Failing to clear wrapped line

[Liamolucko]

Environment

Windows build number: 10.0.18362.657
Windows Terminal version (if applicable): Not sure; latest version on Microsoft Store as of 2020-03-01

Any other software?
PowerShell 6.2.4
Python 3.7.6
TensorFlow 1.14.0

Steps to reproduce

Run a TensorFlow program which attempts to clear a line that exceeds the size of the window and is wrapped onto the next line.
If the window is wide enough to fit the entire line, this is not an issue.

Expected behavior

The entire 'line' should be deleted

212/723 [=======>......................] - ETA: 25:52 - loss: 1.6581 - rpn_class_loss: 0.0278 - rpn_
bbox_loss: 0.3465 - mrcnn_class_loss: 0.3752 - mrcnn_bbox_loss: 0.4602 - mrcnn_mask_loss: 0.4483

Actual behavior

Only the last line of the wrapped line is cleared, leaving behind lots of cut off lines.

190/723 [======>.......................] - ETA: 27:10 - loss: 1.6521 - rpn_class_loss: 0.0285 - rpn_
191/723 [======>.......................] - ETA: 27:06 - loss: 1.6524 - rpn_class_loss: 0.0284 - rpn_
192/723 [======>.......................] - ETA: 27:02 - loss: 1.6509 - rpn_class_loss: 0.0283 - rpn_
193/723 [=======>......................] - ETA: 26:58 - loss: 1.6483 - rpn_class_loss: 0.0284 - rpn_
194/723 [=======>......................] - ETA: 26:54 - loss: 1.6486 - rpn_class_loss: 0.0283 - rpn_
195/723 [=======>......................] - ETA: 26:51 - loss: 1.6507 - rpn_class_loss: 0.0282 - rpn_
196/723 [=======>......................] - ETA: 26:47 - loss: 1.6606 - rpn_class_loss: 0.0281 - rpn_
197/723 [=======>......................] - ETA: 26:43 - loss: 1.6626 - rpn_class_loss: 0.0282 - rpn_
198/723 [=======>......................] - ETA: 26:40 - loss: 1.6598 - rpn_class_loss: 0.0282 - rpn_
199/723 [=======>......................] - ETA: 26:36 - loss: 1.6653 - rpn_class_loss: 0.0281 - rpn_
200/723 [=======>......................] - ETA: 26:33 - loss: 1.6666 - rpn_class_loss: 0.0282 - rpn_
201/723 [=======>......................] - ETA: 26:30 - loss: 1.6665 - rpn_class_loss: 0.0283 - rpn_
202/723 [=======>......................] - ETA: 26:28 - loss: 1.6713 - rpn_class_loss: 0.0283 - rpn_
203/723 [=======>......................] - ETA: 26:24 - loss: 1.6662 - rpn_class_loss: 0.0283 - rpn_
204/723 [=======>......................] - ETA: 26:20 - loss: 1.6652 - rpn_class_loss: 0.0282 - rpn_
205/723 [=======>......................] - ETA: 26:16 - loss: 1.6606 - rpn_class_loss: 0.0281 - rpn_
206/723 [=======>......................] - ETA: 26:13 - loss: 1.6575 - rpn_class_loss: 0.0281 - rpn_
207/723 [=======>......................] - ETA: 26:09 - loss: 1.6591 - rpn_class_loss: 0.0280 - rpn_
208/723 [=======>......................] - ETA: 26:05 - loss: 1.6565 - rpn_class_loss: 0.0279 - rpn_
209/723 [=======>......................] - ETA: 26:03 - loss: 1.6592 - rpn_class_loss: 0.0279 - rpn_
210/723 [=======>......................] - ETA: 25:59 - loss: 1.6577 - rpn_class_loss: 0.0280 - rpn_
211/723 [=======>......................] - ETA: 25:56 - loss: 1.6591 - rpn_class_loss: 0.0279 - rpn_
212/723 [=======>......................] - ETA: 25:52 - loss: 1.6581 - rpn_class_loss: 0.0278 - rpn_
bbox_loss: 0.3465 - mrcnn_class_loss: 0.3752 - mrcnn_bbox_loss: 0.4602 - mrcnn_mask_loss: 0.4483

[Liamolucko]

This is probably related to the issues in #4200

[Liamolucko]

This doesn't seem to be an issue with Windows Terminal - it happens in both PowerShell and Command Prompt in the default windows shell, so it's probably an issue with TensorFlow.

[DHowett-MSFT]

Huh, thanks for following up. Does it reproduce in other terminals that work like xterm? That’s the compatibility level we’re going for :)

[Liamolucko]

I tried it in Git Bash & Cygwin, since they both use mintty, but it doesn't seem to support deleting lines at all.

[Liamolucko]

Okay, I've figured out that the specific problem is that TensorFlow is using carriage return (\r) for this.

My test:

for i in range(5):
    print((str(i) * 200) + '\r', end='')

print()

Note: I added the empty print() statement at the end so that the terminal wouldn't overwrite the last line.

Expected:

444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444
44444444444444444444444444444444444444444444444444444444444444444444444444444444

Output:

000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000
111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111111
222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222222
333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333333
444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444444
44444444444444444444444444444444444444444444444444444444444444444444444444444444

Note: This also occured in VS Code's integrated terminal

Strangely, MinTTY had the same behaviour as Command Prompt and Windows Terminal this time.

[Liamolucko]

Should this issue be fixed? It seems to happen in pretty much every terminal.

It behaves this way in:

• Windows Terminal
• Command Prompt
• MinTTY
• VS Code Integrated Terminal
• Terminal.app

[DHowett-MSFT]

Say, are you using python for Windows or python inside WSL?

[Liamolucko]

Python for Windows, versions 3.7.6 and 3.8.0.

I just tested it on WSL though, same behaviour (Python 3.6.7)

[zadjii-msft]

I'm gonna say that this should be resolved as "by-design". I just tried this out in gnome-terminal and I'm seeing the same behavior with your simple repro case as what happens in conhost and Windows Terminal. I'd bet that this is just a misunderstanding by the TensorFlow maintainers that \r won't take you to the beginning of a wrapped line if the line started on another row, it'll only take you to the start of the current row.