loading of previously trained models about clstm HOT 7 CLOSED

Please paste here the full commands you used to:

• Train the model
• Load the model

from clstm.

The bash-script I used contains the following commands for training (loading-attempt was made with commented lines)
#!/bin/bash
set -x
set -a
report_every=10
save_every=1000
ntrain=200000
dewarp=center
display_every=10
test_every=1000
testset=mub1test.h5
hidden=100
lrate=1e-4
save_name=mub1
report_time=1
#load=mub1-12000.clstm
#start=12000
./../clstm-master/clstmocrtrain mub1gt mub1test

from clstm.

If you haven't noticed, I already solved the problem. PR #53 has 1-line fix. You can apply it and recompile. It seems to work as expected now.
Thanks for reporting !

from clstm.

The 'start' environment variable isn't needed in the second run, only the 'load'.

from clstm.

Current code still fails with "FATAL missing parameter"
at network_info(clstm.net) just after clstm.load().
The first call is OK. but recurrently-called(2nd time) network_info(net, ".Stacked") fails.(with super == 0x0)

This parameter seems to be "learning_rate". Is it stored in saved .clstm file?
Or net->attr must have super instance as default values.

from clstm.

What do you mean by "current code"?
Tom didn't merge my PR with master yet, so if you want to test it you should apply the patch locally (to a new branch) and recompile.

It works for me and for @stexandev.

from clstm.

It's my fault. patched code works perfectly.
Thanks.

from clstm.