Avoid problem during averaging when parameter-tying exists

[tzyll]

Why?

When parameter-tying exists in a pytorch network, different keys share the same memory which will be processed multiple times if not separated using clone().

[sw005320]

Will it break the compatibility, or can we still use the old model checkpoint?

[tzyll]

Hi, the modification won't affect the old model checkpoint.

Will it break the compatibility, or can we still use the old model checkpoint?

[sw005320]

Thanks!
One more request.
Could you show me that the old one and this do not change the results (model parameters themselves and ASR results)?

[codecov]

Codecov Report

Merging #5533 (e50da1c) into master (0d0ab98) will decrease coverage by 1.06%.
Report is 5 commits behind head on master.
The diff coverage is n/a.

@@            Coverage Diff             @@
##           master    #5533      +/-   ##
==========================================
- Coverage   70.31%   69.25%   -1.06%     
==========================================
  Files         711      710       -1     
  Lines       65757    65670      -87     
==========================================
- Hits        46237    45481     -756     
- Misses      19520    20189     +669     

Flag	Coverage Δ
test_configuration_espnet2	∅ <ø> (?)
test_integration_espnet1	63.02% <ø> (?)
test_integration_espnet2	48.61% <ø> (ø)
test_python_espnet1	?
test_python_espnet2	51.39% <ø> (+0.03%)	⬆️
test_utils	22.19% <ø> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

see 147 files with indirect coverage changes

📣 Codecov offers a browser extension for seamless coverage viewing on GitHub. Try it in Chrome or Firefox today!