Introduction

from PIL import Image
from matplotlib import pyplot as plt
from utils import working_directory, interactive_plot, get_errors_index
with working_directory('..'):
    from polar import load_sample_data, POLAR

kaleido is only required to save Plotly figures as jpeg image, so that they can be displayed in the web documentation.
It is not required when running this notebook.

%pip install -U kaleido

Requirement already satisfied: kaleido in /data/mambaforge/envs/geom/lib/python3.8/site-packages (0.2.1)

views, degradations, R_abs_gt = load_sample_data('outliers')

model = POLAR(**degradations)

model.fit(views.cuda())

:: Initializing model ...
    - Using device: cuda.
Cached SO(3) Lookup Table not found. It will be created on the fly and saved for later use. This will occur extra runtime.
    - Running coarse exhaustive search ...
LOOP 1
:: Running 100 parallel joint optimizations ...
:: Computing restart angles ...
:: Running 400 parallel multistart optimizations ...
Multistart improved 16 objects:
tensor([12, 19, 21, 27, 29, 32, 35, 42, 46, 56, 70, 75, 76, 86, 94, 95])
--------------------------------------------------------------------------------
LOOP 2
:: Running 100 parallel joint optimizations ...
:: Computing restart angles ...
:: Running 400 parallel multistart optimizations ...
Multistart improved 1 objects:
tensor([48])
--------------------------------------------------------------------------------
LOOP 3
:: Running 100 parallel joint optimizations ...
:: Computing restart angles ...
:: Running 400 parallel multistart optimizations ...
Multistart improved 1 objects:
tensor([48])
Multistart improved same objects twice in a row.
Reducing lr to 0.001.
--------------------------------------------------------------------------------
LOOP 4
:: Running 100 parallel joint optimizations ...
:: Computing restart angles ...
:: Running 400 parallel multistart optimizations ...
Multistart improved 1 objects:
tensor([37])
--------------------------------------------------------------------------------
LOOP 5
:: Running 100 parallel joint optimizations ...
:: Computing restart angles ...
:: Running 400 parallel multistart optimizations ...
Multistart converged.
:: Running 100 parallel joint optimizations ...

get_errors_index(model.get_matrix(rotation_only=True), R_abs_gt)

tensor([89])

loss = model.get_mean_losses(multistart=True)

plt.plot(loss); plt.show()

This is just to display the Plotly figure as an image in the web documentation.
If you are running this notebook, just call interactive_plot(model.template.detach())

fig = interactive_plot(model.template.detach(), return_fig=True)

fig.show()

fig.write_image("template.jpeg")
Image.open('./template.jpeg')