See discussion in PR #238 for further details. There appears to have been a change in tensorflow between version 1.3.0 and 1.13.1 which created that bug, and I'm not 100% that not closing the session is the proper way to handle that, but the fix worked as a temporary measure.

Sometimes, one wants to paste a decimal position (e.g., 0.5, 3.3, 4.4) from somewhere into webKnossos which is not supported. However, our internal position is stored as floats, so it would make sense to support this input format. Proposed approach:

• correctly parse float positions
• also jump to that precise float position
• on blur, render the position as integers (as it's done currently

Supplying incorrect pixel sizes can can use so much RAM for registration that crashes can occur. E.g. supplying -x 2 -y 2 -z 5 rather than -x 1 -y 2 -z 5 could (incorrectly) require ~120GB RAM for registration.

This should be avoided by:

• Highlighting the importance of correct pixel sizes to be passed.
• Encouraging the use of --metadata over specifying the pixel sizes (and adding as m