needs to be laid out such that the highest directory is the one that contains all the tasks. Within each task directory is a directory for each building, and within each of those is a bunch of views. So for example, one file is located at: "/taskonomy_data/normal/carpio/point_52_view_5.png"

Here "carpio" is the name of the building.